# Sell My Data To AI: full text Sell My Data To AI is an independent broker that helps companies license their data to vetted AI buyers, and helps AI teams source rights-cleared training data from vetted sellers. --- # How brokering an AI data deal works URL: https://sellmydatatoai.net/how-it-works We sit between companies that hold valuable records and the AI teams that need them. Here is exactly what happens at each step, and what you control. ## Key takeaways - It starts with a free 30-minute consultation, and we reply within two business days to arrange it. - You approve every buyer, every sample and every license term. You can withdraw at any point before signing. - We do not take custody of your raw data. Buyers see redacted samples under a mutual NDA, and full data moves only after you sign a license directly with the buyer. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## For companies licensing data 1. **Free consultation**: Thirty minutes on what you hold, which systems it sits in and how many years it covers. No data leaves your systems. We tell you honestly if we think buyers will not be interested. 2. **Valuation and rights pre-check**: A written view of likely buyers, a realistic range based on published buyer terms and closed deals, and the rights issues that need clearing first. Our fee is set out in writing at this point. 3. **Data manifest and exclusions**: We help you list each data set by system, record count and date range, and decide what to leave out: HR files, privileged material, client work you cannot license. See [how to build a data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest). 4. **Vetting**: We run the eight checks in our [vetting standard](https://sellmydatatoai.net/how-we-vet). Most issues found here can be fixed by excluding records or getting a client's consent. 5. **Redacted sample to verified buyers**: Only buyers who have passed our checks see a representative, de-identified sample, under a mutual NDA. You approve each buyer before they see anything. 6. **Offers compared on the same terms**: Offers are only comparable when they cover the same records, uses, exclusivity and term. We put them side by side so the highest headline is not mistaken for the best deal. 7. **License signed directly with the buyer**: You contract with the buyer, not with us. Your own lawyer reviews it. We help you negotiate scope, deletion, de-identification certification and payment triggers. 8. **Delivery, acceptance and payment**: A fixed export of only the approved data goes through the buyer's approved transfer method. The buyer checks it against agreed acceptance tests and pays on the agreed schedule. ## For AI teams buying data 1. **Free consultation**: What the model needs to learn, what data would teach it, and the license terms you need. We turn that into a written requirement. 2. **Buyer verification**: We verify your company identity, intended use, security and retention commitments and onward transfer limits. You do this once. 3. **Sourcing and vetting**: We search for sellers whose records match, and put each through the same eight-point vetting standard before you hear about them. 4. **Shortlist with vetting files**: For each seller: rights evidence, provenance record, de-identification method, manifest and proposed license terms. 5. **Samples under NDA**: Representative samples, drawn at random and labeled with how they were drawn, so you can test fit with your own pipeline. 6. **Contract, delivery and acceptance**: You contract directly with the seller. We help agree acceptance tests, staged payments and what happens if prohibited material turns up after delivery. ## How long does a deal take? Buyers' own published timelines give a guide. Handshake AI says pricing comes in 12 to 24 hours and its process takes 7 to 14 business days. Polyshares says it makes offers within one to two weeks. micro1 allows up to 30 days to approve delivered data and pays within 30 days of invoice. Allow six to twelve weeks from first consultation to payment for a typical company archive, longer if client consents are needed. ## What we never do - We do not take custody of your raw data. Buyers see redacted samples under a mutual NDA, and full data moves only after you sign a license directly with the buyer. - We are independent of the buyers we work with. We do not own an AI model and we do not buy data for our own use. - We never send your company name or any record to a buyer you have not approved. - We never give a buyer's tools direct access to your live systems. --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How we vet sellers and buyers URL: https://sellmydatatoai.net/how-we-vet Trust in this market breaks down at two points: buyers cannot verify what they are buying, and sellers cannot verify who they are selling to. Our vetting standard closes both gaps. ## Key takeaways - Every seller passes 8 checks: authority to sign, ownership and chain of title, contract restrictions, privacy notice and lawful basis, third-party content, de-identification method, representative sample and manifest accuracy. - Every buyer passes 5 checks: company identity, intended use, security and retention, onward transfer limits and ability to pay. - We do not take custody of your raw data. Buyers see redacted samples under a mutual NDA, and full data moves only after you sign a license directly with the buyer. ## Why a written standard matters A license tag on a dataset does not prove it can be used legally, and a buyer's logo does not prove what will happen to your records. Buyers' lawyers ask the same questions on every deal, so we ask them first and write down the answers. Each side gets the evidence, not just our word. ## The eight seller checks 1. **Authority to sign**: The person signing can bind the company, or the trustee, liquidator or board can act for a closed company. 2. **Ownership and chain of title**: The company created or acquired the records and holds the rights to license them, including IP assignments from staff and contractors. 3. **Contract restrictions**: Client contracts, NDAs, vendor terms and data processing agreements do not prohibit the proposed use. 4. **Privacy notice and lawful basis**: The privacy notice in force when the data was collected, and the applicable law, allow the disclosure, or the personal data is excluded. 5. **Third-party content**: Licensed content, open-source code and material owned by others is identified and either cleared or removed. 6. **De-identification method**: Direct and indirect identifiers are removed by a documented method, and a re-identification test is run on a sample. 7. **Representative sample**: The sample is drawn at random across the archive, not hand-picked, so a buyer can judge the whole package. 8. **Manifest accuracy**: Record counts, date ranges, systems and languages in the data manifest match what is actually delivered. > **Exclude rather than hope** Where a record type carries high risk, such as HR files, health records, privileged legal material or payment card data, we recommend leaving it out entirely rather than relying on de-identification. Law firm Frankfurt Kurnit Klein & Selz gives the same advice in its October 2026 guidance to companies approached to sell data. ## The five buyer checks 1. **Company identity**: Registered entity, beneficial owner where relevant, and a named signatory. 2. **Intended use**: Training, evaluation or fine-tuning, which models, and whether outputs will be commercialized. 3. **Security and retention**: Where data will be stored, who can access it, and when raw copies are deleted. 4. **Onward transfer limits**: Whether data may pass to affiliates, subprocessors or model partners, and on what terms. 5. **Ability to pay**: For larger deals, evidence the buyer can meet the price and payment schedule. ## What a vetting file contains | Item | What it shows | | --- | --- | | Rights evidence | Who signed, on what authority, and the contracts and privacy notices that permit the license | | Provenance record | Where each record came from, which systems and dates, and what was excluded | | De-identification method | What was removed or transformed, by whom, and the result of a re-identification test | | Data manifest | Record counts, date ranges, languages and formats, matched to the delivery | | Representative sample | Drawn at random across the archive, labeled with the sampling method | | Buyer verification summary | Shared with the seller: identity, intended use, security and retention commitments | ## What vetting does not do Vetting reduces risk. It does not remove it, and it is not legal advice. We are a broker, not a law firm. We tell you what buyers will check and help you prepare, but you should take independent legal advice before signing any license. --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Our fees URL: https://sellmydatatoai.net/fees No upfront fees, no listing fees, no retainers. We are paid when a deal you have chosen completes. ## Key takeaways - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. - Buyers pay nothing to submit a requirement or review a shortlist. Any fee is agreed in writing before an introduction. - We are independent of the buyers we work with. We do not own an AI model and we do not buy data for our own use. ## For sellers Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. - The consultation, valuation and rights pre-check are free. - Our fee is a success fee, calculated on the value of a completed license. - The rate and how it is calculated are set out in writing before we introduce you to any buyer. - If no deal completes, you pay nothing. ## For buyers Buyers pay nothing to submit a requirement or review a shortlist. Any fee is agreed in writing before an introduction. ## How this compares Published fees across the market vary widely. Opendatabay's commission is tiered from 5% to 30% of the sale price, according to a February 2026 guide on GISuser. FileYield's FAQ states 8% for sellers and 6% for buyers. SellMyData takes a commission only when a sale closes but does not publish the rate. Some training-data companies such as Mercor charge sellers nothing and cover extraction costs, because they are the buyer. See [AI data marketplaces, brokers and direct deals compared](https://sellmydatatoai.net/guides/ai-data-marketplaces-compared). ## Why a success fee A success fee means we are only paid when you are, and only on a deal you chose. It also means we have a reason to tell you early if your data is unlikely to sell, rather than charging you to find out. --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How to sell your company's data to AI companies URL: https://sellmydatatoai.net/guides/how-to-sell-data-to-ai-companies Reviewed: 2026-11-08 You sell company data to AI companies by licensing a fixed, rights-cleared export of your records to a buyer such as micro1, Mercor or a direct lab. Inventory what you hold, exclude what you cannot license, de-identify personal data, share a representative sample under NDA, compare offers on the same terms, then sign a license with acceptance tests and a payment schedule. ## Key takeaways - Buyers publish real payouts: micro1 lists $100,000 to $3 million or more per company, and Mercor's industry listings run from $20,000 to $1.5 million. - Holding records is not the same as having the right to license them. Client contracts, NDAs, privacy notices and third-party content all need checking first. - Never give a buyer access to your live systems. Export a fixed copy of only the approved records. - Published payment timings run from about 2 weeks to 2 months after signing or delivery, depending on the buyer. - Compare offers only when they cover the same records, uses, exclusivity and term. AI labs and the companies that supply them are paying operating businesses for their internal records. In 2026 the buyers include specialist training-data companies such as micro1, Mercor, Scale AI and Handshake AI, direct buyers such as Polyshares, and, in at least one case, Google itself. This guide covers the whole process: who buys, what they want, the rights checks that decide whether you can sell at all, how to prove quality without handing your data over, what the published payouts and timings are, and where sellers lose money. ## Do AI companies really buy company data? AI companies do buy company data, and the market grew quickly in 2025 and 2026. MIT Technology Review reported in October 2026 that demand appeared suddenly as AI labs hit the limits of public and synthetic datasets, and that buyers now include foundation labs, vendors that build simulated workspaces and mature tech companies training their own models. The clearest public example is Spirit Airlines. Google won an August 2026 bankruptcy auction for Spirit's business data with a $10 million bid, according to MIT Technology Review. The archive held about 100 million emails and 500 million Microsoft Teams messages. Fortune reported that micro1 then proposed $12.5 million, and The Next Web reported on October 9, 2026 that the sale still needed the court's approval. Smaller deals are far more common. SimpleClosure, which helps startups wind down, told MIT Technology Review it had handled more than 150 data and code deals through its AssetHub platform. Its chief executive, Dori Yona, told Fortune that payments are typically $10,000 to $100,000 per company. ## Who buys company data for AI training? Company data is bought by five kinds of organization, and most sellers deal with a training-data company rather than a lab. The type of buyer decides how the deal is structured, who sees the data and how you are paid. | Buyer type | Examples | How they work | Published payout | | --- | --- | --- | --- | | Training-data company | micro1, Mercor, Scale AI, Handshake AI | License your records, process them and supply AI labs | micro1: $100,000 to $3 million or more. Mercor: $20,000 to $1.5 million by industry | | Direct buyer | Polyshares | Buys licenses outright, exclusive by default | $100,000 to $2 million typical | | Aggregator | Endless Data | Licenses non-exclusively to several labs | Up to $4 million | | Licensing agent | Troveo | Sells your data repeatedly and shares revenue | Full-company deals start at six figures | | Wind-down broker | SimpleClosure AssetHub | Sells code and data from closing startups | Tens to hundreds of thousands of dollars | Payout figures for micro1 come from its own data partnerships page, checked November 2026. The others come from each buyer's published terms as compiled by Don't Do It Yourself, checked October 26, 2026. Our [data buyers directory](https://sellmydatatoai.net/data-buyers) profiles each one. ## What company data do AI companies want? AI companies want records that show how real work gets done, with enough history and context to follow decisions from start to finish. micro1's data partnerships page lists standard operating procedures, process documentation, internal knowledge bases, workflows, templates, CRM records and support tickets. Mercor's industry listings name the records it wants by sector: complete files from closed matters for law firms, project records, drawings and site photos for construction companies, and private code with its history for software companies. DataVendor notes that labs pay more for human-written code, complex products and systems languages such as Rust, Go, C and C++. Contact lists, scraped web data and data you bought from someone else rarely sell. Our guide to [what data AI companies buy](https://sellmydatatoai.net/guides/what-data-do-ai-companies-buy) covers each record type in detail. ## How do you sell company data to an AI company, step by step? Selling company data to an AI company takes nine steps, and the first four happen before any buyer sees anything. Skipping the early steps is the most common reason deals stall at the contract stage. 1. **Inventory what you hold**: List each system (email, Slack or Teams, Google Drive, Jira, GitHub, your CRM), what it contains, how many records and which years. Frankfurt Kurnit Klein and Selz recommends knowing this before any offer arrives. 2. **Check your right to license**: Read the client contracts, NDAs, vendor terms and privacy notices that cover each system. Mark each dataset as included, excluded or needing consent. 3. **Decide the exclusions**: Take out HR and payroll files, privileged legal material, health information, union communications and other parties' confidential data. MIT Technology Review reports that suppliers routinely leave HR records out. 4. **Build a data manifest**: Record counts, date ranges, systems, languages and file formats for every dataset you will license. Buyers price from this document. See [how to build a data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest). 5. **De-identify personal data**: Remove direct and indirect identifiers by a documented method and test a sample for re-identification. Agree the method with the buyer before work starts. 6. **Share a representative sample under NDA**: Draw a random sample across the whole archive, redact it, and share it only with verified buyers under a mutual NDA. 7. **Compare offers on equal terms**: Line up each offer by records covered, permitted uses, exclusivity, term and payment triggers before you look at the headline price. 8. **Sign the license**: Have your own lawyer review it. Check permitted uses, rights in derived assets, deletion, de-identification certification and caps on your liability. 9. **Deliver a fixed export and get paid**: Send only the approved data through the buyer's secure transfer method. The buyer runs the agreed acceptance tests and pays on the agreed schedule. ## What must you check before you agree to sell? Before you agree to sell, check that you actually have the right to license each dataset for AI use. Nixon Peabody's October 2026 diligence alert puts it in three tests: whether you can transfer the data, whether the buyer can use it for AI, and whether the buyer can keep and commercialize what training produces. Frankfurt Kurnit Klein and Selz lists the usual blockers. Email and chat contain other companies' confidential information under NDAs. Licensed material, open-source code and vendor-generated data can carry their own restrictions. Slack's API terms, it notes, bar third-party apps from using API data to train large language models. A sale can also waive legal privilege or expose trade secrets. Personal data adds a second layer. You must handle it in line with the privacy notice in force when you collected it, and the US Federal Trade Commission has warned that quietly adding AI training to a notice may be unfair or deceptive. Our guide on [whether you can legally sell company data](https://sellmydatatoai.net/guides/can-you-legally-sell-company-data-for-ai) goes through each law. > **Rule of thumb** If a dataset needs a lawyer's opinion to justify including it, exclude it. Frankfurt Kurnit Klein and Selz advises leaving high-risk data out entirely rather than relying on de-identification to make it safe. ## Which records should you leave out of a sale? The records to leave out of a sale are the ones where the risk is high and the training value is low. Frankfurt Kurnit Klein and Selz lists privileged material, HR and payroll records, health information, union communications, third-party confidential data, licensed data and data from regulated systems as categories to exclude. | Record type | Why it is usually excluded | Example from a reported deal | | --- | --- | --- | | HR and payroll files | Sensitive personal data with little training value | Suppliers routinely leave HR records out (MIT Technology Review, October 2026) | | Customer profiles and lists | Personal data that buyers do not need for training | Excluded from the Spirit Airlines sale (Fortune, September 2026) | | Booking and payment records | Consumer data with high privacy risk | Spirit booking, refund and inflight purchase records kept out of the Google deal (The Next Web, October 2026) | | Privileged legal advice | Disclosure can waive privilege | Listed as a high-risk category by Frankfurt Kurnit Klein and Selz | | Other companies' confidential material | Usually covered by NDAs you signed | Listed by Frankfurt Kurnit Klein and Selz and Nixon Peabody | | Licensed or third-party content | You do not hold the rights to license it on | Vendor code, stock images, paid research reports | Exclusions lower the volume but rarely kill a deal. Nixon Peabody notes that a smaller, well-documented dataset with clear provenance may be worth more than a large one with significant exclusions. Agree in the license what happens if excluded material turns up after delivery. ## How much will an AI company pay for your data? An AI company will typically pay between $10,000 and $4 million for a company's records, with most published ranges between $100,000 and $2 million, according to Don't Do It Yourself's October 2026 comparison of buyer terms. Where you land depends on years of history, linked context, how clean the rights are and whether the license is exclusive. micro1 publishes three tiers: $100,000 to $500,000 for a qualified enterprise data partnership, $500,000 to $3 million for large-scale datasets or ongoing participation, and $3 million or more for highly unique data. It also lists agent trajectory data at $1 million to $2 million per year. Actual closed sales are often lower. A shut-down crypto startup received about $25,000 for its codebase, according to MIT Technology Review. For a full breakdown by industry and deal type, read [how much your data is worth to AI companies](https://sellmydatatoai.net/guides/how-much-is-your-data-worth-to-ai), or try the [data value estimator](https://sellmydatatoai.net/data-valuation#estimator). ## How long does it take to sell data to an AI company? Selling data to an AI company takes from about 2 weeks to 2 months after signing or delivery under buyers' published terms, plus the time your own rights checks take beforehand. Published timings are measured from different points, so read each one carefully. | Buyer | Published timing | Source | | --- | --- | --- | | micro1 | Up to 30 days to review and accept the data, then payment within 30 days of invoice | micro1 data partnerships page, November 2026 | | Mercor | Typically paid 2 to 4 weeks after signing | Mercor terms via Don't Do It Yourself, October 2026 | | Polyshares | Offer within 1 to 2 weeks, payment net 30 to 60 days after approval | Polyshares terms via Don't Do It Yourself, October 2026 | | Handshake AI | Pricing in 12 to 24 hours, process 7 to 14 business days, milestone payments | Handshake AI terms via Don't Do It Yourself, October 2026 | | Endless Data | Fixed sum agreed before sharing, paid after transfer | Endless Data terms via Don't Do It Yourself, October 2026 | The slowest part is rarely the buyer. Reviewing contracts, agreeing exclusions and de-identifying unstructured messages can take longer than the commercial negotiation. ## How do you prove data quality without giving it away? You prove data quality without giving it away by sharing evidence about the whole archive rather than the archive itself. A Hugging Face forum thread from 2026 on why selling datasets is hard describes the deadlock: sellers cannot prove quality without exposing data, and buyers cannot trust quality without seeing enough. Contributors to that thread suggested several fixes that work in practice: - A stratified, random sample labeled as such, so buyers know it was not hand-picked - A locked preview where the buyer runs fixed queries without downloading records - A version-bound evidence pack with file hashes, so the delivered data provably matches what was described - A paid pilot on a small slice before the full license - Staged payments tied to acceptance tests, with clear refund terms A license tag on a file does not prove the data is legally usable. Buyers increasingly ask for the rights evidence behind it, which is why a representative sample and an accurate manifest matter as much as the data. ## Should you sell direct, through a marketplace or through a broker? Selling direct suits companies with legal resources and one obvious buyer. A marketplace suits standard, packaged datasets. A broker suits companies that want several buyers to compete and help with rights checks. | Route | Best for | Who finds buyers | Typical cost to seller | | --- | --- | --- | --- | | Direct to a training-data company | Records that match a buyer's published listing | You | Usually none stated, buyer often covers extraction | | Marketplace | Packaged, repeatable datasets | Buyers browse listings | Commission on each sale | | Licensing agent | Archives that can be licensed many times | The agent | Revenue share | | Independent broker | Unusual or high-value archives, several competing buyers | The broker | Success fee on completion | Our [comparison of AI data marketplaces, brokers and direct deals](https://sellmydatatoai.net/guides/ai-data-marketplaces-compared) goes into fees and trade-offs. At Sell My Data To AI, sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## What mistakes cost sellers money? The mistakes that cost sellers most are signing broad licenses too early and letting buyers set the terms of evidence. These are the patterns that recur in published legal guidance and buyer terms. - Giving a buyer's tools direct access to your systems, which Frankfurt Kurnit Klein and Selz warns against, instead of exporting a fixed copy - Accepting a perpetual exclusive license at a non-exclusive price. Exclusive deals fetch about 2.5 to 5 times a non-exclusive one, according to AIData.Broker - Sharing a hand-picked sample, which leads to disputes at acceptance - Ignoring rights in derived assets such as annotations, embeddings, evaluation sets and fine-tuned models, which Nixon Peabody lists as separate items to negotiate - Assuming you can withdraw data later. Once data is used to train a model, Frankfurt Kurnit Klein and Selz says it cannot practically be pulled back out - Forgetting staff and clients. Spirit Airlines' flight attendants' and pilots' unions both objected to its data sale, according to The Next Web Read [exclusive vs non-exclusive data licenses](https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses) before you accept exclusivity. Use a clause-by-clause checklist when the contract arrives. ## What happens after you sign the license? After you sign the license, you deliver the data, the buyer checks it and you get paid. micro1 describes this sequence on its data partnerships page: the company transfers data through micro1's APIs or another approved method, micro1 checks it for completeness, then payment follows acceptance. Three things protect you at this stage. A manifest that matches the delivery exactly, so acceptance is a formality. Written acceptance tests with a deadline, so the buyer cannot hold payment open. A deletion clause that says when raw copies are destroyed, which micro1 proposed for raw employment records in its Spirit Airlines bid, according to Fortune. Keep a record of what you delivered, the exclusions and the de-identification method. Frankfurt Kurnit Klein and Selz notes that prior data sales come up in later M&A due diligence, so an acquirer will ask for this file. ## What should you tell staff and clients? Tell staff and clients what is being licensed, what has been excluded and how personal data is protected, before the data leaves your systems. Workplace messages are written by people who did not expect them to train AI models, and reactions affect your reputation as much as your legal position. In the US, Frankfurt Kurnit Klein and Selz points out that employee discussions about pay and conditions can be protected activity under Section 7 of the National Labor Relations Act. Agree the message before the data leaves your systems, not after a buyer announces the deal. ## How Sell My Data To AI helps We are an independent broker. We do not own an AI model and we do not buy data for our own use. Every seller passes our eight-point [vetting standard](https://sellmydatatoai.net/how-we-vet) before any buyer sees their data, and every buyer is verified before they see a sample. We do not take custody of your raw data. Buyers see redacted samples under a mutual NDA, and full data moves only after you sign a license directly with the buyer. You approve every buyer, every sample and every license term, and you can withdraw at any point before signing. We are a broker, not a law firm, so take independent legal advice before you sign any license. To find out what your records could be worth and which rights need clearing, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### Can a small company sell its data to AI companies? A small company can sell its data, but some buyers set minimum sizes. micro1 looks for established companies with 30 or more full-time employees, while Mercor's listings start at 5 staff for law firms and accounting firms. Polyshares publishes no headcount minimum. Years of records and linked context usually matter more than headcount alone. ### Do I lose ownership of my data when I license it? You keep ownership under a license. The buyer gets the right to use a copy for the purposes the license allows, such as training or evaluation. An exclusive license can stop you licensing the same data to anyone else, so read the exclusivity and term clauses before signing. ### Will the buyer get access to my systems? The buyer should not get access to your live systems. Legal guidance from Frankfurt Kurnit Klein and Selz warns against giving buyer tools direct access to company systems. Export a fixed copy of only the approved records and send it by a secure transfer method. That keeps excluded material, credentials and current work out of the buyer's hands. ### Can I sell data that contains personal information? You can only include personal information if the privacy notice in force when you collected it, and the applicable law, allow the disclosure. Most sellers de-identify personal data or exclude it. UK GDPR, EU GDPR and the CCPA each set their own standard, so take independent advice for your case. ### How long until I am paid? Published timings range from about 2 weeks to 2 months. Mercor says it typically pays 2 to 4 weeks after signing. micro1 takes up to 30 days to accept the data and pays within 30 days of invoice. Polyshares pays net 30 to 60 days after approval. ### Do I need a lawyer to sell company data? You should have your own lawyer review any license before you sign. The contract decides what the buyer can do with your data and anything built from it, who carries the risk if a problem appears later, and whether you can license the same data again. A broker can prepare you but cannot replace legal advice. ## Sources - [Company data partnerships](https://www.micro1.ai/data-partnerships), micro1, accessed November 2026 - [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), Don't Do It Yourself, accessed November 2026 - [How much is company data worth to AI? (2026 prices and real deals)](https://ddiy.co/how-much-is-company-data-worth-to-ai/), Don't Do It Yourself, accessed November 2026 - [AI labs are buying data from dead startups to train their models](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models), MIT Technology Review, accessed November 2026 - [micro1 tries to beat Google's bid to buy bankrupt Spirit Airlines' data for $12.5 million](https://fortune.com/2026/09/14/micro1-tries-to-beat-googles-bid-to-buy-bankrupt-spirit-airlines-data-for-12-5-million-highlighting-value-of-data-to-ai-companies/), Fortune, accessed November 2026 - [Spirit's privacy ombudsman backs Google's $10m data deal](https://thenextweb.com/news/google-spirit-airlines-data-ombudsman-approval), The Next Web, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 - [Shopping for AI training data: five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody, accessed November 2026 - [Is selling datasets way harder than building them, or is it just me?](https://discuss.huggingface.co/t/is-selling-datasets-way-harder-than-building-them-or-is-it-just-me/177440), Hugging Face forum, accessed November 2026 - [AI training data marketplaces](https://datavendor.ai/blog/ai-training-data-marketplaces), DataVendor, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How much is your data worth to AI companies? URL: https://sellmydatatoai.net/guides/how-much-is-your-data-worth-to-ai Reviewed: 2026-11-08 Company data is worth between $10,000 and $4 million to AI companies on buyers' published terms, and most published ranges sit between $100,000 and $2 million. Closed sales are often lower: AIData.Broker puts the median public sale at $31,600 against a median published offer of $280,000. Years of linked records, clean rights and exclusivity push the price up. ## Key takeaways - Published buyer ranges run from $10,000 to $4 million per company, according to Don't Do It Yourself's October 2026 comparison. - Real sales are often well below headline offers. AIData.Broker's median public sale is $31,600, against a median published offer of $280,000. - Polyshares has published four exclusive deals between $110,000 and $450,000. - An exclusive license fetches about 2.5 to 5 times a non-exclusive one, according to AIData.Broker. - Personal data you cannot clear, and unclear rights, lower the price more than low volume does. The honest answer is a range, and a wide one. Buyers publish ceilings in the millions, while most closed sales from small and closing companies land in the tens of thousands. This guide sets out what buyers publish, what deals have actually closed and what moves a price up or down, with a source for every figure. ## How much do AI companies pay for company data? AI companies pay between $10,000 and $4 million per company in their published terms, with most ranges between $100,000 and $2 million, according to Don't Do It Yourself's comparison of buyer sites, checked October 26, 2026. The bottom of that range is Scale AI's illustrative $10,000 and the top is Endless Data's up to $4 million. | Buyer | Published payout per company | Source | | --- | --- | --- | | micro1 | $100,000 to $3 million or more, in three tiers | micro1 data partnerships page, November 2026 | | Mercor | $20,000 to $1.5 million, by industry | Mercor listings via Don't Do It Yourself, October 2026 | | Polyshares | $100,000 to $2 million typical | Via Don't Do It Yourself, October 2026 | | Endless Data | Up to $4 million | Via Don't Do It Yourself, October 2026 | | Handshake AI | $100,000 to $4 million paid to date | Via Don't Do It Yourself, October 2026 | | Labelbox Prism | About $50,000 to $1 million or more | Via Don't Do It Yourself, October 2026 | | Scale AI | $10,000 to $1 million or more (illustrative) | Via Don't Do It Yourself, October 2026 | | SimpleClosure AssetHub | Typically $10,000 to $100,000 | Chief executive Dori Yona to Fortune, September 2026 | Treat these as ceilings rather than quotes. A published range tells you what a buyer is willing to pay for its best-fitting records. It says nothing about what it will pay for yours until it has seen the manifest and a sample. ## What have real company data sales fetched? Real company data sales have fetched between about $25,000 and $10 million in reported deals from 2026. The figures below are the ones where the price has been published by the buyer or reported by a named news outlet. | Seller | Price | Data | Source | | --- | --- | --- | --- | | Roofing company | $450,000 | Company records, perpetual exclusive license | Polyshares, via Don't Do It Yourself | | General contractor | $300,000 | Company records, perpetual exclusive license | Polyshares, via Don't Do It Yourself | | Solar company | $275,000 | Company records, perpetual exclusive license | Polyshares, via Don't Do It Yourself | | Global marketing agency | $110,000 | Company records, perpetual exclusive license | Polyshares, via Don't Do It Yourself | | Shut-down crypto startup | About $25,000, plus about $15,000 pending | Codebase, then workspace data | MIT Technology Review, October 2026 | | Cielo24 | Hundreds of thousands of dollars | 13 years of workplace data | Former chief executive to Fortune, September 2026 | | Spirit Airlines (in bankruptcy) | $10 million winning bid, $12.5 million competing proposal | About 100 million emails and 500 million Teams messages | MIT Technology Review and Fortune, 2026 | Turing's chief executive told Fortune it had bought five to ten failed-startup codebases at an average of tens of thousands of dollars each. Fortune also stated that there is no reliable market average for a failed company's data, and that $12.5 million was the largest figure it found. AIData.Broker's figures, quoted by Don't Do It Yourself, sum up the gap: a median of $280,000 across buyers' published offers, and a median of $31,600 across public sales. You can see more examples on our [AI data deals tracker](https://sellmydatatoai.net/ai-data-deals). ## How much is data worth in my industry? Data value by industry is clearest in Mercor's listings, which publish a company payout range and a size band for each sector. The ranges below come from Mercor's listings as compiled by Don't Do It Yourself, checked October 26, 2026. | Industry | Company size | Company earns | | --- | --- | --- | | Manufacturing and industrial | 20 to 2,000 staff | $100,000 to $1.5 million | | Law firm | 5 to 300 attorneys and staff | $50,000 to $1.2 million | | Construction | 50 to 2,000 staff | $115,000 to $1 million | | Design and web agency | 10 to 200 staff | $20,000 to $890,000 | | Accounting and tax firm | 5 to 200 staff | $25,000 to $865,000 | | Medical practice | 10 or more clinicians or 2,500 or more patients | $80,000 to $820,000 | | Cybersecurity | 20 to 2,000 staff | $20,000 to $750,000 | | Software company | 15 to 500 staff | $50,000 to $500,000 | | Publisher or education company | 10 to 500 staff | $35,000 to $370,000 | | Consulting and advisory | 10 to 500 staff | $150,000 to $300,000 | | Call center or BPO | 50 to 5,000 agents | $20,000 to $250,000 | Mercor's listings show a separate referral fee for the person who introduces the company. That figure is not the company payout, so do not confuse the two. Our industry pages for [law firms](https://sellmydatatoai.net/sell-your-data/law-firms), [construction companies](https://sellmydatatoai.net/sell-your-data/construction-companies) and [software companies](https://sellmydatatoai.net/sell-your-data/software-companies) explain what buyers ask for in each sector. ## What makes company data more valuable to AI buyers? Company data becomes more valuable when it shows complete pieces of work over many years, with the context linked together and the rights settled. The factors below are the ones buyers name in their own terms. - Years of history. AIData.Broker and micro1 both say more years of records raise the price - Linked records. A ticket joined to the chat thread, the code change and the outcome is worth more than each item alone, according to AIData.Broker - Rare, proprietary knowledge. micro1 reserves its top tier, $3 million or more, for highly unique data - Size, quality and workflow complexity, which micro1 lists as price drivers - Settled rights. AIData.Broker and Mercor both say clear rights raise the price and unclear rights lower it - Situations with judgment in them. MIT Technology Review reported that Surge AI looks for archives showing disagreements and judgment calls to build reinforcement learning environments Sunset, a rival to SimpleClosure, values assets by headcount, years in operation, data volume and how the team used its tools, according to MIT Technology Review. That is a useful checklist for your own first estimate. ## What lowers the price of company data? The price of company data falls when a buyer has to carry risk or cleaning cost. Personal data you cannot clear lowers the price, according to AIData.Broker, and so do large exclusions, a hand-picked sample and a manifest that does not match the delivery. Buyers also discount data they cannot use straight away. Mostly non-English records fall outside micro1's stated preference for data that is primarily in English. Public material, contact lists and data you obtained from someone else rarely add value, because labs can usually get them elsewhere. ## Does the size of my company affect the price? The size of your company affects the price mainly because several buyers set minimums. micro1 looks for established companies with 30 or more full-time employees. Handshake AI asks for 20 or more staff and 3 or more years in business, Scale AI prefers 40 or more staff and Endless Data looks for about 30 or more staff with several years of records, according to Don't Do It Yourself. Polyshares publishes no headcount minimum. Headcount is a rough proxy for volume and variety. A 15-person firm with ten years of well-linked records can be worth more than a 200-person firm with two years of scattered files. ## How much more is an exclusive license worth? An exclusive license is worth about 2.5 to 5 times a non-exclusive deal, according to AIData.Broker as quoted by Don't Do It Yourself. All four of Polyshares' published deals were perpetual exclusive licenses. The multiple only makes sense if you are giving something up. Troveo, a licensing agent, says its datasets sell an average of four times, so a non-exclusive archive licensed repeatedly can earn as much as a single exclusive sale. Read [exclusive vs non-exclusive data licenses](https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses) before you choose. ## How do AI buyers decide what to offer? AI buyers decide what to offer by estimating how many useful training tasks a dataset will yield. MIT Technology Review reported in October 2026 that buyers quote very different prices because each values data against its own research thesis, and that no company has published tests comparing training on real-world company data with alternatives. That is why two buyers can look at the same archive and offer amounts several times apart. It is also why competing offers matter more in this market than in most. MIT Technology Review also reported that founders often have little visibility into who buys their data or whether an offer is fair. ## How is data priced per unit? Some data is priced per unit rather than per company. Video and audio archives are often priced per minute or per hour, and packaged datasets are often sold as annual licenses. | Type | Reported price | Source | | --- | --- | --- | | Video, direct licensing to OpenAI | $1 to $4 per minute | Versos (no primary citation given) | | Video, Adobe partnerships | About $3 per minute | Versos (no primary citation given) | | Agent trajectory data | $1 million to $2 million per year | micro1 data partnerships page | | Research license | $5,000 to $15,000 | Phyllo, reported market range | | Commercial license | $25,000 to $500,000 or more | Phyllo, reported market range | | Exclusive or custom license | $1 million or more | Phyllo, reported market range | Phyllo notes that research licenses usually do not permit production training, so a low research price is not a fair comparison with a commercial offer. Large platform deals sit in a different bracket again: Reuters reported in February 2024 that Reddit's license with Google was worth about $60 million a year, and Shutterstock reported $104 million from AI licensing in 2023, according to Business Insider. ## How do you estimate your own data's value? You estimate your own data's value by matching your records to the closest published buyer listing, then adjusting for history, rights and exclusivity. These steps give you a defensible starting range before you talk to anyone. 1. **Find the closest listing**: Match your sector and size to Mercor's industry ranges and micro1's tiers. Note the size band, because ranges assume you fit it. 2. **Count what you hold**: Years of records, number of systems, record counts and whether items link to each other. 3. **Subtract what you cannot license**: Remove client work under restrictive contracts, HR files and personal data you cannot clear. Estimate what share of the archive is left. 4. **Decide on exclusivity**: Price a non-exclusive license first, then treat exclusivity as a separate item with its own price. 5. **Test the range with buyers**: Only competing offers on the same records and terms tell you the market price. Our [data value estimator](https://sellmydatatoai.net/data-valuation#estimator) does the first three steps for you. For a written view of likely buyers, a realistic range and the rights issues to clear, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### What is the average price AI companies pay for company data? There is no reliable average. Fortune said in September 2026 that there is no reliable market average for a failed company's data. AIData.Broker, quoted by Don't Do It Yourself, puts the median of buyers' published offers at $280,000 and the median of public sales at $31,600, which shows how far real sales fall below headline offers. ### Is my small company's data worth anything? A small company's data can be worth something if it shows real work over several years. Mercor's listings start at 5 staff for law firms and accounting firms, and SimpleClosure says closing startups typically receive $10,000 to $100,000. Contact lists and public material on their own are rarely worth selling. ### Why do published payouts differ so much from real sales? Published payouts describe what a buyer will pay for its best-fitting records. Real sales reflect exclusions, unclear rights, smaller archives and weaker negotiating positions, especially for companies that are closing. MIT Technology Review reported that buyers value data against their own research theses, so offers for the same archive can vary widely. ### Do I get paid once or over time? Most company data deals pay a one-off fee. Polyshares and micro1 pay a single sum after approval. Labelbox Prism pays an upfront fee at signing plus a revenue share on later use, and Troveo pays a share each time your data is licensed again. Handshake AI sets milestone payments in each agreement. ### Does exclusivity always mean more money? Exclusivity usually means a higher single payment, about 2.5 to 5 times a non-exclusive deal according to AIData.Broker. It does not always mean more money in total. Troveo says its datasets sell an average of four times, so repeated non-exclusive licenses can match or beat one exclusive sale. ## Sources - [How much is company data worth to AI? (2026 prices and real deals)](https://ddiy.co/how-much-is-company-data-worth-to-ai/), Don't Do It Yourself, accessed November 2026 - [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), Don't Do It Yourself, accessed November 2026 - [Company data partnerships](https://www.micro1.ai/data-partnerships), micro1, accessed November 2026 - [AI labs are buying data from dead startups to train their models](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models), MIT Technology Review, accessed November 2026 - [micro1 tries to beat Google's bid to buy bankrupt Spirit Airlines' data for $12.5 million](https://fortune.com/2026/09/14/micro1-tries-to-beat-googles-bid-to-buy-bankrupt-spirit-airlines-data-for-12-5-million-highlighting-value-of-data-to-ai-companies/), Fortune, accessed November 2026 - [How to sell video content as AI training data](https://www.versos.ai/blog/sell-video-content-ai-training-data), Versos, accessed November 2026 - [Best public data providers for AI training and grounding](https://www.getphyllo.com/post/best-public-data-providers-ai-training-grounding), Phyllo, accessed November 2026 - [Reddit in AI content licensing deal with Google, sources say (Reuters report)](https://www.niemanlab.org/reading/reddit-in-ai-content-licensing-deal-with-google-sources-say/), Nieman Lab, accessed November 2026 - [Shutterstock earned over $100 million in revenue thanks in part to its AI-powered image generator tool](https://www.aol.com/shutterstock-earned-over-100-million-170917396.html), Business Insider, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # What data do AI companies buy? URL: https://sellmydatatoai.net/guides/what-data-do-ai-companies-buy Reviewed: 2026-11-08 AI companies buy private records that show how real work gets done: workplace email and chat, support tickets, standard operating procedures, CRM records, source code with its change history, completed project files and agent trajectories. They rarely buy public data, contact lists or scraped material. Years of history, linked context and settled rights make records worth more. ## Key takeaways - Buyers want records of real work with the context attached, not isolated documents. - micro1 lists SOPs, process documentation, knowledge bases, workflows, templates, CRM records and support tickets. - DataVendor says labs pay more for human-written code for complex products, with pull requests, issues, runbooks and incident notes. - Public text, contact lists and data you obtained from someone else rarely sell. - HR files, customer profiles and payment records are usually excluded even from large deals. AI labs trained their first large models mostly on public text. The data they are now paying for is private: the records companies create while doing their work. This guide lists the record types buyers name in their own terms, explains why they want them and covers what rarely sells. ## Why do AI companies want private company data? AI companies want private company data because public and synthetic data no longer cover what they need. MIT Technology Review reported in October 2026 that demand for company archives appeared suddenly as labs hit the limits of public and synthetic datasets. Fortune cited a 2025 Epoch AI report projecting that public human-generated text could be fully used before 2030 on current trends. Labs are also training models to carry out work, not only to write about it. DataVendor notes that post-training now creates stronger demand for evaluations, environments and real workflow traces. Company records are where those workflows live. The shift shows in deal volumes. SimpleClosure told MIT Technology Review that 47% of its wind-down deals now include code or data sales, a practice its chief executive described as once rare. Fortune reported that SimpleClosure's number of AI buyers grew ninefold compared with 2025. ## Which types of company data are in demand? The types of company data in demand are the ones buyers name on their own sites. The table below lists each record type, what makes it useful and who has said they want it. | Record type | What makes it useful | Who names it | | --- | --- | --- | | [Workplace communications](https://sellmydatatoai.net/buy-ai-training-data/workplace-communications-data) | Email, Slack and Teams threads that show decisions, disagreements and hand-offs | Google's Spirit Airlines bid, MIT Technology Review on Surge AI | | SOPs and process documentation | Step-by-step records of how work is done | micro1 | | Internal knowledge bases and templates | Expert know-how written down | micro1 | | [Support tickets](https://sellmydatatoai.net/buy-ai-training-data/support-ticket-data) | Problem, conversation and resolution in one record | micro1 | | CRM records | Sales activity linked to outcomes | micro1 | | [Source code with history](https://sellmydatatoai.net/buy-ai-training-data/source-code-datasets) | Human-written code plus pull requests, issues, runbooks and incident notes | DataVendor, Mercor | | [Agent trajectories](https://sellmydatatoai.net/buy-ai-training-data/agent-trajectory-data) | Recorded steps an AI agent or person took to complete a task | micro1, at $1 million to $2 million per year | | [Evaluation datasets](https://sellmydatatoai.net/buy-ai-training-data/evaluation-datasets) | Tasks, expected answers and scoring rules | DataVendor | | Completed project files | Drawings, photos, workpapers and deliverables from finished jobs | Mercor industry listings | | Recorded customer conversations | Real calls between agents and customers | Mercor's call center listing | ## What do AI companies buy from specific industries? AI companies buy the core working records of each industry. Mercor's listings, compiled by Don't Do It Yourself in October 2026, name the records it licenses sector by sector. | Industry | Records named in Mercor's listing | | --- | --- | | Law firm | Complete files from closed matters and completed deals | | Construction company | Project records, drawings and site photos from completed projects | | Manufacturer | Production, maintenance and inspection records | | Medical practice | De-identified patient records with years of physician notes | | Accounting and tax firm | Tax returns, workpapers and bookkeeping records | | Design and web agency | Design systems, shipped websites and related code | | Publisher or education company | Journals, textbooks, courses and video the company holds rights to | | Software company | Private code and its history | | Cybersecurity company | Alerts, investigations and case write-ups | | Consulting firm | Client deliverables, models and drafts | | Call center | Recorded customer conversations | Several of these contain client material. A law firm's closed matter files or a consulting firm's deliverables can only be licensed where client contracts and professional rules allow it, which is why rights checks come before pricing. Our pages on selling data by industry, starting with [law firms](https://sellmydatatoai.net/sell-your-data/law-firms), cover the specific restrictions. ## What makes company records useful for training? Company records are useful for training when a model can follow a piece of work from start to finish. A single document teaches little. A support ticket linked to the chat where the fix was discussed, the code change that shipped it and the customer's reply teaches a lot. - Linked context. AIData.Broker, quoted by Don't Do It Yourself, says linked records raise the price - Years of history. micro1 asks for substantial historical data from companies with mature operations - Judgment calls. MIT Technology Review reported that Surge AI looks for archives showing varied situations, including disagreements, to build reinforcement learning environments - Human authorship. DataVendor says labs pay more for code written mostly by people - Language. micro1 says data should be primarily in English - Clean rights. Buyers discount anything they might have to remove later DataVendor explains why engineering history matters: public repositories rarely preserve the full operational record of how a system was run and changed. Code-adjacent material such as pull requests, issue tickets, runbooks and incident notes explains why the code changed. Our guide on [selling source code to AI companies](https://sellmydatatoai.net/guides/sell-source-code-to-ai-companies) covers this in detail. ## Do AI companies buy evaluation data and agent traces? AI companies do buy evaluation data and agent traces, and these are among the most specialized items on the market. An evaluation dataset is a set of tasks with expected answers and scoring rules, used to measure how well a model performs. DataVendor notes that evaluations command more attention when independent reviewers can reproduce the scores. Agent trajectories record each step taken to complete a task: the clicks, tool calls, observations and result. micro1 lists agent trajectory licenses at $1 million to $2 million per year, subject to its approval. DataVendor adds that replayable environments, the software or sandbox needed to run a task again, let buyers generate new trajectories and check results against verified outcomes. Most operating companies do not hold these as finished products. They can sometimes be built from existing records, for example by turning resolved support tickets into tasks with known correct answers. That work is usually done with the buyer, not before. ## Do AI companies buy images, video and audio? AI companies buy images, video and audio, usually priced per unit rather than per company. Versos reports that OpenAI licenses video directly at $1 to $4 per minute and that Adobe pays about $3 per minute, though Versos gives no primary citation for either figure. Media archives carry layers of rights. Phyllo points out that video can contain music, images, performances and trademarks, each needing its own clearance. Shutterstock reported $104 million from AI licensing in 2023, according to Business Insider, which shows the scale when a library's rights are already settled. Mercor's listings also include construction site photos and publishers' video that the company holds rights to. ## How should you package records for a buyer? You should package records for a buyer as defined datasets, each with a description, counts, dates and exclusions. Buyers price what they can understand quickly, so a clear package often earns more than a larger, undocumented one. - Split the archive by system and record type, such as support tickets from 2018 to 2025 or closed matter files from 2015 to 2024 - Give record counts, date ranges, languages and file formats for each dataset - List what was excluded and why - Describe how records link together, for example ticket IDs that appear in chat threads and commits - Prepare a random sample drawn across the whole period, redacted and labeled as random Nixon Peabody's October 2026 alert makes the same point from the buyer's side: a defensible inventory of systems, record types, time periods, jurisdictions and individuals represented is the first thing a careful buyer asks for. ## What do AI product companies sell? AI product companies sell the data their own products generate. That includes prompt and response logs, agent run traces, user feedback and ratings, product event data and evaluation sets. Mercor's listings put AI product companies with 5 to 200 staff at $20,000 to $480,000, according to Don't Do It Yourself. Much of this data was created by your customers, not by you. Check your terms of service and privacy notice before treating it as yours to license, and see our guide on [UK GDPR, EU GDPR and CCPA](https://sellmydatatoai.net/guides/gdpr-and-selling-data-for-ai-training) for the personal data rules. ## What data do AI companies not buy? AI companies rarely buy data they could get elsewhere or data that carries more risk than value. These categories seldom sell on their own: - Public data, including your website, published reports and social media posts - Contact lists and other personal data with no work content attached - Data you bought, downloaded or scraped from someone else, because you do not hold the rights to license it - Generic text, which DataVendor says carries limited value unless the seller adds clean rights, filtering or rare coverage - Every private file you hold. DataVendor warns sellers against treating every private file as inventory ## Which records are usually left out of deals? Records usually left out of deals are those holding sensitive personal data with little training value. MIT Technology Review reported that suppliers employ data scientists to scrub personal information and exclude HR records. The Spirit Airlines sale shows how far exclusions go even in a $10 million deal. Fortune reported that customer profiles and lists were excluded and passenger data in included systems had to be de-identified. The Next Web reported that booking and transaction records, refunds, inflight purchases and WiFi sales were kept out entirely. > **Before you count your records** Subtract what you cannot license first. A buyer prices the archive that is left after exclusions, not the one on your servers. ## How do you find out if your data will sell? You find out if your data will sell by matching it to a buyer's published listing and checking your rights to license it. If your sector appears in Mercor's listings or your records match micro1's list, there is a buyer for that type of data. Whether there is a buyer for your copy depends on history, volume and rights. Read [how much your data is worth to AI companies](https://sellmydatatoai.net/guides/how-much-is-your-data-worth-to-ai) for published ranges. If you want an honest view of whether buyers will be interested, including when they will not, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### Do AI companies buy emails and Slack messages? AI companies do buy workplace email and chat. Google's winning bid for Spirit Airlines' data covered about 100 million emails and 500 million Microsoft Teams messages, according to MIT Technology Review. Check vendor terms first: Frankfurt Kurnit Klein and Selz notes that Slack's API terms bar third-party apps from using API data to train large language models. ### Will AI companies buy my customer list? AI companies rarely buy customer lists. A list of names and contact details has little training value and high privacy risk. Even in the Spirit Airlines sale, Fortune reported that customer profiles and lists were excluded. Buyers want records of work, such as support conversations, with personal details removed. ### Is old data still valuable? Old data is often more valuable, not less. Buyers including micro1 and AIData.Broker say more years of records raise the price, because long histories show how decisions and processes change over time. The records must still be readable, complete and covered by rights you can license. ### Do AI companies want non-English data? Some do, but the largest published programs favor English. micro1 says data should be primarily in English and that demand is strongest in the US, the UK and Canada. Multilingual data has its own buyers, so non-English records are worth assessing separately rather than writing off. ### Can I sell data my software product collected from users? You can only sell user-generated data if your terms of service, privacy notice and the law allow it. The privacy notice in force when the data was collected sets the limit, and the US Federal Trade Commission has warned against quietly changing notices to add AI training. Take independent advice before including it. ## Sources - [Company data partnerships](https://www.micro1.ai/data-partnerships), micro1, accessed November 2026 - [AI training data marketplaces](https://datavendor.ai/blog/ai-training-data-marketplaces), DataVendor, accessed November 2026 - [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), Don't Do It Yourself, accessed November 2026 - [How much is company data worth to AI? (2026 prices and real deals)](https://ddiy.co/how-much-is-company-data-worth-to-ai/), Don't Do It Yourself, accessed November 2026 - [AI labs are buying data from dead startups to train their models](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models), MIT Technology Review, accessed November 2026 - [micro1 tries to beat Google's bid to buy bankrupt Spirit Airlines' data for $12.5 million](https://fortune.com/2026/09/14/micro1-tries-to-beat-googles-bid-to-buy-bankrupt-spirit-airlines-data-for-12-5-million-highlighting-value-of-data-to-ai-companies/), Fortune, accessed November 2026 - [Spirit's privacy ombudsman backs Google's $10m data deal](https://thenextweb.com/news/google-spirit-airlines-data-ombudsman-approval), The Next Web, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 - [Shopping for AI training data: five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody, accessed November 2026 - [How to sell video content as AI training data](https://www.versos.ai/blog/sell-video-content-ai-training-data), Versos, accessed November 2026 - [Best public data providers for AI training and grounding](https://www.getphyllo.com/post/best-public-data-providers-ai-training-grounding), Phyllo, accessed November 2026 - [Shutterstock earned over $100 million in revenue thanks in part to its AI-powered image generator tool](https://www.aol.com/shutterstock-earned-over-100-million-170917396.html), Business Insider, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # AI data licensing explained: license, sale or revenue share URL: https://sellmydatatoai.net/guides/ai-data-licensing-explained Reviewed: 2026-11-08 AI data licensing means granting an AI company the right to use a copy of your data for agreed purposes, such as training or evaluation, while you keep ownership. Most deals are one-off license fees, exclusive or not. Some pay a revenue share each time data is resold. Outright sales of the data itself are rare outside bankruptcy and closures. ## Key takeaways - A license grants use rights. You keep ownership unless the contract assigns it. - The main structures are a one-off fee, an upfront fee plus revenue share, a pure revenue share and milestone payments. - Polyshares' four published deals were all perpetual exclusive licenses. Endless Data licenses non-exclusively to several labs. - The rights that matter most are permitted uses, rights in derived assets such as fine-tuned models, exclusivity, term and deletion. - Once data trains a model it cannot practically be removed, so the end of a license rarely undoes the training. When companies say they sold their data to an AI company, they usually mean they licensed it. The difference decides what you keep, what the buyer can do and whether you can make money from the same records again. This guide explains each structure with examples from buyers' published terms. ## What is AI data licensing? AI data licensing is a contract in which you give an AI company permission to use a copy of your data for defined purposes, in return for payment. The purposes are typically training, fine-tuning or evaluating models. You keep ownership of the data and anything the license does not grant. A [data license](https://sellmydatatoai.net/glossary/data-license) is the same legal tool used for software, music and photographs. What makes AI licenses different is that the buyer's use creates new assets, such as trained models, and the data cannot be taken back out of them afterward. The license is also where the price is really set. Two offers with the same headline figure can be worth very different amounts once you compare the records covered, the permitted uses, the exclusivity and the term. A lower fee for a narrow, time-limited, non-exclusive license can beat a higher fee for a perpetual exclusive one. ## Is selling data to AI the same as licensing it? Selling data to AI is usually not a sale in the legal sense. Most company deals are licenses: the buyer gets rights to use a copy, and you keep the original and the ownership. Outright sales do happen, mainly when a company is closing. The Spirit Airlines data is being sold as an asset in bankruptcy, with Google's $10 million bid subject to court approval, according to MIT Technology Review and The Next Web. Even there, the deal carried conditions: The Next Web reported that the buyer must publicly commit not to re-identify the data and may pass it on only to recipients who make the same commitment. Our guide to [selling data from a closed company](https://sellmydatatoai.net/guides/selling-data-from-a-closed-company) covers those cases. ## How are AI data deals structured? AI data deals are structured in five main ways, and each published buyer tends to favor one. The table shows the structures, how you are paid and a buyer that uses each one, taken from published terms compiled by Don't Do It Yourself in October 2026 unless stated. | Structure | How you are paid | Example | Main trade-off | | --- | --- | --- | --- | | One-off non-exclusive license | Fixed sum, paid after transfer | Endless Data, which licenses to several labs | You can license the same data again, but each fee is lower | | Perpetual exclusive license | Fixed sum after approval, net 30 to 60 days | Polyshares, exclusive by default | Higher fee, but you cannot license the data to anyone else, ever | | Upfront fee plus revenue share | Fee at signing, plus a share of later use | Labelbox Prism | Ongoing income depends on the buyer's sales | | Revenue share through an agent | A share each time data is licensed | Troveo, which says its datasets sell an average of four times | No upfront certainty, but repeated income | | Milestone payments | Staged payments set in each agreement | Handshake AI | Payment tied to delivery and acceptance stages | micro1, checked November 2026, pays in full within 30 days of acceptance after a review period of up to 30 days. Its contracting stage includes limits on reselling data before personal information is anonymized, which is a sensible clause to look for from any buyer. ## What rights does an AI data license grant? An AI data license grants only the rights written into it, so the scope clauses decide the deal. Nixon Peabody's October 2026 alert lists the questions a license must answer: - Which models, systems and affiliates may use the data - Whether use covers training only, or also testing, product development and commercial deployment - Whether the data may be combined with other datasets or shared with subprocessors or research partners - Who owns or may use derived artifacts, such as cleaned datasets, annotations, embeddings, evaluation sets, fine-tuned models, weights and outputs - Whether the buyer must stop outputs reproducing confidential or personal information - Retention, deletion, security, audit and incident response obligations Derived assets are the point sellers most often miss. A license that allows training also lets the buyer keep the trained model, unless it says otherwise. If you care about a competitor benefiting, restrict the field of use rather than relying on deletion. ## What happens when an AI data license ends? When an AI data license ends, the buyer usually has to stop using and delete its copies of your raw data. Models already trained on that data normally continue to exist. Frankfurt Kurnit Klein and Selz states that once data is used to train a model, it cannot practically be pulled back out. Nixon Peabody adds that deleting source files may not remove the data's influence from a trained model, and recommends saying in the contract whether deletion, retraining or model retirement applies if rights expire or prohibited material surfaces. Phyllo lists withdrawal terms, including what happens to models already trained, as one of six things a training data license should cover. > **Price the license as if it were permanent** Because training cannot be undone, a five-year license and a perpetual license often have similar practical effects on models trained in year one. Negotiate the price on that basis. ## How long do AI data licenses last? AI data licenses last anywhere from a single training run to forever. Polyshares' published deals were all perpetual, which means no end date. Troveo describes its exclusivity as time-limited. Annual licenses are common for packaged datasets, and Phyllo reports professional annual licenses at $25,000 to $50,000 and enterprise annual licenses at $50,000 to $500,000 as market ranges. A [perpetual license](https://sellmydatatoai.net/glossary/perpetual-license) is not automatically bad. It suits archives you will never add to, such as records from a closed business line. For data you keep creating, a fixed term with an option to renew lets you reprice later. ## What is the difference between a research and a commercial license? The difference between a research and a commercial license is whether the buyer may use your data to build products it sells. Phyllo reports research licenses at $5,000 to $15,000 and commercial licenses at $25,000 to $500,000 or more, and notes that research licenses usually do not permit production training. Academic and lab research teams sometimes ask for a research license first. If you accept one, define research narrowly, ban use in commercial models and require deletion at the end of the project. Otherwise a cheap research license can become a commercial one in practice. ## What should an AI training data license cover? An AI training data license should cover the rights you are granting and the evidence that you can grant them. Phyllo lists six things a training data license should cover, which work as a checklist for sellers too: 1. Rights documentation for each asset 2. Explicit permission for AI training 3. A documented chain of acquisition 4. Clearance for each layer of content, since video can contain music, images, performances and trademarks 5. Evidence that rights holders are paid where they need to be 6. Withdrawal terms, including what happens to models already trained Buyers will expect you to back the first three with documents. A data manifest, the contracts that cover each dataset and a note of your exclusions usually satisfy them. ## Who carries the risk under an AI data license? The risk under an AI data license sits wherever the warranties and indemnities put it. Nixon Peabody lists the warranties a buyer will ask for: authority, privacy notices, contractual restrictions, IP rights, regulatory compliance, litigation holds, security and completeness of disclosed exclusions. As the seller, you want those warranties limited by your knowledge and by the exclusions you disclosed, with a cap on liability. Nixon Peabody also mentions holdbacks, where part of the price is kept back for a period, and insurance. Agree these before you share the full data, not after. ## When does a revenue share make sense? A revenue share makes sense when your data can be licensed to many buyers and you can wait for the income. A [revenue share](https://sellmydatatoai.net/glossary/revenue-share) pays you a percentage of what the licensing party earns from your data, rather than a single fee. Troveo says its datasets sell an average of four times, which is the case for revenue sharing in one number. The risks are the opposite of a fixed fee: you do not know the total, you depend on the agent's reporting and you need audit rights to check it. Labelbox Prism's model, an upfront fee plus a revenue share, splits the difference. ## Will the buyer say publicly that it used your data? The buyer may have to disclose that it used licensed data, depending on where its models are offered. Article 53 of the EU AI Act requires providers of general-purpose AI models to publish a sufficiently detailed summary of the content used for training, using a template the European Commission published in July 2025. Frankfurt Kurnit Klein and Selz also notes that California's AB 2013 requires generative AI developers to disclose their training data sources, including whether data was purchased or licensed. If confidentiality about the deal matters to you, agree in the license how you will be described. Our guide to [the EU AI Act and training data transparency](https://sellmydatatoai.net/guides/eu-ai-act-training-data-transparency) explains the template. ## Who signs the license? You sign the license directly with the buyer. When we broker a deal, Sell My Data To AI is not a party to the license and does not take custody of your raw data. Full data moves only after you sign. Before you sign, work through our [AI data licensing agreement checklist](https://sellmydatatoai.net/guides/ai-data-licensing-agreement-checklist) with your own lawyer. We are a broker, not a law firm. To compare structures for your own records, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### Do I still own my data after licensing it to an AI company? You still own your data after a license, unless the contract assigns ownership to the buyer. A license grants the right to use a copy for defined purposes. Read the grant clause carefully, because some contracts called licenses contain wording that transfers rights in practice, especially over derived assets and trained models. ### What is the difference between an exclusive and a non-exclusive license? An exclusive license stops you licensing the same data to anyone else for the agreed scope and term. A non-exclusive license lets you license it again. AIData.Broker, quoted by Don't Do It Yourself, says an exclusive deal fetches about 2.5 to 5 times a non-exclusive one. ### Can I license the same data to more than one AI company? You can license the same data to several AI companies if every license is non-exclusive. Endless Data licenses non-exclusively to several labs, and Troveo says its datasets sell an average of four times. Check that no earlier license contains an exclusivity or first-refusal clause before you sign another. ### Can the buyer resell my data? The buyer can only resell or sublicense your data if the license allows it. Nixon Peabody recommends stating whether data may be shared with affiliates, subprocessors or research partners. micro1's process includes limits on reselling data before personal information is anonymized. Ask for onward transfer to be prohibited or tightly defined. ### Is a revenue share better than a fixed fee? A revenue share is better when your data can be licensed many times and you can wait for income. A fixed fee is better when you want certainty or are closing the business. Troveo pays a share each time data resells, while Polyshares and micro1 pay one sum after approval. ## Sources - [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), Don't Do It Yourself, accessed November 2026 - [How much is company data worth to AI? (2026 prices and real deals)](https://ddiy.co/how-much-is-company-data-worth-to-ai/), Don't Do It Yourself, accessed November 2026 - [Company data partnerships](https://www.micro1.ai/data-partnerships), micro1, accessed November 2026 - [Shopping for AI training data: five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 - [Best public data providers for AI training and grounding](https://www.getphyllo.com/post/best-public-data-providers-ai-training-grounding), Phyllo, accessed November 2026 - [AI labs are buying data from dead startups to train their models](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models), MIT Technology Review, accessed November 2026 - [Spirit's privacy ombudsman backs Google's $10m data deal](https://thenextweb.com/news/google-spirit-airlines-data-ombudsman-approval), The Next Web, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # AI data licensing agreement checklist URL: https://sellmydatatoai.net/guides/ai-data-licensing-agreement-checklist Reviewed: 2026-11-08 An AI data licensing agreement should define exactly which records are licensed, what the buyer may do with them and with anything trained on them, how personal data is de-identified and certified, when you are paid, when raw copies are deleted, and how far your liability goes. Use this checklist with your own lawyer before you sign any license. ## Key takeaways - Attach the data manifest and the exclusions list to the contract as schedules, so the licensed data is defined precisely. - Deal with derived assets separately: cleaned data, annotations, embeddings, evaluation sets, fine-tuned models, weights and outputs. - Tie payment to written acceptance tests with a deadline. - Require de-identification certification before each transfer and a contractual ban on re-identification. - Limit your warranties to your knowledge and disclosed exclusions, and cap your liability. Most AI data licenses are drafted by the buyer. That is normal, but it means the first draft protects the buyer. This checklist sets out each clause a seller should look for, what a fair position looks like and why it matters, drawing on Nixon Peabody's October 2026 diligence alert and guidance from Frankfurt Kurnit Klein and Selz. > **Not legal advice** This checklist explains what to look for. It is not advice on your contract. Sell My Data To AI is a broker, not a law firm, and you should take independent legal advice before signing any license. ## What should an AI data licensing agreement include? An AI data licensing agreement should include at least the 14 clauses below. If a draft is missing one, ask why before negotiating anything else. | Clause | What to look for as the seller | | --- | --- | | Parties and authority | The right legal entities, and signatories who can bind them | | Licensed data | Defined by an attached manifest, not by a general description | | Exclusions | A schedule of what is not licensed, and what happens if it turns up | | Permitted uses | Training, evaluation or fine-tuning, stated precisely | | Prohibited uses | Re-identification, resale, competing products and any field you want to protect | | Derived assets | Who owns or may use annotations, embeddings, evaluation sets, models and outputs | | Exclusivity and term | Scope, length and what exclusivity actually covers | | Price and payment | Amount, triggers, deadlines and late payment terms | | Delivery and acceptance | Transfer method, acceptance tests and a deadline for acceptance | | Personal data | De-identification standard, method, provider and certification | | Security and deletion | Storage, access controls, retention period and deletion of raw copies | | Onward transfer | Limits on affiliates, subprocessors and model partners | | Warranties, indemnities and caps | Knowledge qualifiers, disclosed exceptions and a liability cap | | Termination | What ends, what survives and what happens to trained models | ## How should the licensed data be defined? The licensed data should be defined by a schedule that lists each dataset by system, record type, record count, date range and format. A vague definition such as all business records lets a buyer argue later that something was promised and not delivered. - Attach your [data manifest](https://sellmydatatoai.net/glossary/data-manifest) as a schedule and make it the definition of the licensed data - Attach a separate exclusions schedule, such as HR files, privileged material and named client matters - Record file hashes or version numbers so both sides can prove what was delivered - Agree in advance what happens if prohibited material is found after delivery, which Nixon Peabody recommends - State whether future records are included, or only the archive at a fixed date Our guide to [building a data manifest buyers will trust](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest) shows the fields to include. ## What uses should the license permit and prohibit? The license should permit only the uses you have priced and prohibit the uses you are worried about. Nixon Peabody suggests stating which models, systems and affiliates may use the data, and whether use covers training only or also testing, product development and commercial deployment. - Permitted: name the purposes, such as pre-training, fine-tuning or evaluation - Permitted: say whether commercial deployment of trained models is included - Prohibited: any attempt to re-identify individuals - Prohibited: resale or sublicensing of the raw data without your written consent - Prohibited: use to build a product that competes with you, if that matters to your business - Required: steps to stop model outputs reproducing confidential or personal information, which Nixon Peabody lists as a term to address ## Who owns what the buyer builds from your data? Who owns what the buyer builds from your data depends entirely on the derived assets clause. Nixon Peabody lists the assets to address separately: cleaned datasets, annotations and embeddings, evaluation sets, fine-tuned models, model weights and outputs. Most buyers will insist on owning trained models. That is usually reasonable. The items worth negotiating are cleaned and annotated copies of your data and evaluation sets built from it, because these can be resold or reused in ways that work like a second license of your records. ## How should exclusivity and term be written? Exclusivity and term should be written narrowly enough that you know exactly what you are giving up. An exclusive license fetches about 2.5 to 5 times a non-exclusive one, according to AIData.Broker, so make sure the price reflects it. - Define what exclusivity covers: all AI uses, a field of use, a type of model or a list of named competitors - Set a term in years, or confirm in writing that the license is perpetual - Check for first refusal or matching rights hidden elsewhere in the contract - Agree whether you may still use the data internally, including for your own AI tools Read [exclusive vs non-exclusive data licenses](https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses) for the options in between. ## How should payment and acceptance work? Payment and acceptance should work to a fixed timetable, so the buyer cannot hold payment open by delaying acceptance. Published terms show the range: micro1 takes up to 30 days to review and accept the data and then pays within 30 days of invoice, and Polyshares pays net 30 to 60 days after approval. - Written acceptance tests based on the manifest, such as record counts, date ranges and file integrity - A deadline for acceptance, after which the data is treated as accepted - A deadline for payment after acceptance or invoice - Milestones for staged deliveries, if the deal uses them - Interest or a right to suspend delivery if payment is late - For revenue shares, reporting dates and a right to audit the figures ## What should the license say about personal data? The license should say exactly how personal data is removed, who removes it and how that is proved. Nixon Peabody recommends that transaction documents set the standard and the process, including an independent de-identification provider informed by NIST guidance, re-identification risk testing against external data and certification before each transfer. - The de-identification standard, by reference to the law that applies, such as UK GDPR or the CCPA - Who chooses and pays the provider. Frankfurt Kurnit Klein and Selz recommends the seller selects or jointly selects it and approves the protocol - Certification before each transfer - A ban on re-identification, backed by controls, and the same ban passed to any recipient - Record-level exclusions for privileged or highly sensitive material Under the CCPA, data only counts as de-identified if the holder publicly commits not to re-identify it and binds recipients to the same, according to Nixon Peabody. See our guide to [de-identification for AI training data](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data). ## What should the license say about security and deletion? The license should say where your data will be stored, who can see it and when raw copies are deleted. Frankfurt Kurnit Klein and Selz advises against giving buyers direct access to your systems and recommends secure transfer methods. - A fixed export delivered by an agreed secure method, never live system access - Named storage locations and access limited to named teams - A retention period for raw copies and written confirmation of deletion - Incident notification within a set number of hours or days - A statement of what deletion means for models already trained, since Nixon Peabody notes deletion may not remove the data's influence ## Can the buyer pass your data to anyone else? The buyer can pass your data to someone else only if the license allows it, so the onward transfer clause needs to be specific. Nixon Peabody recommends stating whether data may be combined with other datasets or shared with subprocessors or research partners, and notes that restrictions must flow through the buyer's whole technical stack, including cloud and model vendors. - A list of permitted affiliates, or a ban on sharing outside the licensee entity - Subprocessors bound by terms at least as strict as the license - No transfer to model partners or research partners without your written consent - The buyer remains liable for anything its subprocessors do with the data Training-data companies such as micro1, Mercor and Scale AI supply AI labs, so onward supply is their business model. With them, the question is not whether data passes on but in what form, with which protections and under whose name. ## How far should your warranties and liability go? Your warranties and liability should go only as far as you can control. Buyers will ask for warranties on authority, privacy notices, contractual restrictions, IP rights, regulatory compliance, litigation holds, security and completeness of disclosed exclusions, according to Nixon Peabody. | Buyer asks for | A fair seller position | | --- | --- | | Unqualified warranty that no personal data is included | Warranty that the agreed de-identification method was applied, with the provider's certificate | | Warranty that you hold all rights in all records | Warranty given to your knowledge, subject to the exclusions schedule | | Uncapped [indemnity](https://sellmydatatoai.net/glossary/indemnity) | Indemnity capped at the fee received, or a multiple agreed in writing | | Large holdback for a long period | A smaller holdback released on a fixed date | | Cost of retraining models if a problem appears | Removal of affected records from future use, with retraining costs excluded or capped | Nixon Peabody notes that buyers may seek remedies covering the cost of removing data or retraining a model. That exposure can exceed the price of the license, so the cap matters more than almost any other term. ## What about confidentiality and publicity? Confidentiality and publicity clauses should say whether the buyer may name you, and how. Article 53 of the EU AI Act requires providers of general-purpose AI models to publish a summary of training content, and California's AB 2013 requires disclosure of training data sources, according to Frankfurt Kurnit Klein and Selz. You cannot stop a buyer meeting a legal disclosure duty, but you can agree how you are described, ban marketing use of your name and get notice before any announcement. The [mutual NDA](https://sellmydatatoai.net/glossary/mutual-nda) you signed for the sample stage should be replaced or extended by these terms. ## What should happen on termination? On termination, the buyer should stop using your raw data and certify deletion, and the contract should say clearly which rights survive. Trained models usually survive. Confidentiality, the re-identification ban, the security obligations and payment terms should survive too. Before you sign, read [AI data licensing explained](https://sellmydatatoai.net/guides/ai-data-licensing-explained) for how each deal structure affects these clauses. If you want help comparing offers clause by clause, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### Who usually drafts an AI data license? The buyer usually drafts an AI data license, because it signs many and has standard terms. That means the first draft favors the buyer. You can and should negotiate it. Ask your lawyer to focus first on scope, derived assets, warranties and the liability cap, since those carry the most money. ### Should the data manifest be part of the contract? The data manifest should be attached to the contract as a schedule and used as the definition of the licensed data. That makes acceptance testable and protects you from claims that more was promised. Attach an exclusions schedule too, listing what is deliberately left out. ### Can I cap my liability under a data license? You can usually negotiate a cap on liability, and you should. Nixon Peabody notes that buyers may seek remedies covering the cost of removing data or retraining a model, which can exceed the license fee. A common position is a cap at the fee received, though any cap is a matter for negotiation. ### What is a holdback in a data deal? A holdback is part of the price that the buyer keeps for an agreed period, to cover claims if a problem with the data appears. Nixon Peabody lists holdbacks alongside indemnities, caps and insurance as ways of allocating risk. Keep any holdback small and release it on a fixed date. ### Do I need a data processing agreement as well? You may need one if personal data is still included and the buyer processes it on your behalf. In most AI data licenses the buyer acts for its own purposes, so it is a separate controller, and a data sharing agreement is more typical. Your lawyer should decide which applies. ## Sources - [Shopping for AI training data: five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 - [Company data partnerships](https://www.micro1.ai/data-partnerships), micro1, accessed November 2026 - [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), Don't Do It Yourself, accessed November 2026 - [How much is company data worth to AI? (2026 prices and real deals)](https://ddiy.co/how-much-is-company-data-worth-to-ai/), Don't Do It Yourself, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Exclusive vs non-exclusive data licenses URL: https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses Reviewed: 2026-11-08 An exclusive data license lets only one buyer use your data for the agreed scope and term, while a non-exclusive license lets you license the same data again. Exclusive deals pay more upfront, about 2.5 to 5 times a non-exclusive one according to AIData.Broker, but repeat non-exclusive licenses can earn as much. Field-of-use and time-limited terms sit in between. ## Key takeaways - AIData.Broker puts an exclusive deal at about 2.5 to 5 times a non-exclusive one. - Polyshares and Labelbox Prism license exclusively by default. Endless Data licenses non-exclusively to several labs. - Troveo says its datasets sell an average of four times, which is the main argument for staying non-exclusive. - Exclusivity can be limited by field of use, model type, named competitors or a fixed number of years. - A perpetual exclusive license is effectively a sale of the data's AI value. Price it that way. Exclusivity is often the biggest single lever on price in an AI data deal, and the one sellers agree to most casually. This guide explains what each type of license means, what the published buyers offer and how to choose. ## What is the difference between exclusive and non-exclusive data licenses? The difference between exclusive and non-exclusive data licenses is whether you can license the same data to anyone else. Under an [exclusive license](https://sellmydatatoai.net/glossary/exclusive-license), only the licensee may use the data within the agreed scope and term, and you promise not to license it to others. Under a [non-exclusive license](https://sellmydatatoai.net/glossary/non-exclusive-license), you keep the right to license it again. | | Exclusive license | Non-exclusive license | | --- | --- | --- | | Who else can use the data | No one within the agreed scope, sometimes including you | Any other licensee you choose | | Typical price | About 2.5 to 5 times a non-exclusive deal (AIData.Broker) | Lower per deal, can be repeated | | Income pattern | One larger payment | Several smaller payments over time | | Published examples | Polyshares, Labelbox Prism (exclusive by default) | Endless Data, Shutterstock's AI licensing | | Main risk | Locks you out of later, better offers | Each buyer may pay less knowing others have the data | | Best for | Rare data with one strong buyer, or a one-off exit | Data many labs want, or a growing archive | ## How much more does an exclusive license pay? An exclusive license pays about 2.5 to 5 times a non-exclusive deal for the same data, according to AIData.Broker as quoted by Don't Do It Yourself in October 2026. On that ratio, data that would fetch $100,000 non-exclusively would fetch roughly $250,000 to $500,000 exclusively. Polyshares, which buys exclusive licenses by default, has published four deals: a roofing company at $450,000, a general contractor at $300,000, a solar company at $275,000 and a global marketing agency at $110,000. All four were perpetual exclusive licenses, according to Don't Do It Yourself. The multiple reflects what the buyer gains and what you give up. The buyer gets data its rivals cannot train on, and you lose every later license of the same records within the agreed scope. If the data is rare, that is worth a great deal to the buyer. If several companies hold similar records, exclusivity protects the buyer far less and the premium should be smaller. Our guide to [how much your data is worth to AI companies](https://sellmydatatoai.net/guides/how-much-is-your-data-worth-to-ai) has the full range of published payouts. ## Which buyers want exclusivity? The buyers that want exclusivity are mostly direct buyers and platforms that resell to labs at a premium. Published terms, compiled by Don't Do It Yourself in October 2026, show the split. | Buyer | Exclusivity | Payment model | | --- | --- | --- | | [Polyshares](https://sellmydatatoai.net/data-buyers/polyshares) | Exclusive by default, published deals perpetual | One-off fee, net 30 to 60 days | | Labelbox Prism | Exclusive by default | Upfront fee plus revenue share | | [Endless Data](https://sellmydatatoai.net/data-buyers/endless-data) | Non-exclusive, licensed to several labs | Fixed sum paid after transfer | | [Troveo](https://sellmydatatoai.net/data-buyers/troveo) | Time-limited | Revenue share, no fees | | micro1, Mercor, Scale AI, Handshake AI | Not published | Varies by agreement | Where a buyer does not publish its position, ask before you share a sample. The answer changes how you should price the deal. ## When does a non-exclusive license earn more? A non-exclusive license earns more when several buyers want the same data. Troveo says its datasets sell an average of four times, according to Don't Do It Yourself. Four non-exclusive fees can match or beat a single exclusive fee at 2.5 to 4 times the price. Large rights holders mostly license non-exclusively. Reddit licensed content to Google in February 2024, in a deal Reuters reported at about $60 million a year, and announced a separate partnership with OpenAI in May 2024. Business Insider reported that Shutterstock earned $104 million from AI licensing in 2023 and named its AI partners as including Meta, Alphabet, Amazon, Apple, OpenAI and Nvidia. Non-exclusive licensing works best for data that keeps growing. If you create new support tickets, code or project files every month, you can license fresh tranches to new buyers each year. ## When does an exclusive license make sense? An exclusive license makes sense when one buyer will pay well above what the open market would and you do not expect to sell the data again. Three situations fit. - Closing or selling the business, when a single clean payment matters more than future income - Rare data where only one or two buyers are likely to be interested - A closed archive you will never add to, such as records from a discontinued service Even then, check whether the exclusive price really reflects the 2.5 to 5 times range. An exclusive offer at the same level as non-exclusive offers is a poor deal. Exclusivity also has a cost after the deal. If a better buyer appears in year two, you cannot accept its offer for the same records within the agreed scope. A buyer that later goes out of business may still hold the exclusive rights, unless the contract ends them on insolvency. Ask for a clause that does, and for exclusivity to lapse if the buyer fails to pay on time. ## What are the options between exclusive and non-exclusive? The options between exclusive and non-exclusive limit exclusivity to what the buyer actually needs. Each one lets you keep some value while still giving the buyer something worth paying for. | Option | How it works | When to use it | | --- | --- | --- | | Time-limited exclusivity | Exclusive for a fixed period, such as 12 or 24 months, then non-exclusive | The buyer wants a head start, not permanent control | | Field-of-use exclusivity | Exclusive only for one use, such as legal drafting models | Your data has value in more than one domain | | Named-competitor exclusivity | You may not license to a listed set of the buyer's rivals | The buyer fears a specific competitor | | Model-type exclusivity | Exclusive for one kind of model, such as speech | Multimedia archives with several uses | | First-look or first-refusal right | The buyer sees new data first or can match other offers | Ongoing data where the buyer wants priority | | Tranche exclusivity | Exclusive for the current archive, non-exclusive for future records | Growing datasets | Troveo's time-limited model is a published example of the first option. The others are standard licensing tools, so a buyer's lawyer will recognize them. ## Can exclusivity be combined with a revenue share? Exclusivity can be combined with a revenue share, and Labelbox Prism publishes exactly that model: exclusive by default, with a fee at signing plus a revenue share on later use, according to Don't Do It Yourself. The buyer gets sole control of the data and you share in what it earns from it. This structure only works if you can see the numbers. Ask for regular statements, a right to audit and a minimum annual payment, so exclusivity does not leave you with a single upfront fee and nothing after it. ## What does exclusivity mean when the buyer resells to labs? Exclusivity with a training-data company usually means that company is the only route by which your data reaches AI labs. It does not mean only one lab uses it. Firms such as micro1, Mercor and Scale AI process data and supply it to labs, so an exclusive license to them may still put your records in several models. That matters for two reasons. You should be paid for the breadth of use the buyer gains, and you should know the onward transfer terms. Nixon Peabody recommends stating whether data may be shared with subprocessors or research partners, and that restrictions flow through the buyer's whole technical stack. ## What should an exclusivity clause say? An exclusivity clause should say exactly what you may not do, for how long and with which data. A vague clause gives the buyer more than it paid for. - The datasets covered, by reference to the manifest schedule - The field of use or model type the exclusivity applies to - The start date and end date, or a clear statement that it is perpetual - Whether future records you create are covered - Your right to keep using the data internally, including for your own AI tools - What happens to exclusivity if the buyer fails to pay or is acquired ## Is a perpetual exclusive license the same as a sale? A perpetual exclusive license is not legally a sale, but commercially it comes close. You keep ownership of the records, but you cannot license their AI value to anyone else, ever, within the agreed scope. Training cannot be reversed either. Frankfurt Kurnit Klein and Selz states that once data is used to train a model, it cannot practically be pulled back out. A [perpetual license](https://sellmydatatoai.net/glossary/perpetual-license) combined with exclusivity should therefore be priced as the last AI deal you will do with those records. > **Check what exclusive actually covers** Ask whether exclusivity stops you using your own data for internal AI tools. Some drafts prevent the licensor from any AI use, including its own. Most sellers want that carved out. ## How do you negotiate exclusivity? You negotiate exclusivity by treating it as a separate item with its own price. Get non-exclusive offers first, so you know the base value, then ask what each buyer will pay for each form of exclusivity. 1. **Price the non-exclusive license**: Collect at least two offers on identical records and terms without exclusivity. 2. **Ask what the buyer needs**: Most buyers want protection from specific rivals or a head start, not permanent control of your data. 3. **Offer the narrowest form that meets the need**: Time-limited or field-of-use exclusivity usually satisfies the buyer at a lower cost to you. 4. **Compare the premium with the ratio**: Check the exclusive price against AIData.Broker's 2.5 to 5 times range. 5. **Write the scope into the contract**: Define the field, the term and what you may still do. Use our [licensing agreement checklist](https://sellmydatatoai.net/guides/ai-data-licensing-agreement-checklist). Competing offers make this possible. When we broker a deal, offers are compared on the same records, uses, exclusivity and term so the highest headline is not mistaken for the best deal. To see what your data could fetch each way, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### Can I use my own data after granting an exclusive license? You can usually keep using your data for running your business, but an exclusive license may stop you using it for AI training, including your own internal models. It depends on the drafting. Ask for an express carve-out for internal use before you sign, so the position is clear. ### Is exclusivity always perpetual? Exclusivity is not always perpetual. Polyshares' four published deals were perpetual exclusive licenses, but Troveo describes its exclusivity as time-limited. You can negotiate a fixed term, such as 12 or 24 months, after which the license becomes non-exclusive and you may license the data to others. ### Does non-exclusive mean the buyer pays less? Non-exclusive usually means a lower fee per deal, because the buyer knows others may get the same data. AIData.Broker puts exclusive deals at about 2.5 to 5 times a non-exclusive one. The total can still be higher if you license the data several times, as Troveo's average of four sales suggests. ### What is a first-refusal right in a data license? A first-refusal right lets the buyer match any offer you receive from another buyer before you accept it. It is weaker than exclusivity but can still deter other bidders, who may not want to spend time on an offer that can be matched. Treat it as a form of exclusivity and price it. ### Should a closing company grant exclusivity? A closing company often grants exclusivity, because it will not be around to license the data again and a single larger payment helps creditors and investors. The Spirit Airlines data is being sold as an asset in bankruptcy, subject to court approval. Check the price still reflects exclusivity rather than distress. ## Sources - [How much is company data worth to AI? (2026 prices and real deals)](https://ddiy.co/how-much-is-company-data-worth-to-ai/), Don't Do It Yourself, accessed November 2026 - [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), Don't Do It Yourself, accessed November 2026 - [Reddit in AI content licensing deal with Google, sources say (Reuters report)](https://www.niemanlab.org/reading/reddit-in-ai-content-licensing-deal-with-google-sources-say/), Nieman Lab, accessed November 2026 - [Shutterstock earned over $100 million in revenue thanks in part to its AI-powered image generator tool](https://www.aol.com/shutterstock-earned-over-100-million-170917396.html), Business Insider, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 - [Shopping for AI training data: five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody, accessed November 2026 - [OpenAI and Reddit partnership](https://openai.com/index/openai-and-reddit-partnership/), OpenAI, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Can you legally sell company data for AI training? URL: https://sellmydatatoai.net/guides/can-you-legally-sell-company-data-for-ai Reviewed: 2026-11-08 You can legally sell or license company data for AI training if your company owns the records, no contract or professional rule forbids the use, and privacy law allows any personal data to be disclosed or it is removed. Holding data is not the same as having the right to sell it. Most companies can license some of their records, rarely all of them. ## Key takeaways - Holding records does not give you the right to license them. Ownership, contracts and privacy law each need to allow it. - Under UK law, employers own copyright in work staff create in the course of employment, but contractors keep theirs unless they assign it. - If you hold client data as a processor, UK GDPR Article 28 limits you to the client's documented instructions, so you cannot license it. - Client NDAs, vendor terms and professional confidentiality rules often rule out specific records. - Excluding high-risk records is usually safer and cheaper than trying to clear them. This guide sets out the legal questions buyers and their lawyers will ask before they sign, so you can answer them first. It describes what the law and regulators say. It is not advice on your situation, and Sell My Data To AI is a broker, not a law firm, so take independent legal advice before you license anything. ## Is it legal to sell company data for AI training? Selling company data for AI training is legal in principle, and companies do it openly. Frankfurt Kurnit Klein and Selz, in an October 2026 briefing by partner Daniel Goldberg, says such a sale can be lawful but that most companies cannot sell everything a buyer asks for, and that the work needed to sell safely is substantial. Nixon Peabody's October 2026 alert puts the core point plainly: possessing data does not mean you have permission to transfer it. It suggests three tests. Can the seller transfer the data? Can the buyer use it for the AI purpose? Can the buyer keep and commercialize what training produces? ## Does your company own the records it holds? Your company owns the records it created through its own staff, in most cases, but not everything on its systems. Ownership has to be checked for each source. Under section 11 of the Copyright, Designs and Patents Act 1988, the author of a work is the first owner of copyright, except that an employer owns copyright in literary, dramatic, musical or artistic works and films made by an employee in the course of employment, subject to any agreement to the contrary. Contractors and freelancers are not employees for this purpose, so their work stays theirs unless a contract assigns it. Collections of records can also carry database right. Regulation 13 of the Copyright and Rights in Databases Regulations 1997 says the right exists where there has been a substantial investment in obtaining, verifying or presenting a database's contents. Other countries have different rules, so check the law that applies to each dataset. ## Can you sell data you hold for clients? You usually cannot sell data you hold for clients without their agreement. Whether you can depends on your contract and on whether you act as a controller or a processor of any personal data in it. Under Article 28(3)(a) of the UK GDPR, a processor may process personal data only on documented instructions from the controller. Article 28(10) adds that a processor which decides the purposes and means of processing itself is treated as a controller for that processing, and takes on the liability that goes with it. Licensing a client's data to an AI company is a new purpose that a processor has no authority to choose. This affects agencies, software companies, call centers, outsourcers and IT providers most. Frankfurt Kurnit Klein and Selz advises vendors to check whether their customer agreements and data processing agreements would permit a sale before taking an offer. ## What do your contracts allow? Your contracts allow only what they say, and many say nothing about AI, which is not the same as permission. Read each agreement that covers the records you plan to license. - Client contracts, which may restrict use of deliverables, client materials or information learned during the work - NDAs, which Frankfurt Kurnit Klein and Selz notes often cover other companies' confidential information sitting in your email and chat - Vendor and platform terms. Frankfurt Kurnit Klein and Selz points out that Slack's API terms bar third-party apps from using API data to train large language models - Inbound licenses for content, software and research reports you use but do not own - Open-source licenses on code in your repositories - Employee policies and contracts, which may describe how staff communications are used Where a contract blocks a dataset, the usual answers are to exclude it or ask the other party for written consent. Consent requests should describe the use, the buyer type and the protections in plain terms. ## Does privacy law stop you selling the data? Privacy law does not stop you selling data outright, but it sets strict conditions on any personal data in it. You must handle personal data in line with the privacy notice in force when you collected it, and Frankfurt Kurnit Klein and Selz notes that the US Federal Trade Commission has warned that quietly adding AI training to a notice may be unfair or deceptive. Which laws apply depends on whose data it is and where they are. UK GDPR and EU GDPR cover people in the UK and EU. The California Consumer Privacy Act treats a transfer of personal information for value as a sale unless the data meets its de-identification standard, and Nixon Peabody notes that it covers employee and applicant data too. US sector laws such as HIPAA for health data and GLBA for financial data, state biometric laws and call recording consent rules can also apply. Our guide to [UK GDPR, EU GDPR and CCPA when selling data for AI training](https://sellmydatatoai.net/guides/gdpr-and-selling-data-for-ai-training) covers each regime, and [de-identification for AI training data](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data) explains how personal data is removed. ## Do professional confidentiality rules apply? Professional confidentiality rules apply to regulated firms on top of contract and privacy law. For solicitors in England and Wales, paragraph 6.3 of the SRA Code of Conduct for Solicitors requires them to keep the affairs of current and former clients confidential unless disclosure is required or permitted by law or the client consents. Legal privilege is a separate risk. Frankfurt Kurnit Klein and Selz warns that a sale can waive privilege and put trade secrets at risk. Accountants, medical practices and financial firms have their own rules. Our industry pages for [law firms](https://sellmydatatoai.net/sell-your-data/law-firms) and [medical practices](https://sellmydatatoai.net/sell-your-data/medical-practices) cover the specific restrictions. ## Do you need to involve staff or unions? You may need to involve staff or unions, and it is wise to even where the law does not require it. Workplace email and chat are written by employees, and some of it is about them. In the US, Frankfurt Kurnit Klein and Selz notes that employee discussions about pay, scheduling and conditions may be protected under Section 7 of the National Labor Relations Act, and that unionized workplaces may face collective bargaining restrictions. In the Spirit Airlines case, The Next Web reported that unions for Spirit's flight attendants and pilots objected to the data sale. UK and EU employers should check employment contracts, staff privacy notices and any collective agreements. ## Who can sign for a closed or closing company? For a closed or closing company, the person who can sign is usually the insolvency practitioner, trustee or board acting under the relevant process. Their authority and duties differ from those of normal directors. In US bankruptcy, Nixon Peabody notes that section 363(b)(1) of the Bankruptcy Code restricts sales of personally identifiable information that conflict with the debtor's privacy policy, and that compliance generally needs consistency with that policy or court approval after a consumer privacy ombudsman is appointed under section 332. In the Spirit Airlines case, the ombudsman recommended on October 5, 2026 that the court approve the sale of de-identified data to Google, according to The Next Web. Our guide to [selling data from a closed company](https://sellmydatatoai.net/guides/selling-data-from-a-closed-company) covers this in more detail. ## Which legal risks matter most, and how are they fixed? The legal risks that matter most are the ones a buyer cannot fix after delivery. The table summarizes each risk area with the question a buyer will ask and the usual fix. | Risk area | Question a buyer will ask | Usual fix | | --- | --- | --- | | Authority | Can the signatory bind the company? | Board resolution or insolvency practitioner's authority | | Ownership | Did the company create or acquire the records? | Confirm IP assignments, exclude contractor work without one | | Client contracts | Does any contract forbid this use? | Exclude or get written consent | | Processor status | Is any data held for a client as processor? | Exclude it | | Privacy notice | Did the notice at collection allow this disclosure? | Exclude or de-identify personal data | | Third-party content | Is licensed or open-source material included? | Identify and remove or clear it | | Confidentiality and privilege | Is privileged or NDA-covered material included? | Exclude whole categories | | Staff and unions | Have employees' rights been considered? | Exclude HR records, consult where required | > **Exclude rather than clear** Frankfurt Kurnit Klein and Selz recommends excluding high-risk categories outright: privileged material, HR and payroll records, health information, union communications, third-party confidential data, licensed data and data from regulated systems. ## Can a license expose your own confidential information? A license can expose your own confidential information, which is a commercial risk rather than a legal bar. Frankfurt Kurnit Klein and Selz notes that a buyer could use the data to build a competing product, and that models can reproduce parts of their training data in outputs. Nixon Peabody separately lists vendor pricing, security vulnerabilities and legal strategy as confidential material that needs its own review, apart from privacy and privilege. Treat pricing files, security reports, credentials and strategy documents as exclusions by default. The training value of these records is low, and the harm if they reappear in a model's output is high. ## Will the buyer disclose that it used your data? The buyer may have to disclose that it used licensed data. Frankfurt Kurnit Klein and Selz notes that California's AB 2013 requires generative AI developers to disclose training data sources, including whether data was purchased or licensed, and that the EU AI Act has similar requirements. A past data sale will also come up in later M&A due diligence, so keep a record of what you licensed and why. ## How do you check your right to sell before talking to buyers? You check your right to sell by working through each dataset against a fixed list of questions, before any sample leaves your systems. Our eight-point [vetting standard](https://sellmydatatoai.net/how-we-vet) covers authority to sign, ownership and chain of title, contract restrictions, privacy notice and lawful basis, third-party content, de-identification method, sample quality and manifest accuracy. Most issues found at this stage can be fixed by excluding records or getting a client's consent. For a rights pre-check on your records, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### Can I sell my company's emails to an AI company? You can license company email only after removing or clearing what you do not have the right to share. Email contains personal data, other companies' confidential information under NDAs and sometimes privileged advice. Google's bid for Spirit Airlines' data covered about 100 million emails, which The Next Web reported must be de-identified and certified before each transfer. ### Do I own work my contractors created? Under UK law you do not automatically own copyright in work created by contractors. Section 11 of the Copyright, Designs and Patents Act 1988 gives first ownership to the author, with an exception only for employees acting in the course of employment. You need a written assignment from the contractor. ### Can a software company sell its customers' data? A software company usually cannot sell data it holds for customers. Where it acts as a processor, UK GDPR Article 28 limits it to the customer's documented instructions. Its own records, such as internal code, tickets and documentation, may be licensable. Check customer agreements and data processing agreements first. ### Can I update my privacy notice to allow AI training? You can update a privacy notice for data collected in future, but a new notice does not automatically cover data you already hold. Frankfurt Kurnit Klein and Selz notes the US Federal Trade Commission has warned that quietly adding AI training may be unfair or deceptive, and that consent is needed for past data. ### Is anonymized data free of legal restrictions? Anonymized data falls outside data protection law, but other restrictions still apply. Contracts, NDAs, copyright, professional confidentiality and privilege are not affected by removing personal details. Frankfurt Kurnit Klein and Selz notes that de-identification does not resolve confidentiality problems, so each restriction needs its own check. ### What happens if I license data I did not have the right to license? If you license data without the right to do so, you may face claims from the true rights holder and from the buyer under the license warranties. Nixon Peabody notes that buyers seek indemnities covering privacy claims, confidentiality breaches, IP disputes and the cost of removing data or retraining a model. ## Sources - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 - [Shopping for AI training data: five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody, accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 11](https://www.legislation.gov.uk/ukpga/1988/48/section/11), legislation.gov.uk, accessed November 2026 - [Copyright and Rights in Databases Regulations 1997, regulation 13](https://www.legislation.gov.uk/uksi/1997/3032/regulation/13), legislation.gov.uk, accessed November 2026 - [UK GDPR Article 28: processor](https://www.legislation.gov.uk/eur/2016/679/article/28), legislation.gov.uk, accessed November 2026 - [SRA Code of Conduct for Solicitors, RELs and RFLs](https://www.sra.org.uk/solicitors/standards-regulations/code-conduct-solicitors/), Solicitors Regulation Authority, accessed November 2026 - [Spirit's privacy ombudsman backs Google's $10m data deal](https://thenextweb.com/news/google-spirit-airlines-data-ombudsman-approval), The Next Web, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # UK GDPR, EU GDPR and CCPA when selling data for AI training URL: https://sellmydatatoai.net/guides/gdpr-and-selling-data-for-ai-training Reviewed: 2026-11-08 UK GDPR and EU GDPR apply to any personal data in records you license for AI training, and the CCPA treats a transfer of personal information for value as a sale. Data that is truly anonymous falls outside GDPR, but the ICO says pseudonymized data is still personal data for anyone holding the key. Most sellers de-identify or exclude personal data before licensing. ## Key takeaways - Anonymous information falls outside UK and EU GDPR. Pseudonymized data does not, for anyone who holds the additional information. - The ICO's anonymization guidance tests whether identification is reasonably likely, using the motivated intruder test, and treats anonymizing as processing in its own right. - EDPB Opinion 28/2024 says models trained on personal data are not always anonymous, and that deployers should check whether the data was processed lawfully. - Under the CCPA, handing over personal information for value is a sale unless the data meets a three-part de-identification standard. - The privacy notice in force when you collected the data limits what you can now do with it. Most company archives contain personal data: names in emails, customers in support tickets, staff in chat. This guide explains how UK GDPR, EU GDPR and the California Consumer Privacy Act treat licensing those records to an AI buyer, citing the ICO's anonymization guidance and the European Data Protection Board's Opinion 28/2024. > **Not legal advice** This guide summarizes what the law and regulators say. It is not advice on your situation. Sell My Data To AI is a broker, not a law firm. Take independent legal advice before licensing any data that contains or was derived from personal data. ## Does GDPR apply when you sell data for AI training? GDPR applies when the data you sell contains [personal data](https://sellmydatatoai.net/glossary/personal-data), meaning information about an identified or identifiable person. If the data is truly anonymous, UK GDPR and EU GDPR do not apply to it. Under Recital 26, identifiability is judged by all the means reasonably likely to be used to identify someone. In practice, almost every company archive starts as personal data. The question is whether you can make it anonymous before licensing, or whether you need a lawful basis to disclose it as personal data. Most sellers aim for the first and exclude what they cannot anonymize. ## When is data anonymous under the ICO's guidance? Data is anonymous under the ICO's guidance when the risk of identifying anyone from it is sufficiently remote, taking into account all the means reasonably likely to be used. The ICO published its anonymization guidance in March 2025, according to Taylor Wessing's summary, and says the guidance is under review following the Data (Use and Access) Act, which became law on June 19, 2025. The guidance makes four points that matter for a data sale: - Identifiability sits on a spectrum. There is a blurred band between clearly personal and clearly anonymous data, and most disputes arise in that band - The motivated intruder test asks whether a determined person with access to resources and expertise could identify individuals from the data - Identifiability depends on context and can change over time, so the assessment should be reviewed, including when new recipients are given access - The risk does not have to be zero, but the ICO expects objective factors such as cost, time and available technology to be assessed Taylor Wessing's summary also records that the ICO treats the process of [anonymization](https://sellmydatatoai.net/glossary/anonymization) as a processing activity in itself. That means you need a lawful basis and a defined purpose for anonymizing the data, and the ICO expects a data protection impact assessment to structure the decisions. ## Is pseudonymized data still personal data? Pseudonymized data is still personal data for anyone who holds the additional information needed to re-identify it. The ICO's pseudonymization page states this directly, and adds that if pseudonymized data is shared without the additional information, it may be anonymous information in the recipient's hands. The EU's top court reached a similar view. In EDPS v SRB, decided on September 4, 2025, the Court of Justice held that pseudonymized data is not personal data in all cases and for every person, and that it may not be personal data for a recipient who cannot re-identify individuals by means reasonably likely to be used, according to Clifford Chance's analysis. The court also held that the duty to tell people about recipients is assessed from the original controller's side, at the time of collection. For a seller, this cuts both ways. Replacing names with codes and keeping the key means the data stays personal data in your hands, so your disclosure still needs a lawful basis. [Pseudonymization](https://sellmydatatoai.net/glossary/pseudonymization) reduces risk, but it is not a way out of GDPR for the company doing it. ## What lawful basis do you need to disclose personal data for AI training? You need one of the lawful bases in Article 6 of UK GDPR or EU GDPR to disclose personal data for AI training, and the new purpose must be compatible with the one you collected it for. Most commercial sellers would look at legitimate interests, which requires a purpose test, a necessity test and a balancing test. Transparency duties apply at the same time. Article 14(1)(e) requires controllers to tell people about the recipients or categories of recipients of their data. Article 14(4), like Article 13(3) for data collected directly, requires controllers to tell people before further processing their data for a new purpose. The privacy notice in force when you collected the data is therefore the starting point for any [lawful basis](https://sellmydatatoai.net/glossary/lawful-basis) assessment. Special category data, such as health information, needs an additional condition under Article 9. Frankfurt Kurnit Klein and Selz recommends excluding health information from company data sales entirely, which is what most sellers do. ## What does EDPB Opinion 28/2024 say about AI training data? EDPB Opinion 28/2024, adopted by the European Data Protection Board on December 17, 2024 at the request of the Irish supervisory authority, says AI models trained on personal data cannot always be treated as anonymous, sets out how legitimate interests applies to AI development and explains the consequences when training data was processed unlawfully. The opinion is addressed to supervisory authorities rather than to companies, but it shows how EU regulators are likely to assess AI cases. | Topic | What the opinion says | Why it matters to a seller | | --- | --- | --- | | Anonymous models | A model is anonymous only if the likelihood of extracting personal data, directly or through queries, is insignificant. This is assessed case by case | Buyers cannot assume training removes personal data, so they will ask what was in yours | | Legitimate interests | A three-step test: a lawful, clearly stated and real interest, necessity, and a balancing test | The same test applies to your decision to disclose | | Reasonable expectations | Meeting the transparency rules is not enough on its own to show people expect the processing (paragraph 92) | An updated privacy notice may not make AI use expected | | Source of the data | Expectations differ depending on whether people gave data directly or it came from a third party (paragraph 94) | Records bought in from others carry more risk | | Unlawful development | A different controller deploying a model should check that it conducted an appropriate assessment of how the data was processed | Buyers will ask sellers for evidence that collection and disclosure were lawful | | Anonymization before deployment | If a model is anonymized before deployment, unlawful initial processing should not affect its later operation | Buyers favor data that is anonymized before it reaches them | The practical effect is that EU buyers now carry more of the risk of a seller's mistakes. That makes them ask harder questions about notices, lawful basis and de-identification, and pay less for data where the answers are weak. ## How does the CCPA treat selling data for AI training? The CCPA treats a transfer of personal information to an AI company for money or other value as a sale, unless the data is de-identified to the standard in California Civil Code section 1798.140(m), according to Frankfurt Kurnit Klein and Selz. A sale triggers notice, opt-out and risk assessment duties. Nixon Peabody summarizes the de-identification standard as three conditions: the data cannot reasonably be linked to an individual, the holder publicly commits not to re-identify it, and recipients are contractually bound to the same. It also notes that the CCPA generally gives employees and job applicants consumer-style rights, so staff email and chat are covered, not only customer data. The Spirit Airlines case shows how buyers apply this. The Next Web reported that the buyer must publicly commit not to re-identify the data and may pass it on only to recipients who make the same commitment. ## How do UK GDPR, EU GDPR and the CCPA compare? UK GDPR, EU GDPR and the CCPA all restrict selling personal data, but they draw the line in different places. The table compares the points that matter for a data license. | Point | UK GDPR | EU GDPR | CCPA | | --- | --- | --- | --- | | Regulator | Information Commissioner's Office | National supervisory authorities, coordinated by the EDPB | California Privacy Protection Agency and California Attorney General | | Outside the law when | Data is anonymous, judged by means reasonably likely to be used | Same test under Recital 26 | Data meets the 1798.140(m) de-identification standard | | Pseudonymized data | Personal data for anyone holding the key | Personal data for the holder, possibly not for a recipient who cannot re-identify (EDPS v SRB) | Not de-identified unless all three conditions are met | | Key seller duties | Lawful basis, compatible purpose, transparency about recipients | Same, plus EDPB guidance on AI | Notice, opt-out and risk assessment for a sale | | Main guidance | ICO anonymization guidance | EDPB Opinion 28/2024 | Statute and regulations | ## What has the Data (Use and Access) Act 2025 changed? The Data (Use and Access) Act 2025 has amended the UK GDPR, and the ICO says its anonymization guidance is under review as a result. legislation.gov.uk shows Article 14(5)(b), one of the exemptions from telling people about indirectly collected data, as omitted from February 5, 2026. Check the current text of the UK GDPR and the ICO's latest guidance before relying on any exemption. The rules on transparency and further processing are the ones most likely to affect a data license. ## What should you do before licensing records that contain personal data? Before licensing records that contain personal data, decide what to exclude, document how the rest is de-identified and make the buyer contractually responsible for not re-identifying it. These steps follow the ICO, EDPB and US guidance cited above. 1. **Map the personal data**: List whose data is in each dataset (staff, customers, contacts at other companies), which countries they are in and which privacy notice applied when you collected it. 2. **Exclude the highest-risk records**: Take out HR files, health information, children's data, payment data and anything you cannot de-identify reliably. 3. **Run a data protection impact assessment**: The ICO expects one for anonymization. Record the purpose, the risks and the measures you chose. 4. **De-identify by a documented method**: Remove direct and indirect identifiers, then apply the motivated intruder test and a [re-identification risk](https://sellmydatatoai.net/glossary/re-identification-risk) test on a sample. 5. **Check notices and lawful basis**: Confirm that any data that remains personal can lawfully be disclosed, and that people have been told about recipients. 6. **Bind the buyer**: Prohibit re-identification, require the same from any recipient, and set deletion dates for raw copies. 7. **Keep the file**: Store the assessment, the method, the test results and the certification. Buyers and acquirers will ask for them. Our guides on [de-identification for AI training data](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data) and [whether you can legally sell company data](https://sellmydatatoai.net/guides/can-you-legally-sell-company-data-for-ai) go further. EU-facing buyers also face disclosure duties of their own, explained in [the EU AI Act and training data transparency](https://sellmydatatoai.net/guides/eu-ai-act-training-data-transparency). For a rights pre-check on your records, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### Is anonymized data covered by UK GDPR? Anonymized data is not covered by UK GDPR, provided it is truly anonymous. The ICO judges this by whether identification is reasonably likely, using tests such as the motivated intruder test. The ICO also treats the process of anonymizing personal data as processing, so that step needs a lawful basis. ### Can I sell pseudonymized data without worrying about GDPR? You cannot treat pseudonymized data as outside GDPR if you keep the key. The ICO states that pseudonymized data is personal data in the hands of anyone holding the additional information. It may be anonymous for a recipient without the key, but your own disclosure still needs a lawful basis. ### What is EDPB Opinion 28/2024? EDPB Opinion 28/2024 is guidance from the European Data Protection Board, adopted on December 17, 2024, on personal data and AI models. It says models trained on personal data are not always anonymous, explains how legitimate interests applies to AI development, and sets out the consequences of unlawful processing during training. ### Does the CCPA apply to employee data in a data sale? The CCPA generally applies to employee and job applicant data as well as customer data, according to Nixon Peabody. Licensing staff email or chat containing personal information for value can therefore be a sale under the CCPA, unless the data meets its de-identification standard. ### Do I need consent to sell data for AI training? Consent is not the only lawful basis under UK or EU GDPR, but your privacy notice and the purpose you collected data for limit what you can do without it. Frankfurt Kurnit Klein and Selz notes that consent is needed for past data in the US where notices did not cover the new use. Take advice for your case. ### Does licensing data abroad raise transfer issues? Licensing personal data to a buyer outside the UK or EU can be a restricted international transfer, which needs its own safeguard such as an adequacy decision or standard contractual clauses. Data that is truly anonymous before it leaves avoids this, which is one reason most sellers anonymize first. ## Sources - [Anonymisation guidance](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/), Information Commissioner's Office, accessed November 2026 - [Pseudonymisation](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/pseudonymisation/), Information Commissioner's Office, accessed November 2026 - [New ICO guidance on anonymisation and pseudonymisation](https://www.taylorwessing.com/de/global-data-hub/2025/spotlight-on-the-uk-data-landscape/gdh---new-ico-guidance-on-anonymisation-and-pseudonymisation), Taylor Wessing, accessed November 2026 - [Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models](https://www.edpb.europa.eu/our-work-tools/our-documents/opinion-board-art-64/opinion-282024-certain-data-protection-aspects_en), European Data Protection Board, accessed November 2026 - [Opinion 28/2024 (full text, PDF)](https://www.edpb.europa.eu/system/files/2024-12/edpb_opinion_202428_ai-models_en.pdf), European Data Protection Board, accessed November 2026 - [Pseudonymized data after EDPS v SRB](https://www.cliffordchance.com/insights/resources/blogs/talking-tech/en/articles/2025/09/pseudonymized-data-after-edps-v-srb.html), Clifford Chance, accessed November 2026 - [UK GDPR Article 14: information to be provided where personal data have not been obtained from the data subject](https://www.legislation.gov.uk/eur/2016/679/article/14), legislation.gov.uk, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 - [Shopping for AI training data: five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody, accessed November 2026 - [Spirit's privacy ombudsman backs Google's $10m data deal](https://thenextweb.com/news/google-spirit-airlines-data-ombudsman-approval), The Next Web, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # De-identification for AI training data URL: https://sellmydatatoai.net/guides/de-identification-for-ai-training-data Reviewed: 2026-11-08 De-identification for AI training data means removing or transforming everything in your records that identifies a person, directly or indirectly, before a buyer sees them. Removing names is not enough. The ICO says stripping direct identifiers alone is insufficient for anonymization, so you need exclusions, a documented method, a random review sample and a re-identification test before transfer. ## Key takeaways - Removing names is not de-identification. The ICO says removing direct identifiers alone is insufficient for effective anonymization. - Pseudonymized data stays personal data for anyone who holds the key, so destroy the mapping table or keep it under separate control. - Exclude HR, payroll, health and privileged material entirely rather than trying to redact it line by line. - Prove the method with a written process, a random sample, a measured error rate and a re-identification test. - De-identify before transfer. Data used to train a model cannot practically be pulled back out. AI buyers want real working records: email threads, support tickets, call notes and code reviews. Those records are full of people. De-identification is the work that turns them into something you can license without handing over personal data, and it is often the step that decides whether a buyer signs. This guide covers what UK, EU and US law means by de-identification, the techniques that work on the unstructured records AI buyers want, and how to prove to a buyer that your method worked. It describes what regulators say. It is not legal advice for your records, so take independent advice before you rely on it. ## What does de-identification mean for AI training data? De-identification for AI training data means removing or transforming the information in your records that identifies a person before anyone outside your company sees them. The word covers a range of outcomes, and the legal result depends on where your data ends up on that range. The ICO describes [anonymization](https://sellmydatatoai.net/glossary/anonymization) as "the way in which you turn personal data into anonymous information", and states that "data protection law does not apply to anonymous information". [Pseudonymization](https://sellmydatatoai.net/glossary/pseudonymization) is different. Under Article 4(5) of the UK GDPR it means data can no longer be attributed to a person without additional information that is kept separately, and the ICO says pseudonymized data is still personal data in the hands of anyone who holds that additional information. | Term | What it means | Still personal data? | | --- | --- | --- | | Anonymization | Personal data turned into information that no longer relates to an identifiable person (ICO) | No, if the anonymization is effective | | Pseudonymization | Identifiers replaced, with the key kept separately (UK GDPR Article 4(5)) | Yes, for anyone who holds the key | | De-identified personal data | Data that cannot be attributed to a person without more (Data Protection Act 2018, section 171) | Yes, and re-identifying it can be a criminal offense | | De-identified health information | US health data meeting HIPAA's Safe Harbor or Expert Determination method (45 CFR 164.514) | Not protected health information under HIPAA | Section 171 of the [Data Protection Act 2018](https://www.legislation.gov.uk/ukpga/2018/12/section/171) makes it an offense to knowingly or recklessly re-identify de-identified personal data without the consent of the controller that de-identified it. That offense gives a UK license some extra protection, but it does not turn pseudonymized data into anonymous data. ## Why is removing names not enough? Removing names is not enough because people can be identified from combinations of ordinary facts. The ICO's anonymization guidance says "simply removing direct identifiers from a dataset is insufficient to ensure effective anonymization", and warns about the "mosaic or jigsaw effect" of linking records together. Research supports the warning. A 2019 study in Nature Communications by Luc Rocher, Julien Hendrickx and Yves-Alexandre de Montjoye estimated that 99.98% of Americans would be correctly re-identified in any dataset using 15 demographic attributes. Workplace records make the problem harder. An email that mentions a job title, an office and a date can point to one person with every name removed. The ICO notes that techniques for quantitative data "are not generally applicable" to qualitative or unstructured data, and that anonymizing it "can require careful human judgment to determine if people are still identifiable". ## Which legal tests apply to de-identified training data? The legal test for de-identified training data depends on whose data it is and where they live. Most sellers deal with at least two regimes, because a UK company's records usually include EU, US or other overseas contacts. - UK GDPR: Recital 26 asks whether identification is possible by means "reasonably likely to be used", taking account of cost, time and the technology available. The ICO's test is the motivated intruder, a reasonably competent person with internet access and public records who wants to identify someone. - EU GDPR: the Court of Justice ruled on September 4, 2025 in EDPS v SRB (Case C-413/23 P) that pseudonymized data is not personal data "in all cases and for every person". A recipient without means reasonably likely to re-identify may not be handling personal data, but the disclosing company's duties are judged from its own position when it collected the data. - EU guidance: the European Data Protection Board's Guidelines 01/2025, adopted in January 2025, say pseudonymized data "remains information related to an identifiable natural person". - US health data: HIPAA allows Safe Harbor, which removes 18 types of identifier, or Expert Determination, where a qualified expert concludes the risk of identification is "very small" and documents the method. - California: the CCPA treats a transfer of personal information for value as a sale unless the data is de-identified, which requires technical measures, a public commitment not to re-identify and contractual obligations on recipients, according to Frankfurt Kurnit Klein & Selz. HIPAA's Safe Harbor list shows how far de-identification has to go. According to the [US Department of Health and Human Services](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html), it removes all date elements except the year, keeps the first three digits of a ZIP code only where that area holds more than 20,000 people, and groups everyone over 89 into a single category of 90 or older. ## How do you de-identify company records for AI training? You de-identify company records for AI training in a fixed order: decide what to leave out, find every identifier in what remains, transform it, then test the result. Skipping the first step is the quickest way to fail a buyer's review. 1. **Exclude the highest-risk material**: Daniel Goldberg of Frankfurt Kurnit Klein & Selz advises sellers to "exclude high-risk data entirely rather than relying on de-identification", naming privileged material, HR and personnel files, payroll and tax records, and health and benefits information. Removing a whole folder is cheaper and safer than redacting it line by line. 2. **List direct and indirect identifiers**: Direct identifiers include names, email addresses, phone numbers, account numbers and signatures. Indirect identifiers include job titles, office locations, rare events, dates, usernames, file paths and document metadata. List them per source system, because a Slack export and a scanned contract hide identifiers in different places. 3. **Choose a technique for each field**: Suppress what adds nothing, replace names with consistent placeholders where conversation structure matters, and generalize dates and places to the level a buyer needs, such as month and region. 4. **Run automated detection, then human review**: Open-source tools such as Microsoft Presidio detect and replace common identifier types in text. Automated detection misses context, so have people review a random sample from each source and record what they find. 5. **Separate or destroy the key**: If you replace names with tokens, the table that maps tokens back to people keeps the data pseudonymized for whoever holds it. Destroy it, or keep it under separate control and never share it with the buyer. 6. **Test re-identification**: Ask someone who did not do the redaction to try to identify people in a sample, using public sources such as LinkedIn profiles and your own website. That is the ICO's motivated intruder test in practice. 7. **Document the method and the results**: Record the tools, rules, reviewers, sample sizes and error rates. Buyers ask for this, and it is your evidence if anyone later questions the release. ## Which techniques suit which types of record? Different types of record need different techniques because identifiers sit in different places. The table shows where they usually hide and the technique most sellers use. | Record type | Where identifiers hide | Usual technique | | --- | --- | --- | | Email and chat | Names, signatures, quoted replies, attachments, mentions | Consistent name tokens, strip signatures and headers, review attachments separately | | Support tickets | Customer names, order numbers, addresses, pasted screenshots | Replace identifiers with tokens, drop images or review them by hand | | Source code | Commit author names and emails, credentials, test fixtures with real customer data | Rewrite commit metadata, scan for secrets, replace fixture data | | Call recordings | Voices, names spoken aloud, card numbers read out | Transcribe and redact, drop audio where the voice identifies, cut payment segments | | Video | Faces, number plates, screens, name badges | Blur or exclude, which the ICO lists as a method for video | | CRM and spreadsheets | Names, emails, rare combinations of fields | Suppress direct identifiers, generalize fields, check group sizes | Voice needs extra care. The ICO's biometric guidance, published in March 2024, lists voice among the characteristics that can underlie biometric data once it is processed to identify someone, and biometric data used for unique identification is special category data. If a buyer wants raw audio, treat it as personal data unless you can show otherwise. Our page on [call recording data](https://sellmydatatoai.net/buy-ai-training-data/call-recording-data) covers what buyers check. ## Does consistent pseudonymization keep the data useful? Consistent pseudonymization keeps conversational data useful because a model can still follow who said what to whom. Replacing every mention of one colleague with the same token, such as PERSON_14, preserves the thread in a way that deleting names does not. The trade-off is legal. The ICO says that if you share pseudonymized data without the additional information, "it may be anonymous information in their hands", and the SRB judgment points the same way in the EU. The data stays personal data for you while you hold the key, and a buyer that could link tokens to public sources may still be processing personal data. Consistency also raises linkage risk. A token that appears in thousands of messages builds a profile of one person's work, opinions and movements. Review the most frequent tokens by hand before release, and generalize or remove the details that would let an outsider attach a name to them. ## How do you prove de-identification worked? You prove de-identification worked with evidence a buyer can check: a written method, a random sample, a measured error rate and an independent test. A statement that the data "has been anonymized" carries little weight on its own. Buyers are starting to offer outside review. In the Spirit Airlines bankruptcy sale, micro1 offered to cover de-identification and independent review costs and to give an independent reviewer a 1% sample of the de-identified data, Fortune reported in September 2026. Our own seller vetting includes the same principle: direct and indirect identifiers are removed by a documented method, and a re-identification test is run on a sample. The full standard is on [how we vet](https://sellmydatatoai.net/how-we-vet). Record the method and its results in your [data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest) so the buyer sees them before the sample. ## Can de-identified data become identifiable later? De-identified data can become identifiable later, because new public data and better tools change what a motivated intruder can do. The ICO describes identifiability as a spectrum and says information can move along it as techniques and available data change, so assessments need to be kept under review. Training adds its own risk. Nicholas Carlini and colleagues showed in 2021 that they could extract hundreds of verbatim sequences from GPT-2's training data, including names, phone numbers and email addresses, and found that "larger models are more vulnerable than smaller models". A 2023 paper by Milad Nasr and others described an attack that made ChatGPT emit training data at a rate 150 times higher than normal. The European Data Protection Board's Opinion 28/2024, adopted on December 17, 2024, says AI models trained with personal data "cannot, in all cases, be considered anonymous". Goldberg puts the seller's position plainly: once data is used to train a model, "it cannot practically be pulled back out". De-identify before transfer, not after. > **Rule of thumb** If removing the identifiers from a record type would leave nothing a buyer wants, exclude that record type. Redaction is for data that keeps its value without the people in it. ## What should the license say about de-identification? The license should say who de-identified the data, by what method, and what the buyer may not do with it. The terms most often used are: - A buyer promise not to attempt re-identification or to link the data with other sources to identify people - A ban on onward transfer except to named parties bound by the same terms - Deletion of raw copies by a fixed date - A procedure if either side finds personal data that slipped through - Who pays for a re-review if the buyer's own checks find identifiers The [AI data licensing agreement checklist](https://sellmydatatoai.net/guides/ai-data-licensing-agreement-checklist) covers these clauses in more detail. For the wider legal position, read our guide to [UK GDPR, EU GDPR and CCPA when selling data](https://sellmydatatoai.net/guides/gdpr-and-selling-data-for-ai-training). If you hold records with people in them and want to know whether de-identification would leave enough value to sell, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). We will tell you which record types buyers want and which are better left out. ## Frequently asked questions ### Is pseudonymized data the same as anonymized data? No. Pseudonymized data has identifiers replaced, but someone holding the key can reverse it, so the ICO treats it as personal data in their hands. Anonymized data no longer relates to an identifiable person, and data protection law does not apply to it. The EU Court of Justice ruled in September 2025 that pseudonymized data may not be personal data for a recipient with no realistic means of re-identifying people. ### Can I use HIPAA Safe Harbor for UK or EU records? HIPAA Safe Harbor is a US standard for health information held by covered entities. It is a useful checklist of identifier types, but meeting it does not show that data is anonymous under the UK GDPR or EU GDPR. Those regimes ask whether identification is reasonably likely by any means, which the ICO tests with its motivated intruder approach. Take advice on which test applies. ### Can software de-identify my records automatically? Software can do most of the first pass. Open-source tools such as Microsoft Presidio find and replace common identifiers like names, emails and phone numbers in text. They miss context, such as a job title that points to one person, so a human review of a random sample is still needed. The ICO says unstructured data needs careful human judgment. ### Who should run the re-identification test? Someone who did not carry out the redaction, ideally with no inside knowledge of the people in the records. They should try to identify individuals in a random sample using public sources, which reflects the ICO's motivated intruder test. For larger or sensitive datasets, an independent reviewer adds credibility with buyers, as micro1 offered in the Spirit Airlines sale. ### Will de-identification reduce what buyers pay? It can, if identifiers carry the value. More often the opposite holds, because data with personal information you cannot clear is harder to sell. AIData.Broker lists uncleared personal data as a factor that lowers price, as reported by ddiy.co in October 2026. Consistent pseudonymization keeps conversations readable, which preserves most of the training value. ### Do I need to tell people their data will be anonymized? The ICO says anonymization is itself processing of personal data, so it must comply with data protection law, including transparency. That usually means your privacy information should cover preparing data for licensing before you start. The exact duty depends on your notices and lawful basis, so take independent advice. ## Sources - [Introduction to anonymisation](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/), Information Commissioner's Office, accessed November 2026 - [How do we ensure anonymisation is effective?](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/how-do-we-ensure-anonymisation-is-effective), Information Commissioner's Office, accessed November 2026 - [Data Protection Act 2018, section 171](https://www.legislation.gov.uk/ukpga/2018/12/section/171), legislation.gov.uk, accessed November 2026 - [Estimating the success of re-identifications in incomplete datasets using generative models](https://pmc.ncbi.nlm.nih.gov/articles/PMC6650473), Nature Communications (Rocher, Hendrickx and de Montjoye, 2019), accessed November 2026 - [Guidance regarding methods for de-identification of protected health information](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html), US Department of Health and Human Services, accessed November 2026 - [CJEU clarifies concept of personal data for a transfer of pseudonymised data to third parties](https://www.aoshearman.com/en/insights/ao-shearman-on-data/cjeu-clarifies-concept-of-personal-data-for-a-transfer-of-pseudonymised-data-to-third-parties), A&O Shearman, accessed November 2026 - [EDPB adopts pseudonymisation guidelines](https://edpb.europa.eu/news/news/2025/edpb-adopts-pseudonymisation-guidelines-and-paves-way-improve-cooperation_en), European Data Protection Board, accessed November 2026 - [Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models](https://edpb.europa.eu/system/files/2024-12/edpb_opinion_202428_ai-models_en.pdf), European Data Protection Board, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein & Selz, accessed November 2026 - [Google wants Spirit Airlines' data. Micro1 wants to pay more](https://fortune.com/2026/09/14/micro1-tries-to-beat-googles-bid-to-buy-bankrupt-spirit-airlines-data-for-12-5-million-highlighting-value-of-data-to-ai-companies/), Fortune, accessed November 2026 - [Extracting training data from large language models](https://www.usenix.org/conference/usenixsecurity21/presentation/carlini-extracting), USENIX Security Symposium 2021, accessed November 2026 - [Scalable extraction of training data from (production) language models](https://arxiv.org/abs/2311.17035v1), arXiv (Nasr and others, 2023), accessed November 2026 - [ICO publishes biometric data guidance](https://www.ropesgray.com/en/insights/viewpoints/102j5cq/ico-publishes-biometric-data-guidance), Ropes & Gray, accessed November 2026 - [Presidio: data protection and de-identification SDK](https://microsoft.github.io/presidio), Microsoft, accessed November 2026 - [How much is company data worth to AI?](https://ddiy.co/how-much-is-company-data-worth-to-ai/), ddiy.co, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How to build a data manifest buyers will trust URL: https://sellmydatatoai.net/guides/how-to-build-a-data-manifest Reviewed: 2026-11-08 A data manifest is the structured inventory a buyer reads before deciding whether your data is worth a sample or an offer. It lists sources, date ranges, record counts, formats, languages, exclusions, de-identification method, rights per source and file hashes. Buyers trust a manifest when it is tied to one exact version of the data and every count can be checked on delivery. ## Key takeaways - A data manifest lets a buyer judge value and risk without seeing the full data, which breaks the trust deadlock between seller and buyer. - Count records from the export you will deliver, never from a tool's admin dashboard. - Tie the manifest to one version of the data with SHA-256 file hashes and a version number. - Borrow structure from Datasheets for Datasets, Croissant and the Data Provenance Standards rather than starting from scratch. - List exclusions and known defects. Disclosed gaps rarely become disputes, hidden ones often do. A data manifest is the document a buyer reads before deciding whether your data is worth a sample, a pilot or an offer. A buyer cannot see your full data before signing, and you cannot show it. The manifest is how both sides get past that. In a July 2026 Hugging Face forum thread on why selling datasets is so hard, the original poster put the problem in two sentences: "The seller can't prove quality without exposing too much." And: "The buyer can't trust quality without seeing enough." A precise manifest tied to an exact version of the data is the most practical answer sellers have. ## What is a data manifest? A [data manifest](https://sellmydatatoai.net/glossary/data-manifest) is a structured inventory of a dataset: what it contains, where each part came from, how much there is, what was removed and what rights attach to it. It describes the data without exposing it, so a buyer can judge fit and price before seeing a sample. It overlaps with other documentation formats, such as the [dataset card](https://sellmydatatoai.net/glossary/dataset-card), but it has a different job. The table shows how they compare. | Document | Main purpose | Who usually writes it | | --- | --- | --- | | Data manifest | Inventory for a sale: contents, counts, sources, rights, exclusions, file hashes | The seller or broker | | Dataset card | Public description of a dataset on a hub such as Hugging Face | The dataset publisher | | Datasheet for datasets | Answers to set questions on motivation, composition, collection, preprocessing, uses, distribution and maintenance | The dataset creator | | Croissant file | Machine-readable metadata built on schema.org, readable by ML tools | The publisher or platform | | EU AI Act training summary | Public summary of training content that general-purpose model providers must publish | The model provider, not the seller | ## What should a data manifest include? A data manifest should include enough detail for a buyer to estimate value and spot legal risk without sending you a long list of questions. These are the fields buyers ask about most. | Section | Fields | Why buyers ask | | --- | --- | --- | | Identity | Dataset name, version number, release date, seller legal entity | Ties every claim to one exact package | | Sources | Each system exported, such as Slack, Microsoft 365, Jira, GitHub or Zendesk | Shows what kind of work the data captures | | Coverage | Date range per source, teams or departments covered, gaps | Years of history and continuity affect price | | Volume | Record counts by type, file counts, total size, tokens or hours | The basis for most pricing | | Format | File formats, schema, encoding, how threads and attachments link | Shows how much processing the buyer will need | | Language | Languages and the approximate share of each | Many buyers want English first, others want specific languages | | Exclusions | What was removed and why, such as HR, payroll or privileged material | Shows the high-risk material is gone | | De-identification | Method, tools, review sample size, error rate, re-identification test | Evidence the data can be transferred lawfully | | Rights | Who created each source, IP assignments, contract restrictions, third-party content | Buyers will not sign without a clear chain of title | | Known issues | Duplicates, corrupt files, missing months, quality limits | Disclosed defects do not become disputes later | | Sample | How the sample was drawn and its size | Shows whether the sample reflects the whole | | Integrity | File list with SHA-256 hashes and a hash of the manifest itself | Proves the delivered files match the description | Several of these fields drive price directly. AIData.Broker, as reported by ddiy.co in October 2026, lists years of records, linked records and settled rights among the factors that raise what buyers pay. Our guide to [what your data is worth](https://sellmydatatoai.net/guides/how-much-is-your-data-worth-to-ai) covers the rest. ## How do you count records accurately? You count records accurately by counting the exported files you will actually deliver, not the figures in a tool's admin dashboard. Dashboards include deleted items, bots, system messages and data outside your export. - Define the counting unit for each source: message, thread, ticket, file, commit or minute of audio - Deduplicate before counting, and state the method - Count after exclusions and de-identification, not before - Report counts per year so gaps are visible - Keep the script that produced the counts so the buyer can rerun it on delivery Manifest accuracy is one of the eight checks in our [seller vetting standard](https://sellmydatatoai.net/how-we-vet): record counts, date ranges, systems and languages in the manifest must match what is actually delivered. A buyer who receives fewer records than the manifest promised has grounds to reprice or reject the delivery. ## Which standards can you borrow from? You can borrow structure from four published frameworks rather than designing a manifest from scratch. None is a legal requirement for a private license, but buyers recognize them and some will map your manifest onto them. - Datasheets for Datasets, by Timnit Gebru and six co-authors and first posted in March 2018, sets out questions in seven sections: motivation, composition, collection process, preprocessing and labeling, uses, distribution and maintenance. - Croissant, announced by Google Research and MLCommons on March 6, 2024, is a machine-readable metadata format built on schema.org. Kaggle, Hugging Face and OpenML support it. - The Data & Trust Alliance's Data Provenance Standards were developed by a working group from 19 enterprises, including American Express, IBM, Mastercard, Pfizer and Walmart. They track a dataset's origin, how it was created and whether it can be legally used. - The European Commission's template for the public summary of training content, published on July 24, 2025, asks general-purpose model providers to say whether they used commercially licensed content. The last point matters even though the obligation falls on the buyer. Under Article 53 of the EU AI Act, providers of general-purpose AI models must publish the summary from August 2, 2025, and models already on the market before that date have until August 2, 2027, according to WilmerHale. A buyer that has to describe your data publicly will want your manifest to support it. See [the EU AI Act and training data transparency](https://sellmydatatoai.net/guides/eu-ai-act-training-data-transparency). ## How do you make a manifest verifiable? You make a manifest verifiable by binding it to one exact version of the data with file hashes, publishing how the sample was drawn and letting a third party check the claims. A manifest that cannot be checked is just a sales document. In the same Hugging Face thread, one reply recommended version-bound evidence packs that tie claims to an exact dataset revision, file manifest and hashes, and warned that "a seller-curated set of unusually good examples is a demonstration, not an audit sample". A September 2026 reply suggested a locked preview, where buyers run fixed queries without downloading the data, and third-party attestation. - Generate a SHA-256 hash for every delivered file and list it in the manifest - Hash the manifest itself and record that hash in the license - Draw the sample at random across the archive and state the method - Raise the version number if anything changes after a buyer has seen the manifest - Offer an acceptance test in which the buyer checks counts and hashes within a set number of days of delivery A random [representative sample](https://sellmydatatoai.net/glossary/representative-sample) does more for trust than any amount of description. Our guide to [evaluating dataset samples](https://sellmydatatoai.net/guides/how-to-evaluate-dataset-samples) shows what buyers do with one. ## How do you build a data manifest step by step? You build a data manifest step by step from a frozen export, so that every figure in it refers to the same files. The order below avoids the most common rework. 1. **Freeze the export**: Export from each system once, store the export read-only and call it version 1.0. Every count and hash must come from this copy. 2. **Map the sources**: List each system, the account or workspace exported, the date range and the teams covered. Note any months missing because of migrations or retention settings. 3. **Apply exclusions**: Remove HR, payroll, privileged, health and other excluded material, and record what was removed by category rather than in detail. 4. **De-identify and record the method**: Follow the steps in our [de-identification guide](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data) and record the tools, reviewers and error rates. 5. **Count and describe**: Run counting scripts on the final files. Record formats, schema, languages and how records link, such as replies to threads or attachments to tickets. 6. **Document rights per source**: Record who created the material, which staff and contractor IP assignments cover it, and any client contracts or vendor terms that restrict its use. 7. **Hash, sample and sign off**: Hash every file, draw the random sample, and have a named director sign the manifest off as accurate. ## What mistakes make buyers lose interest in a manifest? The mistakes that make buyers lose interest in a manifest are the ones that suggest the seller does not know what is in the data. The most common are: - Counts taken from dashboards rather than the export - Date ranges that hide gaps of several months - No list of exclusions, so the buyer assumes HR and client data are inside - Rights described as "owned by the company" with no detail on contractors or client contracts - A sample drawn from the best team or the best year - A manifest that changes after the sample without a new version number Each of these turns a quick review into a long negotiation, or ends it. If you want a buyer-ready manifest without building it alone, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). We will look at what you hold and tell you what a buyer will expect to see. You can also start with a [free data valuation](https://sellmydatatoai.net/data-valuation). ## Frequently asked questions ### Is a data manifest the same as a dataset card? No. A dataset card is a public description, usually published on a hub such as Hugging Face. A data manifest is a private inventory prepared for a sale, with record counts, exclusions, rights per source and file hashes that a buyer uses to price the data and check delivery. You can produce a dataset card from a manifest, but not the other way round. ### Should the manifest show real examples of records? Keep examples out of the manifest itself. Describe formats and schema in the manifest, and provide real records separately as a redacted sample under a mutual NDA. That way the manifest can circulate to several buyers early without exposing any content, and the sample goes only to buyers who have been vetted. ### What file format should a data manifest use? Most sellers produce a readable document for people and a machine-readable file for tools. JSON or YAML suits the machine-readable part, and the Data & Trust Alliance's metadata generator outputs JSON, XML, YAML or CSV. Croissant, built on schema.org, is supported by Kaggle, Hugging Face and OpenML if the buyer wants a standard format. ### Who should sign off the manifest? A director or another person who can bind the company should sign it off as accurate, because buyers rely on it and the license will usually refer to it. The person who ran the export and counts should sign a technical statement too. In a closed company, the liquidator, trustee or board acting for it signs. ### What happens if the delivered data does not match the manifest? That depends on the license. A well-drafted license sets an acceptance test, such as checking counts and file hashes within a fixed number of days, and says what happens on a mismatch: correction, a price adjustment or rejection. Without one, a mismatch becomes a dispute. Hashing every file makes the check quick and objective for both sides. ## Sources - [Is selling datasets way harder than building them, or is it just me?](https://discuss.huggingface.co/t/is-selling-datasets-way-harder-than-building-them-or-is-it-just-me/177440), Hugging Face forum, accessed November 2026 - [Datasheets for Datasets](https://huggingface.co/papers/1803.09010.md), Gebru and others, arXiv 1803.09010, accessed November 2026 - [Croissant: a metadata format for ML-ready datasets](https://research.google/blog/croissant-a-metadata-format-for-ml-ready-datasets/), Google Research, accessed November 2026 - [Cross-industry standards for data provenance in AI](https://www.helpnetsecurity.com/2024/07/22/saira-jesani-data-trust-alliance-data-provenance-standards/), Help Net Security, accessed November 2026 - [Commission presents template for general-purpose AI model providers to summarise the data used to train their model](https://digital-strategy.ec.europa.eu/en/news/commission-presents-template-general-purpose-ai-model-providers-summarise-data-used-train-their), European Commission, accessed November 2026 - [European Commission releases mandatory template for public disclosure of AI training data](https://wilmerhale.com/en/insights/blogs/wilmerhale-privacy-and-cybersecurity-law/european-commission-releases-mandatory-template-for-public-disclosure-of-ai-training-data), WilmerHale, accessed November 2026 - [How much is company data worth to AI?](https://ddiy.co/how-much-is-company-data-worth-to-ai/), ddiy.co, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Selling data from a closed or closing company URL: https://sellmydatatoai.net/guides/selling-data-from-a-closed-company Reviewed: 2026-11-08 You can sell data from a closed company if someone with legal authority still controls its property: the board while it winds down, or a liquidator, administrator or bankruptcy trustee. Buyers typically pay $10,000 to $100,000 per small company, according to SimpleClosure. Sell before a UK company is dissolved, because its property then passes to the Crown. ## Key takeaways - The board, a liquidator, an administrator or a bankruptcy trustee can sign, depending on the company's status. Former directors of a dissolved UK company cannot. - SimpleClosure's typical payments run $10,000 to $100,000 per company, with a few large estates such as Spirit Airlines drawing bids of $7.5 million to $12.5 million. - Data protection law still applies after closure. The Insolvency Service says selling data for a different use is likely to be incompatible with its original purpose. - Export everything before you cancel any software subscriptions, and check each tool's terms on bulk export. - Complete the license before dissolution, because a dissolved UK company's property passes to the Crown. Closed and closing companies are now a regular source of AI training data. MIT Technology Review reported on October 13, 2026 that AI labs are buying code and workspace data from shut-down startups, and that founder Keelan Miskell's winding-down crypto startup received about $25,000 for its codebase. The money is real, but a closing company has less time, fewer staff and more people with a claim on the proceeds than an operating one. This guide covers who can sign, what buyers pay, what data protection law still requires and the order to do things in. ## Can you sell data from a closed company? You can sell data from a closed company if someone with legal authority still controls the company's property. Who that is depends on the company's status, and a fully dissolved company is the hardest case. | Company status | Who can sign | What to check | | --- | --- | --- | | Solvent and winding down | The directors, with board approval | Articles, shareholder agreements and investor consent rights | | In liquidation (England and Wales) | The liquidator, who can sell any of the company's property under Schedule 4 to the Insolvency Act 1986 | Creditor interests and the liquidator's own approval process | | In administration | The administrator | Whether the sale fits the purpose of the administration | | US bankruptcy | The trustee or debtor in possession, with court approval for asset sales | Section 363 of the Bankruptcy Code and any consumer privacy ombudsman | | Dissolved (UK) | No one until the company is restored | Property passes to the Crown as bona vacantia under section 1012 of the Companies Act 2006 | The last row catches founders out. Section 1012 of the [Companies Act 2006](https://www.legislation.gov.uk/ukpga/2006/46/section/1012) says that when a company is dissolved, "all property and rights whatsoever" it held pass to the Crown as bona vacantia. That includes copyright and database rights in its records, so former directors have no authority to license them. If you are planning a closure, license the data before dissolution, not after. Authority to sign is the first check in our [seller vetting standard](https://sellmydatatoai.net/how-we-vet): the person signing can bind the company, or the trustee, liquidator or board can act for a closed company. ## What do buyers pay for a closed company's data? Buyers usually pay tens of thousands of dollars for a small closed company's data, and a few large estates have drawn bids in the millions. Published figures are thin, so treat them as ranges rather than prices. | Example | Figure | Source | | --- | --- | --- | | Typical SimpleClosure payment per company | $10,000 to $100,000 | SimpleClosure chief executive Dori Yona to Forbes, as reported by Fortune, September 2026 | | Winding-down crypto startup codebase | About $25,000, with a workspace data sale pending | MIT Technology Review, October 2026 | | Spirit Airlines records, Mercor bid | $7.5 million | Fortune, September 2026 | | Spirit Airlines records, Google bid | $10 million | Fortune, September 2026 | | Spirit Airlines records, micro1 proposal | $12.5 million | Fortune, September 2026 | Spirit's archive is not typical. Fortune reported that its asset schedule listed about 100 million emails and 500 million Microsoft Teams items, plus code and operational records, and that no sale had received court approval as of September 14, 2026. Brendan Mahony, chief executive of the closure service Sunset, told MIT Technology Review that his company values assets on headcount, years in operation, data volume and how the team used its tools. The same report said buyers quote very different numbers because their valuations rest on how many useful training tasks they expect the data to produce. Our guide to [what your data is worth](https://sellmydatatoai.net/guides/how-much-is-your-data-worth-to-ai) explains the main price factors. ## What data can a closed company sell? A closed company can usually sell its source code and its workspace data, and buyers value both most when they record real decisions. SimpleClosure's AssetHub lists two main categories: source code with repositories, documentation and tests, and workspace data from tools such as Slack, Google Drive, email, Notion and Figma. MIT Technology Review reported that buyers want varied, deep records of real situations, such as disagreements and judgment calls, and that suppliers leave HR records out. In the Spirit case, Fortune reported that customer profiles and lists were excluded and that passenger fields in included systems had to be de-identified. Our pages on [workplace communications data](https://sellmydatatoai.net/buy-ai-training-data/workplace-communications-data) and [source code datasets](https://sellmydatatoai.net/buy-ai-training-data/source-code-datasets) set out what buyers check in each. ## Does data protection law still apply after a company closes? Data protection law still applies after a company closes, and whoever runs the closure becomes responsible for the personal data in its records. Insolvency does not create an exemption. The Insolvency Service's technical manual says data protection law "will not prevent the sale of a database" in an insolvency, but that the buyer must use the data for its original purpose, and that "selling it to a business for a different use is likely to be incompatible with the original purpose". Training an AI model is a different use from the one staff and customers were told about, which is why personal data is usually excluded or de-identified before a closed company's records are licensed. UK regulators have acted on this before. In February 2020 the FCA, the ICO and the FSCS issued a joint statement warning that some insolvency practitioners had attempted to sell clients' personal data to claims management companies unlawfully. In the US, section 363(b)(1) of the Bankruptcy Code stops a trustee selling personally identifiable information against the debtor's privacy policy unless the sale is consistent with that policy, or a consumer privacy ombudsman is appointed and the court approves. In the Spirit case, ombudsman Lucy Thomson asked for more time to investigate micro1 as a possible buyer, Fortune reported. The wider rules are in our guide to [UK GDPR, EU GDPR and CCPA](https://sellmydatatoai.net/guides/gdpr-and-selling-data-for-ai-training). ## Who receives the money from a closed company's data sale? The money from a closed company's data sale goes to the company, and from there to creditors and shareholders in the legal order of priority. In a liquidation, the liquidator distributes it. MIT Technology Review reported that proceeds typically go to senior creditors and investors, not to the employees whose messages make up much of the data. That gap fuels objections. Spirit's flight attendants' unions and consumer privacy advocates filed objections to the sale, and the pilots' union argued that some safety records should stay confidential even after de-identification, according to Fortune. ## What risks are specific to a closing company? The risks specific to a closing company come from time pressure and from the fact that the company will not be around to stand behind its promises. - Lost access: software subscriptions lapse and admin accounts are deleted, so export before you cancel anything. - Tool terms: check each vendor's terms on export and reuse. Hunton Andrews Kurth reported in August 2025 that Slack's revised API terms ban bulk export of data accessed through its APIs and its use in large language models. - Third-party confidentiality: email and chat contain other companies' information covered by NDAs, as Frankfurt Kurnit Klein & Selz warned in October 2026. - Warranties: a dissolved company cannot pay out on an indemnity, so buyers may ask for a lower price, a holdback or insurance instead. - Records retention: some records must be kept for legal or tax reasons, and licensing a copy does not change those duties. ## How do you sell data from a closing company step by step? You sell data from a closing company step by step by securing the data first and the buyer second. The order below keeps your options open while access still exists. 1. **Freeze the data before shutting tools**: Take full exports of code repositories, email, chat, documents and tickets while admin access still exists. Store them read-only with a record of when and how they were made. 2. **Confirm who can sign**: Check whether the board, a liquidator, an administrator or a trustee holds authority, and get that confirmed in writing. 3. **Decide what to exclude**: Remove HR, payroll, health and privileged material, customer lists and anything covered by restrictive client contracts. 4. **De-identify what remains**: Follow a documented method, as set out in our [de-identification guide](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data). 5. **Build the manifest**: Record sources, counts, date ranges and exclusions using our [data manifest guide](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest). 6. **Approach buyers with redacted samples**: Share samples under a mutual NDA and compare offers on price, payment timing and what each buyer will do with the data. 7. **Sign, transfer and close out**: Sign the license, transfer a fixed copy, delete any working copies as the license requires, and only then complete the closure. > **Timing** In the UK, the license and transfer must complete before the company is dissolved. After dissolution its property belongs to the Crown and the company has to be restored before anyone can sell it. If your company is closing or already in an insolvency process, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). We will tell you whether buyers want what you hold and what needs to happen before the deadline. You can also compare closure-focused services such as [SimpleClosure AssetHub](https://sellmydatatoai.net/data-buyers/simpleclosure-assethub). ## Frequently asked questions ### Can I sell my startup's data after it has been struck off? Not directly. In the UK, a dissolved company's property and rights pass to the Crown as bona vacantia under section 1012 of the Companies Act 2006, so former directors have no authority to license them. The company would need to be restored to the register first. Take legal advice on restoration before talking to buyers, and sell before dissolution if you can. ### Can a liquidator sell company data to an AI buyer? A liquidator in England and Wales has power to sell any of the company's property under Schedule 4 to the Insolvency Act 1986, which can include rights in its records. The liquidator still has to comply with data protection law, so personal data is normally excluded or de-identified, and must act in the interests of creditors when choosing a buyer. ### How much will a closed startup get for its data? SimpleClosure's chief executive told Forbes that typical payments run $10,000 to $100,000 per company, as reported by Fortune in September 2026. MIT Technology Review reported one winding-down crypto startup received about $25,000 for its code. Larger estates with decades of records, such as Spirit Airlines, drew bids in the millions. ### Do former employees get any of the money? Usually not. Proceeds go to the company and then to creditors and shareholders in order of priority. MIT Technology Review reported in October 2026 that proceeds typically go to senior creditors and investors. Employees who are owed wages are creditors and rank under insolvency rules, but they have no separate share in a data sale. ### What should I export before shutting down? Export code repositories with full history, email, chat, documents, tickets and design files while admin accounts still work. Keep the exports read-only and note when and how each was made. Check each tool's terms first, because some restrict bulk export or AI use of data accessed through their APIs, as Slack's 2025 API terms do. ## Sources - [AI labs are buying data from dead startups to train their models](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models), MIT Technology Review, accessed November 2026 - [Google wants Spirit Airlines' data. Micro1 wants to pay more](https://fortune.com/2026/09/14/micro1-tries-to-beat-googles-bid-to-buy-bankrupt-spirit-airlines-data-for-12-5-million-highlighting-value-of-data-to-ai-companies/), Fortune, accessed November 2026 - [AssetHub](https://simpleclosure.com/product/assethub/), SimpleClosure, accessed November 2026 - [Companies Act 2006, section 1012](https://www.legislation.gov.uk/ukpga/2006/46/section/1012), legislation.gov.uk, accessed November 2026 - [Insolvency Act 1986, Schedule 4](https://www.legislation.gov.uk/ukpga/1986/45/schedule/4), legislation.gov.uk, accessed November 2026 - [Confidential information: client, customer and patient databases](https://www.insolvencydirect.bis.gov.uk/freedomofinformationtechnical/technicalmanual/Ch25-36/Chapter31/part10A/Part%207/Part%207.htm), The Insolvency Service, accessed November 2026 - [FCA, ICO and FSCS publish joint statement to insolvency practitioners and authorised firms](https://www.fca.org.uk/news/statements/fca-ico-and-fscs-publish-joint-statement-insolvency-practitioners-and-authorised-firms), Financial Conduct Authority, accessed November 2026 - [11 US Code section 363](https://www.law.cornell.edu/uscode/text/11/363), Legal Information Institute, Cornell Law School, accessed November 2026 - [Salesforce locks down Slack data: time to review your Slack API terms](https://natlawreview.com/article/salesforce-locks-down-slack-data-time-review-your-slack-api-terms), The National Law Review (Hunton Andrews Kurth), accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein & Selz, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How to sell source code to AI companies URL: https://sellmydatatoai.net/guides/sell-source-code-to-ai-companies Reviewed: 2026-11-08 You can sell source code to AI companies through code marketplaces, closure brokers, training-data companies and direct programs. Buyers pay most for private, human-written production code with full engineering history, and DataVendor says systems languages such as Rust, Go, C and C++ attract more. First confirm you own every line, strip secrets and customer data, and remove third-party code. ## Key takeaways - Labs pay more for code written mostly by people, for complex production systems and for systems languages, according to DataVendor. - Mercor's published range for software companies is $50,000 to $500,000 for 15 to 500 staff, while one small closed startup received about $25,000 for its code. - Contractors own the code they write unless they assigned it to you in writing, so check every agreement before you list a repository. - Scan the full commit history for secrets, rotate every credential found and replace real customer data in test fixtures. - Non-exclusive licenses let one repository earn from several buyers, which DataVendor says can beat a single exclusive deal. AI companies buy source code because coding models and agents learn more from real engineering history than from public snippets. DataVendor, a marketplace for code repositories run by HUD, says on its blog that "labs pay more for code written mostly by people", and that complex products and systems languages such as Rust, Go, C and C++ attract higher prices. Demand now reaches beyond specialist marketplaces. 404 Media reported in June 2026, as summarized by AI Weekly, that Google was running a confidential pilot to buy non-public production code and archived app projects from Android developers on a non-exclusive basis, with emails telling developers they could "retain 100% of your IP". Our page on [source code datasets](https://sellmydatatoai.net/buy-ai-training-data/source-code-datasets) covers what buyers check. ## Do AI companies buy source code? AI companies do buy source code, through four main routes: code marketplaces, closure brokers, training-data companies and direct programs run by large technology firms. | Route | Example | What it says it buys | | --- | --- | --- | | Code marketplace | DataVendor | Code repositories, raw data archives and RL tasksets, previewed as a file tree | | Closure broker | SimpleClosure AssetHub | Codebases, repositories, documentation and tests from companies shutting down | | Training-data company | Mercor | Software company data, with a published range of $50,000 to $500,000 for 15 to 500 staff | | Direct program | Google (reported pilot) | Non-public production code and archived Android app projects | The Mercor range comes from its referral page as recorded by ddiy.co on October 26, 2026. Code also feeds agent training. MIT Technology Review reported in October 2026 that Surge AI buys data to build reinforcement learning environments where AI agents practice tasks. ## What makes a codebase valuable to AI buyers? A codebase is valuable to AI buyers when it shows how real people built, tested and maintained a working product over time. DataVendor says labs value private engineering history because public repositories rarely preserve the full operational record. | Factor | Worth more | Worth less | | --- | --- | --- | | Authorship | Written mostly by people | Largely generated by AI tools | | Availability | Private and never published | Mirrored on GitHub or similar | | History | Commits, pull requests, review comments, issues, incident notes and runbooks | A single snapshot of the current code | | Complexity | Production systems with real users | Tutorials, prototypes and templates | | Language | Rust, Go, C and C++, according to DataVendor | Languages with plentiful public code | | Tests | Test suites and CI history showing what passed and failed | No tests | The context around the code often matters as much as the code. DataVendor lists pull requests, issue tickets, runbooks and incident notes as the material that explains why code changed, which is exactly what a coding agent needs to learn from. ## How much is source code worth to AI companies? Source code is worth anything from a few thousand dollars to hundreds of thousands, depending on size, history and exclusivity. The published figures are: - Mercor's listing puts software companies at $50,000 to $500,000 for 15 to 500 staff, per ddiy.co in October 2026 - A winding-down crypto startup received about $25,000 for its codebase, according to MIT Technology Review in October 2026 - SimpleClosure's typical payments run $10,000 to $100,000 per company across code and data, according to Fortune in September 2026 DataVendor cautions that a listed price or a large catalog does not establish what an asset will earn. Use the figures to decide whether a sale is worth preparing for, not as a quote. Our guide to [what your data is worth](https://sellmydatatoai.net/guides/how-much-is-your-data-worth-to-ai) explains the other factors. ## Do you own the code you want to sell? You own the code you want to sell only if your company holds the copyright in every part of it. In the UK, section 11(2) of the Copyright, Designs and Patents Act 1988 makes the employer the first owner of work an employee creates in the course of employment, subject to any agreement to the contrary. Contractors are different. Under section 11(1), the author is the first owner, so code written by freelancers or outside agencies belongs to them unless they assigned it to you. Under section 90(3) of the same Act, an assignment of copyright must be in writing and signed by the person assigning it. Check every contractor agreement before listing a repository. Client work is the other trap. Development contracts often assign the code to the client, and that code is not yours to license even if it sits in your repositories. A clean [chain of title](https://sellmydatatoai.net/glossary/chain-of-title) for each repository is what buyers ask for first. See our page for [software companies](https://sellmydatatoai.net/sell-your-data/software-companies) for the contract issues that come up most. ## What about open-source code in your repositories? Open-source code in your repositories has to be identified, recorded and either kept within its license terms or removed. DataVendor advises sellers to trace third-party dependencies and copied code, record each license, and remove material whose terms prohibit redistribution or commercial model training. - Strip vendored dependencies and copied libraries. Buyers can get public code elsewhere, and it adds license risk without adding value. - Run a software composition analysis tool to list the licenses found across the codebase. - Flag snippets copied from forums or documentation, which may carry their own license terms. - Record what you removed in the manifest so the buyer knows the package is your own work. ## What must you remove before selling code? Before selling code you must remove secrets, real personal data and anything that would expose live systems. SimpleClosure says every dataset it handles is processed for personal data before licensing, and that sensitive information and credentials are de-identified. | Item | Where it hides | What to do | | --- | --- | --- | | API keys and passwords | Config files, environment files, old commits | Scan the full history with a secret scanner and rotate every key found | | Customer data | Test fixtures, seed databases, logs, screenshots | Replace with synthetic values | | Staff identities | Commit author names and emails, code comments | Rewrite to consistent pseudonyms | | Infrastructure details | Hostnames, IP addresses, internal URLs | Replace or remove | | Unpatched vulnerabilities | Live production code | Fix them, or exclude the affected service | Rotate any credential that ever appeared in the repository, even after deleting it from the history. Once a copy leaves your control you cannot recall it. Our [de-identification guide](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data) covers the personal data side in detail. ## Should you license code exclusively or non-exclusively? Licensing code non-exclusively usually suits sellers with high-quality codebases, because the same repository can earn from several buyers. DataVendor says a marketplace that licenses to many labs can earn more than a single exclusive deal for such assets, and Google's reported pilot was non-exclusive. Exclusive deals pay more per buyer. AIData.Broker estimates an exclusive deal at about 2.5 to 5 times the price of a non-exclusive one, as reported by ddiy.co. If you are still trading, also consider whether a model trained on your code could help a competitor. Our guide to [exclusive and non-exclusive licenses](https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses) sets out the trade-off. ## How do you sell source code step by step? You sell source code step by step by choosing the right repositories, proving ownership, cleaning the history and only then talking to buyers. 1. **Pick the repositories**: Choose production systems with long histories. List each one with its language, age, number of contributors, commit count and whether it has tests. 2. **Check ownership**: Confirm employee and contractor IP assignments and check client contracts for every repository. 3. **Clean the history**: Remove vendored code, scan for secrets, rotate credentials, replace customer data and pseudonymize author details. 4. **Add the context**: Export pull requests, review comments, issues, runbooks and incident notes, and link them to the code they discuss. 5. **Write the manifest**: Record counts, languages, licenses found and removed, and file hashes. See [how to build a data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest). 6. **Share a preview**: Show a file tree and redacted samples under a mutual NDA before any buyer sees full code. 7. **License and deliver**: Sign the license, deliver a fixed copy rather than repository access, and record the hash of what was delivered. If you have production code with real history and want to know what buyers would pay for it, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). You can also read our profile of the code marketplace [DataVendor](https://sellmydatatoai.net/data-buyers/datavendor). ## Frequently asked questions ### Will AI companies buy code that was partly written with AI tools? They may, but it is worth less. DataVendor says labs pay more for code written mostly by people, because models learn little from output similar to their own. Be honest in the manifest about how much of the codebase was AI-assisted and when that started, since buyers can often detect it and will reprice if they find it. ### Can I sell code I wrote for a client? Usually not. Development contracts commonly assign copyright in the deliverables to the client, so the code is theirs even if it sits in your repositories. Check each contract. If you kept the rights or licensed the code to the client, you may be able to license it again, subject to confidentiality terms. Take advice before including any client project. ### Do I need to include the commit history? You do not have to, but history is often where the value is. DataVendor says labs value private engineering history because public repositories rarely preserve the full operational record. Pull requests, review comments and incident notes show why code changed. Clean the history of secrets and personal data rather than deleting it. ### Is it safe to sell code from a product I still run? It can be, if you remove secrets, rotate every credential, fix or exclude vulnerable services and replace customer data. The commercial question is harder: a model trained on your code may help competitors build similar features. Some sellers license older or retired systems only. Decide this before you approach buyers. ### How is source code usually priced? Pricing varies by buyer. Marketplaces such as DataVendor set a price per repository, training-data companies such as Mercor publish ranges per company, and closure brokers such as SimpleClosure take a percentage of the deal. Exclusive licenses cost buyers more, about 2.5 to 5 times a non-exclusive deal on AIData.Broker's estimate. ## Sources - [AI training data marketplaces](https://datavendor.ai/blog/ai-training-data-marketplaces), DataVendor, accessed November 2026 - [Google pays Play Store devs for code to train AI](https://aiweekly.co/alerts/google-pays-play-store-devs-for-code-to-train-ai), AI Weekly (reporting 404 Media), accessed November 2026 - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 - [How much is company data worth to AI?](https://ddiy.co/how-much-is-company-data-worth-to-ai/), ddiy.co, accessed November 2026 - [AI labs are buying data from dead startups to train their models](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models), MIT Technology Review, accessed November 2026 - [Google wants Spirit Airlines' data. Micro1 wants to pay more](https://fortune.com/2026/09/14/micro1-tries-to-beat-googles-bid-to-buy-bankrupt-spirit-airlines-data-for-12-5-million-highlighting-value-of-data-to-ai-companies/), Fortune, accessed November 2026 - [AssetHub](https://simpleclosure.com/product/assethub/), SimpleClosure, accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 11](https://www.legislation.gov.uk/ukpga/1988/48/section/11), legislation.gov.uk, accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 90](https://www.legislation.gov.uk/ukpga/1988/48/section/90), legislation.gov.uk, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How to license video and audio archives for AI training URL: https://sellmydatatoai.net/guides/sell-video-and-audio-for-ai-training Reviewed: 2026-11-08 You can license video and audio archives for AI training if you hold every rights layer in the recordings: the master, the music, the performances and the people on screen or on mic. Bloomberg reported in January 2025 that AI companies were paying $1 to $4 per minute of unused video. Clear rights, original masters and searchable metadata decide what an archive earns. ## Key takeaways - Reported video rates run from about $1 to about $7.25 per minute, according to Bloomberg's reporting on OpenAI, Google, Moonvalley and Adobe. - Owning the master is not enough. Music, performers, people in shot and third-party clips are separate rights layers. - Voice and faces can become biometric data once processed to identify someone, and Illinois law lists a voiceprint as a biometric identifier. - Start with rushes, b-roll and in-house recordings you own outright, then clear or remove the rest. - Deliver original masters with per-clip metadata and transcripts. Padded or compressed minutes are worth less. Video and speech models need very large volumes of recordings, and the developers building them now pay for licensed footage and audio. Bloomberg reported on January 10, 2025 that OpenAI, Google and Moonvalley were among the AI companies paying creators for unused video, at $1 to $4 per minute of footage. The market is not new. Bloomberg Law reported that Shutterstock's AI licensing business generated $104 million in 2023, and Shutterstock signed a six-year agreement with OpenAI covering images, videos and music in July 2023, according to TechCrunch. This guide covers who buys recordings, what they pay, which rights you need and how to prepare an archive. It describes what the law says, not how it applies to your archive, so take independent legal advice before you license. ## Who buys video and audio for AI training? Video and audio for AI training is bought by model developers directly and through licensing agents, platforms, brokers and marketplaces that assemble datasets for them. The table shows the main routes, using what each company says on its own site. | Route | Example | What it says it handles | | --- | --- | --- | | Licensing agent | Troveo | Video, audio and gaming datasets, citing more than 8 million hours of video and 4 million hours of audio | | Video licensing platform | Versos | Indexes libraries in place in cloud storage such as S3 or Cloudflare and delivers datasets to the licensee | | Data broker | Protege | Video, audio and speech, healthcare and other domain data for model builders | | Data marketplace | Defined.ai | Speech, text, image, video, music and sound effects datasets from vetted partners | | Training-data company | Labelbox Prism | Consented likeness and voice data, among other categories | | Content licensing platform | Shutterstock | Its own contributor library of images, video, music and 3D | Troveo's homepage says it sources from a network of more than 7,000 data providers and that 95% of its content is exclusive to Troveo. Versos's list of places to sell video for AI training names Google, Meta, NVIDIA, Runway, Moonvalley, Synthesia and Getty Images. Our profiles of [Troveo](https://sellmydatatoai.net/data-buyers/troveo) and [Versos](https://sellmydatatoai.net/data-buyers/versos) set out the published terms of each. ## How much do AI companies pay for video and audio? AI companies pay for video by the minute, and the public figures run from about $1 to about $7.25 per minute. Published prices for audio are much rarer. | Buyer | Reported rate | Source | | --- | --- | --- | | OpenAI, Google, Moonvalley and others | $1 to $4 per minute of unused footage | Bloomberg, January 2025 | | Adobe | Average about $2.62 per minute, up to about $7.25 | Bloomberg, April 2024 | | OpenAI | $1 to $4 per minute | Versos blog | | Adobe | About $3 per minute | Versos blog | The arithmetic shows why rights matter at scale. An archive of 10,000 hours holds 600,000 minutes, so the reported $1 to $4 range implies $600,000 to $2.4 million before any footage is excluded. That is a calculation on reported rates, not a quote, and buyers discount footage with open rights questions. For company audio, the closest published figure is Mercor's listing for call centers: $20,000 to $250,000 per company for operations with 50 to 5,000 agents, as recorded by ddiy.co in October 2026. Every published figure on our site is collected in the [AI data price index](https://sellmydatatoai.net/ai-data-price-index). ## What rights do you need to license recordings for AI training? To license recordings for AI training you need rights to every layer in them, not just ownership of the master file. UK copyright law treats the layers as separate works: section 1(1) of the Copyright, Designs and Patents Act 1988 lists musical and dramatic works separately from sound recordings, films and broadcasts. | Layer | Usually owned by | What to check | | --- | --- | --- | | Footage or audio master | The company that shot or commissioned it | Commission, production and terms of trade contracts | | Music | Publishers for the composition, labels for the recording | Whether the license covered only the original production | | Performances | Performers | Whether their consent reaches new uses | | People on screen or on mic | The individuals | Releases, privacy notices and biometric rules | | Inserted clips and stock | The agency or library | The agency's license terms | | Logos and artworks in shot | Third parties | Whether to blur, crop or exclude | Getphyllo's guidance on training data licenses calls this per-layer clearance, because video contains music, images, performances and trademarks that can each belong to someone different. In the UK, [section 182 of the Copyright, Designs and Patents Act 1988](https://www.legislation.gov.uk/ukpga/1988/48/section/182) makes recording a live performance without the performer's consent an infringement, and the consent in a performer's contract may be limited to broadcast or distribution. Raw rushes are often simpler to license than finished programs, because the music, graphics and archive inserts are added in the edit. Our page for [media companies and video archives](https://sellmydatatoai.net/sell-your-data/media-and-video-archives) covers commissioned programs and agency footage in more detail. ## Do speakers and people on screen need to consent? Whether speakers and people on screen need to consent depends on where they were recorded, the lawful basis relied on and what the buyer will train. The rules tighten sharply when voices or faces are processed to identify people. - UK GDPR Article 4 defines biometric data as personal data from specific technical processing that allows or confirms the unique identification of a person, and Article 9 makes biometric data processed for that purpose special category data. - Illinois' Biometric Information Privacy Act lists a voiceprint and a scan of face geometry among its biometric identifiers. - California Penal Code section 632 makes it an offense to record a confidential communication without the consent of all parties, so a recording must have been lawfully made before it can be lawfully licensed. - Tennessee's ELVIS Act, signed on March 21, 2024 and in force from July 1, 2024, added voice to the state's publicity rights, covering both a real voice and a simulation of it, according to Manatt. Buyers reflect these rules. Labelbox Prism says likeness and voice data need explicit, timestamped consent. If a buyer plans voice cloning or speaker identification, agree those uses separately in the license, or exclude them. Our page on [speech and audio data](https://sellmydatatoai.net/buy-ai-training-data/speech-and-audio-data) lists what buyers check. ## Can you license call recordings and meeting audio? You can license call recordings and meeting audio only where they were lawfully recorded and the people on them were told about uses that cover AI training, or the personal data is removed. For most company audio that means licensing redacted transcripts and dropping the audio itself. Demand exists. Polyshares lists call recordings among the operational records it licenses, and Mercor publishes a range for call centers. The risk is concentrated in the content: customers read out card numbers, addresses and health details. Cut payment segments, redact spoken identifiers in the transcript and treat the voice itself as personal data unless you can show otherwise. ## How do you prepare a video or audio archive for buyers? You prepare a video or audio archive by sorting it by rights position first, then describing what remains so a buyer can search and price it. The order below avoids clearing material you will later exclude. 1. **Inventory the archive**: Record total hours by format, year, resolution, language and storage location, and separate original masters from compressed copies. 2. **Sort by rights position**: Group material into owned outright (rushes, b-roll, in-house recordings), cleared with conditions, and uncleared. Offer the first group first. 3. **Remove or flag third-party layers**: Strip music beds, agency inserts and stock clips you do not control, or deliver muted or stem-separated audio where the buyer accepts it. 4. **Deal with the people**: Match releases and consents to clips, and exclude footage of identifiable people where you hold neither a release nor a lawful basis that covers training. 5. **Add searchable metadata**: Write shot descriptions, time-aligned transcripts and speaker labels. Versos notes that without a searchable structure even valuable assets can stay hidden. 6. **Count minutes honestly**: Remove duplicates, slates, black frames and long silences before counting, because buyers price per useful minute. 7. **Build the manifest and sample**: List hours, formats, rights per group and exclusions, then draw a random sample. Our guide to [building a data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest) sets out the fields. ## What formats do AI buyers expect for video and audio? AI buyers expect video and audio as original-resolution or lossless masters with per-clip metadata, not compressed web copies. The table reflects what buyers ask for on our [video training data](https://sellmydatatoai.net/buy-ai-training-data/video-training-data) and speech pages. | Asset | Preferred delivery | Metadata buyers ask for | | --- | --- | --- | | Video | MP4, MOV or ProRes masters at original resolution | Duration, resolution, frame rate, codec, shot description | | Audio | WAV or FLAC lossless files | Sample rate, channels, speaker IDs, language, recording environment | | Transcripts | Time-aligned JSON, SRT or TextGrid | Speaker labels and accuracy measured on a sample | | Rights records | One record per asset | Releases, consents, licenses in and exclusions | ## Should you license footage exclusively or through an agent? Licensing footage exclusively pays more per deal, while an agent or a non-exclusive license lets the same minutes earn more than once. Troveo pays a revenue share on every sale with no fees and asks for time-limited exclusivity, according to its business data page. A direct license with a model developer is usually a per-minute or fixed fee negotiated deal by deal. Check what you have already granted. YouTube's help page says channel owners and rights holders can allow third-party companies to train AI models on their public videos, and that the setting is off by default. If you turned it on, a buyer asking for exclusivity needs to know. Our guide to [exclusive and non-exclusive licenses](https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses) sets out the trade-off. ## What should a video or audio license say? A video or audio license should say which model uses are allowed, which are excluded and what happens if a rights holder later objects. Daniel Goldberg of Frankfurt Kurnit Klein & Selz notes that once data is used to train a model, "it cannot practically be pulled back out", so these terms have to be settled before transfer. - Permitted uses, such as video generation, recognition or captioning - Voice cloning, speaker identification and likeness generation excluded unless separately consented - Outputs that reproduce an identifiable performer or contributor prohibited - Deletion of raw copies by a fixed date - Withdrawal terms covering future training runs and models already trained, which Getphyllo lists among six things a training data license should cover We do not take custody of your recordings. Buyers see redacted samples under a mutual NDA, and full files move only after you sign a license directly with the buyer. If you hold an archive and want to know which parts buyers would take, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### How much is an hour of video worth to AI companies? At the $1 to $4 per minute Bloomberg reported in January 2025, an hour of licensable footage works out at $60 to $240. Bloomberg reported in April 2024 that Adobe paid an average of about $2.62 per minute, rising to about $7.25. Footage with uncleared music, people or inserts earns less or is excluded, so the figure applies to cleared minutes only. ### Can I license footage I have already posted on YouTube? Possibly, if you own every rights layer in it. YouTube's help page says creators and rights holders choose whether third-party companies may train AI models on their public videos, and the setting is off by default. A direct license to a buyer is a separate deal. Tell the buyer about any training permission you have already given, especially if they want exclusivity. ### Does music in my footage need separate clearance? Usually yes. Music has its own copyrights in the composition and the sound recording, and sync licenses are normally limited to the original production. Getphyllo's guidance calls for per-layer clearance for this reason. Many sellers deliver muted or stem-separated audio, or offer rushes recorded before music was added in the edit. ### Is recorded speech biometric data? Recorded speech is personal data where a speaker is identifiable, and it becomes biometric data once processed to identify or verify someone. UK GDPR Article 9 makes biometric data used for unique identification special category data, and Illinois' Biometric Information Privacy Act lists a voiceprint as a biometric identifier. Take advice before licensing audio for voice cloning or speaker identification. ### Can a business license its customer call recordings? Only with care. The calls must have been lawfully recorded, and the privacy information customers received must cover the new use, or the personal data must be removed. Most sellers license redacted transcripts rather than audio, cut payment segments and exclude calls with health or financial detail. California Penal Code section 632 shows how strict recording consent rules can be. ## Sources - [YouTubers Are Selling Their Unused Video Footage to AI Companies](https://news.bloomberglaw.com/ip-law/youtubers-are-selling-their-unused-video-footage-to-ai-companies), Bloomberg Law, accessed November 2026 - [Adobe is buying videos for $3 per minute to build AI model](https://www.bnnbloomberg.ca/adobe-is-buying-videos-for-3-per-minute-to-build-ai-model-1.2057606.amp.html), Bloomberg via BNN Bloomberg, accessed November 2026 - [Shutterstock's AI-licensing business generated $104 million](https://news.bloomberglaw.com/ip-law/shutterstocks-ai-licensing-business-generated-104-million), Bloomberg Law, accessed November 2026 - [Shutterstock expands deal with OpenAI to build generative AI tools](https://techcrunch.com/2023/07/11/shutterstock-expands-deal-with-openai-to-build-generative-ai-tools), TechCrunch, accessed November 2026 - [Top 10 places to sell your video content for AI training data](https://www.versos.ai/blog/sell-video-content-ai-training-data), Versos, accessed November 2026 - [Troveo | Real-World Data Powering Top Frontier Labs](https://www.troveo.ai/), Troveo, accessed November 2026 - [Business data](https://www.troveo.ai/business-data), Troveo, accessed November 2026 - [Prism](https://prism.labelbox.com/), Labelbox, accessed November 2026 - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 1](https://www.legislation.gov.uk/ukpga/1988/48/section/1), legislation.gov.uk, accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 182](https://www.legislation.gov.uk/ukpga/1988/48/section/182), legislation.gov.uk, accessed November 2026 - [Best public data providers for AI training and grounding](https://www.getphyllo.com/post/best-public-data-providers-ai-training-grounding), Getphyllo, accessed November 2026 - [UK GDPR Article 4: definitions](https://www.legislation.gov.uk/eur/2016/679/article/4), legislation.gov.uk, accessed November 2026 - [Biometric Information Privacy Act, 740 ILCS 14/10 definitions](https://lig.ilga.gov/legislation/ilcs/documents/074000140K10.htm), Illinois General Assembly, accessed November 2026 - [California Penal Code section 632](https://law.justia.com/codes/california/code-pen/part-1/title-15/chapter-1-5/section-632), Justia, accessed November 2026 - [Tennessee's ELVIS Act Expands Publicity Rights for Individuals' Voices & Protects Against AI Misuses](https://manatt.com/insights/newsletters/client-alert/tennessees-elvis-act-expands-publicity-rights-for), Manatt, accessed November 2026 - [Your content & third-party training](https://support.google.com/youtube/answer/15509945), YouTube Help, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein & Selz, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Is selling company data to AI safe? The real risks URL: https://sellmydatatoai.net/guides/is-selling-company-data-to-ai-safe Reviewed: 2026-11-08 Selling company data to AI can be done safely, but it is never risk-free, and the decision cannot be undone once a model is trained. The real risks are breaching privacy law or client contracts, licensing content you do not own, people being re-identified, a buyer misusing the data and a public backlash. Exclusions, de-identification, buyer vetting and tight license terms reduce most of them. ## Key takeaways - Training is a new purpose. UK GDPR requires you to tell people about it before the processing starts, and the FTC warns against quiet retroactive policy changes. - Client contracts and NDAs are the risk sellers most often miss. Email and chat hold other companies' confidential information. - You give warranties about rights you license, and the $1.5 billion Anthropic authors' settlement shows why buyers ask for them. - De-identified records can be re-identified, and the EDPB says models trained on personal data cannot in all cases be considered anonymous. - Some data should never be sold. Daniel Goldberg of Frankfurt Kurnit Klein & Selz advises excluding high-risk data entirely. AI buyers are spending heavily on company records. The Next Web reported in October 2026 that micro1 had committed $1bn over 12 months to buying and licensing operational data. The money is real, and so are the risks, because a data license is hard to reverse. Daniel Goldberg of Frankfurt Kurnit Klein & Selz put it plainly in October 2026: "Once data is used to train a model, it cannot practically be pulled back out." This guide sets out each risk honestly, what the law and regulators say about it, and the safeguards that reduce it. It is not legal advice for your company, so take independent advice before you sign anything. ## Is it safe to sell company data to AI companies? Selling company data to AI companies is safe enough for many sellers once the high-risk records are excluded and the rest is prepared properly, but no sale is risk-free. The table summarizes the six risks that come up most and who usually carries each one. | Risk | What can go wrong | Main safeguard | Who usually carries it | | --- | --- | --- | --- | | Privacy law | Personal data disclosed for a purpose people were never told about | Exclude or de-identify, and give notice first | The seller | | Client contracts and NDAs | Breach of confidentiality owed to clients, suppliers or partners | Contract review and exclusion of client material | The seller | | Copyright | Licensing material the company does not own | Chain of title review and removal of third-party content | The seller, through warranties | | Re-identification | People identified from records or from model outputs | Documented method, re-identification test, contract ban | Both sides | | Buyer misuse | Onward transfer, wider use or poor security | Buyer vetting and tight use limits | The seller in practice | | Reputation | Staff, clients or users object in public | Plan disclosure before signing | The seller | ## What are the privacy law risks of selling data to AI? The privacy law risk of selling data to AI is that training is a new purpose, and most records were collected without any mention of it. Goldberg notes that "most privacy policies, employee notices, and handbooks do not contemplate selling workplace communications for AI training". In the UK, Articles 13(3) and 14(4) of the UK GDPR require a controller that intends to process personal data for a new purpose to give people information about it before that processing starts. The ICO's guidance on the right to be informed says: "You must bring any new uses of an individual's personal data to their attention before you start the processing." The ICO's fining guidance puts the higher maximum fine under UK GDPR at £17.5 million or 4% of total worldwide annual turnover, whichever is higher. In the US, FTC staff wrote on February 13, 2024 that it may be unfair or deceptive for a company to adopt more permissive data practices through "a surreptitious, retroactive amendment to its terms of service or privacy policy". The post cites the FTC's 2004 case against Gateway Learning, the maker of Hooked on Phonics. Frankfurt Kurnit Klein & Selz adds that a transfer of personal information for value is a sale under the CCPA unless the data meets its de-identification standard. The safeguard is to keep personal data out of the package where you can. Our guides to [UK GDPR, EU GDPR and CCPA](https://sellmydatatoai.net/guides/gdpr-and-selling-data-for-ai-training) and [de-identification](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data) cover the detail. ## Can client contracts stop you selling company data? Client contracts can stop you selling company data, and they are the risk sellers most often miss. Goldberg warns that "email and chat contain other companies' confidential information, often covered by NDAs and commercial agreements", so an archive you hold is not automatically an archive you can license. Agencies and service firms face a sharper version. If you hold client personal data as a processor, Article 28(10) of the UK GDPR says a processor that determines the purposes and means of processing in breach of the regulation is treated as a controller for that processing. Licensing a client's data for training is a purpose the client never set. Buyers know this: Polyshares says its license does not cover data a company holds on behalf of its own clients, and Endless Data says it never licenses client books or privileged client matter content. - Read the confidentiality, data use and IP clauses in your largest client contracts first - Check supplier and partner NDAs for restrictions on disclosing their information - Check software vendor terms, since Hunton Andrews Kurth reported in August 2025 that Slack's revised API terms ban bulk export of data accessed through its APIs for use in large language models - Exclude client deliverables and client data unless a contract clearly allows the use ## What copyright risks come with selling data for AI? The copyright risk in selling data for AI is licensing material your company does not own, and then standing behind a warranty that says it does. Under section 11 of the Copyright, Designs and Patents Act 1988, the employer owns work made by employees in the course of employment, but contractors own what they create unless they assigned it in writing, as section 90(3) requires. Buyers take copyright seriously because their own exposure is large. A federal court in California gave final approval in July 2026 to a $1.5 billion settlement in the authors' suit over Anthropic's use of pirated books, about $3,000 for each of more than 400,000 books, according to Jurist. Buyers pass that risk back to sellers through contract terms. Troy Lieberman of Nixon Peabody wrote in October 2026 that remedies deserve as much attention as representations: "indemnities, liability caps, holdbacks, insurance, and termination rights", covering privacy claims, confidentiality breaches, IP disputes and the cost of removing data or retraining a model. He also recommends settling whether representations are qualified by knowledge, materiality or time period. Negotiate a cap and a knowledge qualifier before you sign. ## Can de-identified data be traced back to people? De-identified data can sometimes be traced back to people, either from the records themselves or from the model trained on them. The ICO's anonymization guidance says that "simply removing direct identifiers from a dataset is insufficient", and a 2019 study in Nature Communications by Luc Rocher and colleagues estimated that 99.98% of Americans would be correctly re-identified in any dataset using 15 demographic attributes. Models add a second route. Nicholas Carlini and colleagues showed at USENIX Security 2021 that they could extract verbatim training sequences from GPT-2, including names, phone numbers and email addresses. The European Data Protection Board's Opinion 28/2024 says AI models trained with personal data "cannot, in all cases, be considered anonymous". The safeguards are a documented method, a re-identification test run by someone who did not do the redaction, and a license term banning re-identification. In the UK, section 171 of the Data Protection Act 2018 makes knowingly or recklessly re-identifying de-identified personal data an offense. Our glossary entry on [re-identification risk](https://sellmydatatoai.net/glossary/re-identification-risk) explains the test. ## What happens if an AI buyer misuses your data? If an AI buyer misuses your data, your remedy is usually a contract claim after the damage is done, so the protection has to come before transfer. Misuse covers passing data to affiliates or model partners, using it beyond the agreed purpose, holding raw copies longer than agreed and weak security. Derived assets matter as much as the raw data. Nixon Peabody lists "cleaned datasets, annotations, embeddings, evaluation sets, fine-tuned models, model weights, and outputs" as layers a license should address one by one, and warns that deleting source files may not remove their influence from a trained model. It suggests a staged approach that starts with "a limited sample in a controlled environment". Buyers handle access differently. Scale AI says it anonymizes data inside the seller's environment before extraction, while other buyers connect extraction tools to company accounts. Goldberg advises companies to "avoid giving buyer tools direct access to company systems". Ask each buyer which model it uses before you share anything. Our own buyer vetting checks company identity, intended use, security and retention commitments, onward transfer limits and, for larger deals, ability to pay, before a buyer sees a sample. The full standard is on [how we vet](https://sellmydatatoai.net/how-we-vet). ## What are the reputational risks of selling data to AI? The reputational risk of selling data to AI is that staff, clients or users learn about the deal from someone else and object in public. Three published cases show the pattern. - Stack Overflow: after its partnership with OpenAI, some users overwrote their answers in protest and had their accounts suspended, The Register reported on May 9, 2024. - Wiley: The Bookseller reported in August 2024 that Wiley expected $44 million from AI rights deals and had confirmed there was no opt-out for authors. - Spirit Airlines: flight attendants' unions and consumer privacy advocates objected to the sale of the bankrupt airline's data, Fortune reported in September 2026. Staff messages carry a specific US risk. Goldberg notes that employee communications "may be protected concerted activity under Section 7 of the National Labor Relations Act", and that a collective bargaining agreement may restrict the sale. Our guide to [telling staff and clients about a data license](https://sellmydatatoai.net/guides/telling-staff-and-clients-about-a-data-sale) covers what to say and when. ## How do you reduce the risks before you sign? You reduce the risks before you sign by deciding what never leaves the company, then preparing, vetting and contracting in that order. 1. **Decide what never leaves**: Exclude HR and payroll files, health and benefits records, privileged material, payment card data and client data your contracts protect. 2. **Check notices and contracts**: Read the privacy notices in force when the data was collected, your client and supplier contracts, software vendor terms and any collective agreements. 3. **De-identify and test**: Apply a documented method and run a re-identification test on a random sample before any buyer sees it. 4. **Vet the buyer**: Confirm who the buyer is, what models the data will train, where it will be stored, who else may receive it and when raw copies are deleted. 5. **Share samples, not systems**: Show redacted samples under a mutual NDA, and transfer an exported copy rather than giving buyer tools access to live systems. 6. **Negotiate the license**: Set use limits, rules for derived assets, deletion dates, a ban on re-identification and capped, knowledge-qualified warranties. Our [licensing agreement checklist](https://sellmydatatoai.net/guides/ai-data-licensing-agreement-checklist) lists the clauses. 7. **Take independent legal advice**: Have your own lawyer review the license and the privacy position before you sign. ## When should you not sell company data? You should not sell company data when the value sits in material you cannot clear. That includes records whose worth depends on personal data, archives your client contracts protect, code or content you cannot show you own, and deals where the buyer will not accept limits on use or onward transfer. > **Rule of thumb** If you would not be comfortable explaining the sale to the people in the records, fix that first or leave those records out. We are a broker, not a law firm. You approve every buyer, every sample and every license term, and you can withdraw at any point before signing. For a view of the rights issues in your own records, start with a [free data valuation](https://sellmydatatoai.net/data-valuation) or [book a consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). Our guide to [whether you can legally sell company data](https://sellmydatatoai.net/guides/can-you-legally-sell-company-data-for-ai) covers the legal tests in more depth. ## Frequently asked questions ### Will my data leak if I sell it to an AI company? Leaks are possible at two points: in transfer and storage, and through the trained model. Researchers have extracted verbatim training text, including names and email addresses, from GPT-2. Reduce the risk by removing personal data before transfer, sending an exported copy rather than system access, and agreeing security, retention and deletion terms with the buyer in writing. ### Can I get my data back after a sale? You can require the buyer to delete raw copies by a set date, but you cannot practically remove data from a model already trained on it, as Daniel Goldberg of Frankfurt Kurnit Klein & Selz notes. Nixon Peabody recommends that licenses say whether deletion, retraining or model retirement applies if a problem emerges. Agree those terms before transfer, not after. ### Am I liable if the buyer misuses my data? You may be, depending on the law and the contract. Under UK GDPR, the company that discloses personal data needs a lawful basis for the disclosure and must tell people about it. Contractually, your exposure depends on the warranties and indemnities you give. Limit onward transfer and use in the license, vet the buyer first and take independent legal advice on liability. ### Is anonymized data safe to sell? Truly anonymized data falls outside data protection law, but the ICO says removing direct identifiers alone is not enough to achieve it. A 2019 Nature Communications study estimated 99.98% of Americans could be re-identified from 15 demographic attributes. Treat anonymization as a tested result, documented with a re-identification check, rather than a label. ### Do I need a lawyer to sell company data to AI? You should take independent legal advice before signing any data license. The questions involve privacy law, client contracts, copyright and employment law, and the license will contain warranties and indemnities that set your exposure. A broker can tell you what buyers check and help you prepare, but a broker is not a substitute for your own lawyer. ## Sources - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein & Selz, accessed November 2026 - [micro1 commits $1bn to enterprise data for AI training](https://thenextweb.com/news/micro1-1bn-enterprise-data-ai-training-citi-hercules), The Next Web, accessed November 2026 - [UK GDPR Article 13](https://www.legislation.gov.uk/eur/2016/679/article/13), legislation.gov.uk, accessed November 2026 - [UK GDPR Article 14](https://www.legislation.gov.uk/eur/2016/679/article/14), legislation.gov.uk, accessed November 2026 - [UK GDPR Article 28](https://www.legislation.gov.uk/eur/2016/679/article/28), legislation.gov.uk, accessed November 2026 - [The right to be informed](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/individual-rights/the-right-to-be-informed/), Information Commissioner's Office, accessed November 2026 - [The maximum amount of a fine under UK GDPR and DPA 2018](https://ico.org.uk/about-the-ico/our-information/policies-and-procedures/data-protection-fining-guidance/statutory-background/the-maximum-amount-of-a-fine-under-uk-gdpr-and-dpa-2018), Information Commissioner's Office, accessed November 2026 - [AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive](https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/02/ai-other-companies-quietly-changing-your-terms-service-could-be-unfair-or-deceptive), Federal Trade Commission, accessed November 2026 - [Polyshares homepage](https://polyshares.com/), Polyshares, accessed November 2026 - [Endless Data homepage and FAQ](https://endless-data.com/), Endless Data, accessed November 2026 - [Salesforce locks down Slack data: time to review your Slack API terms](https://natlawreview.com/article/salesforce-locks-down-slack-data-time-review-your-slack-api-terms), The National Law Review (Hunton Andrews Kurth), accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 11](https://www.legislation.gov.uk/ukpga/1988/48/section/11), legislation.gov.uk, accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 90](https://www.legislation.gov.uk/ukpga/1988/48/section/90), legislation.gov.uk, accessed November 2026 - [Judge approves record $1.5 billion AI copyright settlement involving Anthropic](https://www.jurist.org/news/2026/07/judge-approves-record-1-5-billion-settlement-involving-anthropic/), Jurist, accessed November 2026 - [Shopping for AI training data? Five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody, accessed November 2026 - [How do we ensure anonymisation is effective?](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/how-do-we-ensure-anonymisation-is-effective), Information Commissioner's Office, accessed November 2026 - [Estimating the success of re-identifications in incomplete datasets using generative models](https://pmc.ncbi.nlm.nih.gov/articles/PMC6650473), Nature Communications (Rocher, Hendrickx and de Montjoye, 2019), accessed November 2026 - [Extracting training data from large language models](https://www.usenix.org/conference/usenixsecurity21/presentation/carlini-extracting), USENIX Security Symposium 2021, accessed November 2026 - [Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models](https://edpb.europa.eu/system/files/2024-12/edpb_opinion_202428_ai-models_en.pdf), European Data Protection Board, accessed November 2026 - [Data Protection Act 2018, section 171](https://www.legislation.gov.uk/ukpga/2018/12/section/171), legislation.gov.uk, accessed November 2026 - [Data partnership](https://scale.com/data-partnership), Scale AI, accessed November 2026 - [Stack Overflow simply bans folks who don't want their advice used to train AI](https://www.theregister.com/2024/05/09/stack_overflow_banning_users_who/), The Register, accessed November 2026 - [Wiley set to earn $44m from AI rights deals, confirms 'no opt-out' for authors](https://thebookseller.com/news/wiley-set-to-earn-44m-from-ai-rights-deals-confirms-no-opt-out-for-authors), The Bookseller, accessed November 2026 - [Google wants Spirit Airlines' data. Micro1 wants to pay more](https://fortune.com/2026/09/14/micro1-tries-to-beat-googles-bid-to-buy-bankrupt-spirit-airlines-data-for-12-5-million-highlighting-value-of-data-to-ai-companies/), Fortune, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Telling staff and clients about a data license URL: https://sellmydatatoai.net/guides/telling-staff-and-clients-about-a-data-sale Reviewed: 2026-11-08 Tell staff and clients about a data license before the data leaves the company, not after. UK GDPR requires people to be told about a new purpose before the processing starts, and FTC staff warn that quiet, retroactive privacy policy changes may be deceptive. Settle what is included and excluded first, then tell each group in a planned order so nobody learns about the deal from the press. ## Key takeaways - UK GDPR Articles 13(3) and 14(4) require information about a new purpose before the new processing starts. - FTC staff say adopting more permissive data practices through a surreptitious, retroactive policy change may be unfair or deceptive. - In the US, staff messages may be protected concerted activity under Section 7 of the National Labor Relations Act, and union agreements may restrict a sale. - Most clients need to hear that their material is excluded, and the manifest should prove it. - You can keep the buyer and price confidential. You cannot keep the processing itself from the people whose personal data is in it. "Will my staff or clients find out?" is one of the first questions sellers ask. The better question is what they should hear, from whom and when, because the people in your records are also the people most likely to object. This guide sets out what UK and US rules say about telling people, which groups need to hear what, and a practical order for the announcement. It describes the law and published cases. It is not legal advice for your company, so take independent advice on your own notices and contracts. ## Do you have to tell staff and clients about a data license? You often have to tell staff and clients about a data license where it involves their personal data, and you usually should even where it does not. The legal duty attaches to personal data, while the commercial duty attaches to anyone whose confidential information or work is in the archive. | Group | Why they are affected | What usually applies | What they need to hear | | --- | --- | --- | --- | | Employees | Their emails, chat, code and tickets | UK GDPR Articles 13 and 14, and NLRA Section 7 in the US | What is included and excluded, how it is de-identified and how to object | | Former employees | Their messages remain in the archive | The same transparency duties for their personal data | Updated privacy information where you can reach them | | Clients | Their confidential information and deliverables | Contract confidentiality clauses, and UK GDPR Article 28 for data you process for them | Usually that their material is excluded | | Customers and users | Their personal data in tickets, CRM records and calls | Your privacy notice, CCPA rules and FTC guidance | Updated privacy information and any opt-out route | | Unions | Members' communications | The collective bargaining agreement | Consultation before signing | | Board and investors | Authority and consent rights | Articles and shareholder agreements | The terms, for approval | ## What does UK GDPR require you to tell people? UK GDPR requires you to tell people about a new purpose before you start processing their data for it. Article 13(3) covers data collected from the person and Article 14(4) covers data obtained from elsewhere, such as client contacts named in your email. Both say the controller must provide information on that other purpose "prior to that further processing". The ICO's guidance on the right to be informed says: "You must bring any new uses of an individual's personal data to their attention before you start the processing." The ICO also treats anonymization as processing of personal data in its own right, so preparing records for a license is itself covered. The information must include people's rights, among them the right to object. Under Article 21(1), where processing rests on legitimate interests, a person can object on grounds relating to their particular situation, and the controller must stop unless it shows compelling legitimate grounds. Build an objection route into the plan. Our guide to [UK GDPR, EU GDPR and CCPA](https://sellmydatatoai.net/guides/gdpr-and-selling-data-for-ai-training) covers lawful basis in more depth. ## What do US rules say about telling customers? US rules focus on the promises you made when you collected the data. FTC staff wrote on February 13, 2024 that "a business that collects user data based on one set of privacy commitments cannot then unilaterally renege on those commitments after collecting users' data". The FTC post cites its 2004 case against Gateway Learning, which changed its privacy policy to share consumer data with third parties without telling consumers, and a 2023 case against a genetic testing company over a retroactive expansion of the third parties it could share data with. Frankfurt Kurnit Klein & Selz adds that companies must comply with the privacy statements in effect when the data was collected. The practical effect is that a new notice rarely fixes old data. Compare each source against the notice in force when it was collected. Where the notice does not cover licensing, exclude the personal data or de-identify it, and remember that the CCPA treats a transfer of personal information for value as a sale unless it meets its de-identification standard, according to Frankfurt Kurnit Klein & Selz. ## Should you tell employees before or after signing? You should tell employees before signing, once the scope is settled but before any data leaves. UK transparency duties apply before the processing starts, and in the US the order can matter for labor law as well as goodwill. Daniel Goldberg of Frankfurt Kurnit Klein & Selz notes that employee communications "may be protected concerted activity under Section 7 of the National Labor Relations Act", and that where staff are unionized, "the collective bargaining agreement may restrict the sale, and the company may have bargaining obligations". He also observes that most employee notices and handbooks do not contemplate selling workplace communications for AI training. Staff objections are not hypothetical. In the Spirit Airlines bankruptcy, flight attendants' unions objected to the sale of the airline's data, and the pilots' union argued that some safety records should stay confidential even after de-identification, Fortune reported in September 2026. ## What should you tell clients about a data license? You should tell clients what you are excluding, and for most sellers that is all of their material. Buyers expect it: Polyshares says its license does not cover data a company holds on behalf of its own clients, and Endless Data says it never licenses client books or privileged client matter content. If client material is in scope, such as de-identified deliverables, check each contract and get written consent where the contract is unclear. Where you process a client's personal data as its processor, UK GDPR Article 28(10) says a processor that determines the purposes and means of processing in breach of the regulation is treated as a controller for it. Contract restrictions are one of the eight checks in our [seller vetting standard](https://sellmydatatoai.net/how-we-vet). Professional firms carry extra confidentiality duties to clients. Our pages for [law firms](https://sellmydatatoai.net/sell-your-data/law-firms) and [marketing agencies](https://sellmydatatoai.net/sell-your-data/marketing-agencies) cover the client issues that come up most in each. ## What do public data deals show about disclosure? Public data deals show that the way people learn about a license shapes their reaction as much as the terms do. These cases involve publishers and platforms rather than company records, but the people affected reacted in similar ways. | Case | What was reported | What it shows | | --- | --- | --- | | Stack Overflow and OpenAI, May 2024 | Some users overwrote their answers in protest and had accounts suspended (The Register) | Contributors who feel unasked will act | | Wiley, August 2024 | $44 million expected from AI rights deals, with no opt-out for authors (The Bookseller) | A no-opt-out position became the headline | | HarperCollins, November 2024 | Author opt-in, $5,000 per title split equally, three-year term (Publishers Weekly) | Even opt-in terms drew questions about who the buyer was | | Spirit Airlines, September 2026 | Unions and consumer privacy advocates objected to the data sale (Fortune) | Staff object even when records are de-identified | The HarperCollins case is worth reading closely. Publishers Weekly reported that Robert Gottlieb of Trident Media said most of his authors were uneasy mainly because they did not know who the buyer was. Saying what kind of buyer will receive the data, and what it may do, answers much of that concern without naming the company. ## Can you keep a data license confidential? You can keep a data license's commercial terms confidential, including the buyer's name and the price, but not the processing itself from the people whose personal data is in it. Those are separate duties. Several buyers offer confidentiality on the commercial side. Troveo says a seller's company is never named publicly and that buyers are introduced confidentially only after an agreement is signed. Polyshares keeps company names confidential on its track record page and does not disclose which labs take which data. With us, buyers see only redacted samples under a mutual NDA before a license is signed. ## How do you plan the announcement? You plan the announcement by fixing the scope first and the message second, then telling each group in order of how directly they are affected. 1. **Settle scope and exclusions**: Decide which systems, date ranges and record types are in, and record what is out, such as HR files, private messages, health data and client material. 2. **Check what people were told**: Collect the privacy notices, employee handbooks and client contracts in force when the data was collected, and note any gaps. 3. **Get internal approvals**: Brief the board and investors with consent rights, and open consultation with any union whose agreement may apply. 4. **Update privacy information**: Publish updated privacy information before any processing for the license begins, including how to object. 5. **Tell employees directly**: Explain the license in a meeting or message from a director, with written answers to common questions and a named contact. 6. **Write to affected clients**: Confirm that their material is excluded, or ask for written consent where it is not. 7. **Prepare a public line and keep records**: Agree a short statement in case the deal becomes public, and keep a record of who was told what and when. ## What should the message to staff say? The message to staff should say plainly what is being licensed, what is not and what protections apply. Staff will read it as a test of whether the company has thought about them. - Why the company is licensing the data, in one or two sentences - Which systems and years are included - What is excluded, such as HR files, payroll, health information and direct messages if you leave them out - How names and other identifiers are removed, and how the method was tested - What kind of buyer will receive the data and what the license lets it do - How to ask questions or object, and by when - Whether staff share in the proceeds, stated honestly either way > **Rule of thumb** Write the staff message before you sign. If it is hard to write honestly, change the scope rather than the message. Our guide to [whether selling company data to AI is safe](https://sellmydatatoai.net/guides/is-selling-company-data-to-ai-safe) covers the wider risks, and the [de-identification guide](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data) explains how to show staff their names have been removed. To plan a license with disclosure built in, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### Do I need employee consent to license their emails for AI training? Not always, because consent is only one lawful basis under UK GDPR. You do need to tell employees about the new purpose before processing starts, and if you rely on legitimate interests they can object under Article 21. In the US, Section 7 of the National Labor Relations Act and any union agreement may also apply. Take independent advice on your position. ### Do former employees need to be told about a data license? Transparency duties under UK GDPR apply to the personal data you process, whether or not the person still works for you. In practice, many sellers exclude or de-identify former staff records and update the privacy information on their website. Where former employees are easy to reach, telling them directly avoids surprises. Take advice on what your notices require. ### Will the buyer be named publicly? Not usually, unless you agree to it. Troveo says a seller's company is never named publicly, and Polyshares keeps company names confidential. Large publisher deals are often announced, but private company licenses rarely are. You can tell staff what kind of buyer will receive the data and what it may do without naming the company. ### What if an employee objects to the license? Take the objection seriously and respond in writing. Where you rely on legitimate interests, UK GDPR Article 21 requires you to stop processing that person's data unless you can show compelling legitimate grounds. The simplest answer is often to exclude that person's messages from the package and record the exclusion in the manifest. ### Should we tell clients even if none of their data is included? Usually yes, for your largest clients at least. Client emails and documents sit in most company archives, and a client who hears about a license secondhand may assume the worst. A short letter confirming that their material is excluded, backed by the exclusions in your data manifest, costs little and protects the relationship. ## Sources - [UK GDPR Article 13](https://www.legislation.gov.uk/eur/2016/679/article/13), legislation.gov.uk, accessed November 2026 - [UK GDPR Article 14](https://www.legislation.gov.uk/eur/2016/679/article/14), legislation.gov.uk, accessed November 2026 - [UK GDPR Article 21](https://www.legislation.gov.uk/eur/2016/679/article/21), legislation.gov.uk, accessed November 2026 - [UK GDPR Article 28](https://www.legislation.gov.uk/eur/2016/679/article/28), legislation.gov.uk, accessed November 2026 - [The right to be informed](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/individual-rights/the-right-to-be-informed/), Information Commissioner's Office, accessed November 2026 - [Introduction to anonymisation](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/), Information Commissioner's Office, accessed November 2026 - [AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive](https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/02/ai-other-companies-quietly-changing-your-terms-service-could-be-unfair-or-deceptive), Federal Trade Commission, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein & Selz, accessed November 2026 - [Google wants Spirit Airlines' data. Micro1 wants to pay more](https://fortune.com/2026/09/14/micro1-tries-to-beat-googles-bid-to-buy-bankrupt-spirit-airlines-data-for-12-5-million-highlighting-value-of-data-to-ai-companies/), Fortune, accessed November 2026 - [Polyshares homepage](https://polyshares.com/), Polyshares, accessed November 2026 - [Track record](https://polyshares.com/track-record), Polyshares, accessed November 2026 - [Endless Data homepage and FAQ](https://endless-data.com/), Endless Data, accessed November 2026 - [Business data](https://www.troveo.ai/business-data), Troveo, accessed November 2026 - [Stack Overflow simply bans folks who don't want their advice used to train AI](https://www.theregister.com/2024/05/09/stack_overflow_banning_users_who/), The Register, accessed November 2026 - [Wiley set to earn $44m from AI rights deals, confirms 'no opt-out' for authors](https://thebookseller.com/news/wiley-set-to-earn-44m-from-ai-rights-deals-confirms-no-opt-out-for-authors), The Bookseller, accessed November 2026 - [Agents, authors question HarperCollins AI deal](https://www.publishersweekly.com/pw/print/20241125/96533-agents-authors-question-harpercollins-ai-deal.html), Publishers Weekly, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # AI data marketplaces, brokers and direct deals compared URL: https://sellmydatatoai.net/guides/ai-data-marketplaces-compared Reviewed: 2026-11-08 AI data marketplaces, brokers and direct deals differ mainly in who sets the price, who prepares the data and who decides where it goes. Marketplaces such as Opendatabay and Datarade let you list and price a dataset for a published commission. Direct buyers such as Polyshares pay a fixed sum with no fee. Brokers and licensing agents sit between, and most are paid only when a deal completes. ## Key takeaways - Marketplaces publish commissions, from Opendatabay's 30% to 5% sliding scale to DataVendor's 20% on withdrawal. Most brokers and buyers do not publish a rate. - A zero fee is not free. Direct buyers that license on to labs earn their margin inside the price they offer. - Exclusivity varies more than price: Polyshares' published deals are all perpetual exclusive, while Endless Data licenses non-exclusively. - Training-data companies and direct buyers usually de-identify for you, while marketplaces such as FileYield leave compliance to the seller. - Ask every route the same questions in writing, about fees, onward transfer, exclusivity, timing and deletion. A company with data to license has four broad routes: list it on a marketplace, use a broker or licensing agent, sell to a training-data company, or sign directly with a buyer that licenses on to AI labs. Each publishes different terms, and the differences matter more than the headline payouts. This guide compares the routes using only what each company publishes on its own site, as checked for our [data buyers directory](https://sellmydatatoai.net/data-buyers) in November 2026. Sell My Data To AI is a broker, so we are one of the options compared. We describe our own terms in the same way as everyone else's and do not rank the routes. ## What is the difference between a data marketplace, a broker and a direct buyer? The difference between a data marketplace, a broker and a direct buyer is who sets the price and who you sign with. On a [data marketplace](https://sellmydatatoai.net/glossary/data-marketplace) you usually set the price and deal with many buyers. A [data broker](https://sellmydatatoai.net/glossary/data-broker) finds a buyer and negotiates for you. A direct buyer signs with you itself and pays a fixed sum. | Route | How it works | Who sets the price | Examples | | --- | --- | --- | --- | | Data marketplace | Lists datasets for buyers to browse, request or buy | Usually the seller | Opendatabay, Datarade, AWS Data Exchange, DataVendor, FileYield, Defined.ai | | Data broker | Prepares or packages data and finds buyers, usually paid on success | Agreed deal by deal | SellMyData, SimpleClosure AssetHub, Protege, Sell My Data To AI | | Licensing agent | Licenses your data to several buyers on your behalf | The agent, with buyers | Troveo | | Training-data company | Licenses records and supplies datasets to AI labs | The buyer, often from published ranges | micro1, Mercor, Scale AI, Handshake AI, Labelbox Prism | | Direct buyer | Pays you a fixed sum and licenses on to labs itself | The buyer | Polyshares, Endless Data | The lines blur. Endless Data says it places data with several labs at once and names micro1, Mercor and Handshake as examples, so it buys from sellers and sells to other buyers. SimpleClosure AssetHub handles both licenses and outright sales. ## How do fees compare across marketplaces, brokers and buyers? Fees compare unevenly because marketplaces publish commission rates, while most brokers and buyers either charge the seller nothing or do not publish a rate. The table uses each company's own wording. | Company | Route | Fee to seller | | --- | --- | --- | | Opendatabay | Marketplace | Free to list, commission from 30% on sales up to $100 falling to 5% above $250,000 | | Datarade | Marketplace | $0 a year plus 30%, $6,000 a year plus 20%, or $12,000 a year plus 15% | | DataVendor | Marketplace | 20% of vendor earnings, charged on withdrawal | | AWS Data Exchange | Marketplace | AWS Marketplace service fees, rate not stated on the pages we checked | | FileYield | Marketplace | Free to list, facilitation fees disclosed when you list, request or accept a deal | | SellMyData | Broker | Commission when a sale closes, rate not published | | SimpleClosure AssetHub | Broker | Percentage on a successful sale or license, rate not published | | Troveo | Licensing agent | No fees, paid from its revenue share | | Polyshares | Direct buyer | No fee and no commission | | Endless Data | Direct buyer | None, Endless is paid by the labs | | Mercor | Training-data company | None, Mercor covers extraction, anonymization and transfer | | Sell My Data To AI | Broker | Success fee only when a deal completes, agreed in writing before any introduction | A zero fee does not mean the intermediary earns nothing. Endless Data says it is paid by the labs, and Polyshares licenses the data on to AI labs itself, so their margin sits inside the price they offer you. Compare the net amount you receive, not the fee line. Our [fees page](https://sellmydatatoai.net/fees) sets out how our own success fee works. ## Which route pays sellers the most? No route reliably pays sellers the most, because every published figure is a range and the final price depends on the data. The ranges below are the companies' own, and most are marked as estimates. | Company | Route | Published payout | | --- | --- | --- | | micro1 | Training-data company | $100,000 to $500,000 for qualified partnerships, up to more than $3 million | | Handshake AI | Training-data company | $100,000 to more than $4 million (headline) | | Endless Data | Direct buyer | Up to $4 million (indicative) | | Polyshares | Direct buyer | $100,000 to $2 million and above typical, four published deals of $110,000 to $450,000 | | Labelbox Prism | Training-data company | Around $50,000 to more than $1 million | | Scale AI | Training-data company | $10,000 to more than $1 million (illustrative) | | SimpleClosure AssetHub | Broker | Tens to hundreds of thousands of dollars | | Troveo | Licensing agent | Full-company deals start at six figures | | Marketplaces | Marketplace | Seller sets the price | Exclusivity moves price as much as route does. AIData.Broker estimates an exclusive deal at about 2.5 to 5 times a non-exclusive one, as reported by ddiy.co in October 2026. Every figure on our site, with its source, is in the [AI data price index](https://sellmydatatoai.net/ai-data-price-index). ## How do exclusivity and payment timing compare? Exclusivity and payment timing compare more sharply than fees, and they are the terms most often left unpublished. Where a company does publish them, they range from perpetual exclusivity to none, and from 7 days to about 60. | Company | Exclusivity | Payment timing | | --- | --- | --- | | Polyshares | All four published deals perpetual and exclusive | Offer in one to two weeks, then net 30 to 60 days | | Labelbox Prism | Exclusive by default | Upfront fee at signing plus revenue share, first call to first wire under 21 days | | Troveo | Time-limited, with exit terms | Revenue share on every sale | | Endless Data | Non-exclusive | Fixed sum paid after signing, preparation and transfer | | FileYield | Seller chooses | Arranged between seller and buyer | | micro1 | Not published | Within 30 days of acceptance, or about 60 days on its FAQ | | Mercor | Not published | Typically 2 to 4 weeks after the signed agreement | | SellMyData | Not published | Target of 7 days after a sale closes | | Opendatabay | Not published | Within 30 days of the end of the month | | AWS Data Exchange | Not published | Monthly, after funds are collected from subscribers | A perpetual exclusive license ends your ability to license the same records again. A non-exclusive one lets a single archive earn from several buyers. Our guide to [exclusive and non-exclusive licenses](https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses) sets out the trade-off. ## Who prepares and de-identifies the data on each route? Training-data companies and direct buyers usually prepare and de-identify the data for you, marketplaces usually leave it to you, and brokers vary. The access model differs too, which affects your security exposure. - Mercor says an automated process masks more than 60 categories of sensitive identifiers, and its process links company accounts to its extraction tooling. - Scale AI says it anonymizes data inside the seller's own environment before anything is extracted. - micro1 says it removes or transforms personal information before sharing data with third parties. - Polyshares says a specialist personal data partner strips identifying fields before Polyshares or any lab sees the record. - SimpleClosure says every AssetHub dataset is processed for personal information and credentials before licensing. - SellMyData says the lab receives the de-identified dataset, never the seller's raw data. - FileYield says sellers handle data transfer, personal data and compliance directly with the buyer. - Datarade says data on its marketplace may not contain unanonymized sensitive personal data. When a buyer de-identifies for you, ask for its written method and test results before signing. We do not take custody of raw data. Buyers see redacted samples under a mutual NDA, and full data moves only after you sign a license directly with the buyer. ## How much control do you keep over who receives the data? How much control you keep over who receives the data depends on who you sign with. If your contract is with an intermediary that licenses on to labs, onward transfer is the point of the deal, so the question becomes which labs and on what terms. - SellMyData says you approve every buyer and every term. - DataVendor's buy-now purchases are anonymous to the vendor, while offers name the buyer organization so you can choose whom to accept. - Polyshares does not disclose which labs take which data, and says use is bounded by the license. - micro1 says it licenses data only to labs and enterprises in the US and allied countries. - Handshake AI says you review who may receive the data and how it may be used before committing. - With us, you approve every buyer, every sample and every license term, and you can withdraw at any point before signing. ## Which route suits which seller? Which route suits you depends on what you hold, how big your company is and how much of the work you want to do yourself. The table matches common situations to the routes whose published terms fit them. | Your situation | Routes whose terms fit | Why | | --- | --- | --- | | A packaged dataset you can price and sell repeatedly | Opendatabay, Datarade, AWS Data Exchange | Published commissions and many buyers | | Private codebases or RL tasksets | DataVendor, SimpleClosure AssetHub | Built for code and engineering assets | | 30 or more staff with years of operational records | micro1, Endless Data, Scale AI (40 or more preferred) | Published size thresholds | | A small company below headcount thresholds | Polyshares, SellMyData | No headcount minimum published | | A startup winding down | SimpleClosure AssetHub | Serves operating and closing companies | | A video or audio archive | Troveo, Versos, Protege | Media-focused licensing | | Ready to run your own diligence and contracts | FileYield | Direct negotiation with buyers | | A sensitive archive where you want to compare buyers | A broker such as SellMyData or Sell My Data To AI | Offers compared before you commit | ## What should you ask any marketplace, broker or buyer? You should ask any marketplace, broker or buyer the same set of questions in writing, so that the answers can be compared line by line. 1. What is the fee or margin, and how is it calculated? 2. Who signs the license with me, and who else will receive the data? 3. Is the license exclusive, for how long, and with what exit terms? 4. When am I paid, and what has to happen first? 5. Who de-identifies the data, by what method, and can I see the test results? 6. Will you take an exported copy, or do your tools need access to our systems? 7. When are raw copies deleted, and what happens to trained models if a problem emerges? 8. Do I approve each buyer and each license term? Our [licensing agreement checklist](https://sellmydatatoai.net/guides/ai-data-licensing-agreement-checklist) covers the clauses behind each question. For a view of which routes and buyers fit your records, start with a [free data valuation](https://sellmydatatoai.net/data-valuation) or [book a consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### Is a data broker better than a data marketplace? Neither is better in general. A marketplace suits a packaged dataset you can price yourself and sell many times, at a published commission such as Opendatabay's 30% to 5% scale. A broker suits a one-off or sensitive archive where finding the right buyer and negotiating terms matters more. Compare the net amount, the exclusivity and who carries the preparation work. ### Do AI data marketplaces check that I own the data? Marketplaces generally place responsibility on the seller. FileYield says sellers handle personal data and compliance directly with the buyer, and Datarade bars unanonymized sensitive personal data from listings. Some review listings before they go live. Whatever the route, buyers will ask about ownership and chain of title, so check your rights before you list. ### Why do some buyers charge sellers no fee? Buyers that charge no fee earn elsewhere. Endless Data says it is paid by the labs it licenses to, and Polyshares licenses data on to AI labs itself, so their margin sits inside the price they offer. Mercor covers extraction and anonymization costs. A zero fee can still be the best net offer, but compare the amount you actually receive. ### Can I use more than one route at the same time? You can if every license you sign is non-exclusive. Endless Data licenses non-exclusively, while all four of Polyshares' published deals were perpetual and exclusive, and Labelbox Prism is exclusive by default. Troveo asks for time-limited exclusivity. Check the exclusivity terms before sharing samples with a second route, and record each license in your data manifest. ### How long does each route take to pay? Published timings range widely. SellMyData targets payment within 7 days of a sale closing, Mercor typically pays within 2 to 4 weeks of a signed agreement, and Labelbox Prism says first call to first wire takes under 21 days. Polyshares pays net 30 to 60 days, and FileYield says enterprise deal cycles run from a few weeks to a few months. ## Sources - [Opendatabay homepage](https://www.opendatabay.com/), Opendatabay, accessed November 2026 - [Fees and pricing](https://www.opendatabay.com/legal/fees), Opendatabay, accessed November 2026 - [Datarade homepage](https://datarade.ai/), Datarade, accessed November 2026 - [Apply as a data provider](https://providers.datarade.ai/apply), Datarade, accessed November 2026 - [FAQ](https://docs.datavendor.ai/overview/faq.md), DataVendor, accessed November 2026 - [Provider financials](https://docs.aws.amazon.com/data-exchange/latest/userguide/provider-financials.html), Amazon Web Services, accessed November 2026 - [Terms of service](https://www.fileyield.com/terms), FileYield, accessed November 2026 - [How it works](https://www.fileyield.com/how-it-works), FileYield, accessed November 2026 - [SellMyData homepage and FAQ](https://www.sellmydata.com/), SellMyData, accessed November 2026 - [AssetHub](https://simpleclosure.com/product/assethub/index.html), SimpleClosure, accessed November 2026 - [Business data](https://www.troveo.ai/business-data), Troveo, accessed November 2026 - [Polyshares homepage](https://polyshares.com/), Polyshares, accessed November 2026 - [Track record](https://polyshares.com/track-record), Polyshares, accessed November 2026 - [Endless Data homepage and FAQ](https://endless-data.com/), Endless Data, accessed November 2026 - [Data monetization](https://www.mercor.com/data/), Mercor, accessed November 2026 - [Data partnerships](https://www.micro1.ai/data-partnerships), micro1, accessed November 2026 - [Data partnership](https://scale.com/data-partnership), Scale AI, accessed November 2026 - [Data partnerships](https://joinhandshake.com/ai/data-partnerships/), Handshake, accessed November 2026 - [Prism](https://prism.labelbox.com/), Labelbox, accessed November 2026 - [How much is company data worth to AI?](https://ddiy.co/how-much-is-company-data-worth-to-ai/), ddiy.co, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How to buy AI training data URL: https://sellmydatatoai.net/guides/how-to-buy-ai-training-data Reviewed: 2026-11-08 To buy AI training data safely, define the exact training use first, then source from sellers who can prove they hold the rights, test a randomly drawn sample in your own pipeline and sign a license covering training, derived assets and deletion. Rights evidence matters more than volume, because data cannot practically be removed from a model once it has been trained on. ## Key takeaways - You are almost always buying a license with defined rights, not the files outright. - The Data Provenance Initiative found license omission rates above 70% and error rates above 50% on popular dataset hosting sites, so a license tag proves little on its own. - Ask for a seeded random sample, an evidence pack with file hashes and a paid pilot before committing to a full purchase. - Your license should cover derived assets such as embeddings, evaluation sets and fine-tuned weights, plus what happens if rights fail after training. - Providers of general-purpose AI models selling into the EU face fines of up to 3% of worldwide turnover from August 2, 2026 if they cannot document their training content. Most AI teams start buying data when public sources run out or stop being safe to rely on. The hard part is rarely finding data. It is finding data whose rights will stand up to scrutiny, and checking its quality before you pay. This guide covers the full buying process, from writing a requirement to recording provenance after delivery. It draws on Nixon Peabody's October 2026 diligence alert, the Commission's guidance under the EU AI Act and published pricing from buyers and marketplaces. ## What does buying AI training data involve? Buying AI training data almost always means buying a license, not the files themselves. The seller keeps ownership and grants you defined rights: which models may be trained, for how long, in which territories and what happens to derived assets such as annotations, embeddings and fine-tuned weights. The distinction matters because training is hard to reverse. Daniel Goldberg of Frankfurt Kurnit Klein and Selz notes that once data has been used for training it cannot practically be pulled back out. Every check you skip before training becomes a risk you carry for the life of the model. A typical purchase runs through six stages: a written requirement, a shortlist of sellers, a sample under NDA, rights and quality review, a pilot, then a signed license with staged delivery. The sections below cover each one. ## What should you define before you approach sellers? Your requirement should define the training use, the shape of the data and the rights you need before you speak to any seller. Sellers package and price differently for pre-training, fine-tuning, evaluation and reinforcement learning, and a vague brief produces samples you cannot compare. | Training use | What to specify | Rights you usually need | | --- | --- | --- | | Pre-training | Modality, volume in tokens, hours or images, languages, date range, deduplication standard | Broad training rights across model versions | | Supervised fine-tuning | Task types, input and output format, expertise of authors, labeling guidelines | Training rights for named models and rights to outputs | | Evaluation | Number of tasks, expected answers, scoring rules, contamination controls | Evaluation use, often exclusive, with no onward publication | | Reinforcement learning | Environment, task set, verifier or reward rules, replay requirements | Rights to run, modify and generate new trajectories | | Domain adaptation | Industry, record types, source systems, de-identification standard | Training rights plus confidentiality and output safeguards | Add the commercial terms you can accept: exclusive or non-exclusive, term, territory, budget range and delivery date. If you can name the models the data will train and whether outputs will be commercialized, write that down as well. Sellers and their lawyers will ask. Our [training data catalog](https://sellmydatatoai.net/buy-ai-training-data) shows what each data type usually includes, which helps you write a requirement sellers can quote against. ## Where can you buy AI training data? AI training data is sold through five main routes: direct deals with data owners, training-data companies, data marketplaces, content licensing platforms and brokers. Each route trades price and speed against how much rights evidence you receive. | Route | How it works | Rights evidence you usually get | Best for | | --- | --- | --- | --- | | Direct from the data owner | You negotiate a license with the company that created the records | Whatever you request in diligence | Unique company data and exclusive deals | | Training-data company | The vendor collects, cleans or licenses data and delivers datasets | Varies by vendor, often summarized | Speed and managed delivery | | Data marketplace | Sellers list datasets with previews and standard terms | Listing terms and seller statements | Off-the-shelf tabular and media datasets | | Content licensing platform | Rights holders license catalogs through a platform or agent | Catalog-level or per-asset rights | Media, publishing and archives | | Broker | An intermediary finds and vets sellers, then you contract directly | A vetting file for each seller | Proprietary data with documented rights | Our [directory of data buyers and platforms](https://sellmydatatoai.net/data-buyers) lists the training-data companies, marketplaces and licensing platforms active in this market, with their published terms. Public and scraped data is the cheapest route and the least certain. The Data Provenance Initiative audited more than 1,800 text datasets and found license omission rates above 70% and error rates above 50% on popular hosting sites, according to its 2024 paper in Nature Machine Intelligence. ## How do you check that a seller has the right to license the data? A seller has the right to license data only if it can show ownership, permission for AI use and no conflicting obligations. Holding the files proves none of these. > "Possession does not simply mean permission." (Troy Lieberman, Nixon Peabody, October 2026) Nixon Peabody's alert frames the rights test as three questions: can the seller transfer the data to you, can you use it for the intended AI purpose and can you keep and commercialize what training produces. Restrictions can sit in privacy notices, employee policies, client contracts, IP assignments, inbound licenses, open-source terms and NDAs. - Ask for the privacy notice in force when the data was collected, not the current one. The same alert notes that regulators treat quietly broadening privacy terms to allow AI training as potentially unfair or deceptive. - Ask for IP assignments from the staff and contractors who created the content, especially for code and design files. - Ask which client contracts and NDAs cover the records, and whether any of them prohibit disclosure or reuse. - Ask how third-party material such as licensed images, open-source code or music in video was identified and then cleared or removed. - Ask who will sign and whether that person can bind the company. For a closed company it may be a liquidator or trustee. Our [due diligence checklist](https://sellmydatatoai.net/guides/ai-training-data-due-diligence-checklist) lists every document to request, and [data provenance explained](https://sellmydatatoai.net/guides/data-provenance-explained) covers how to trace a dataset back to its origin. ## How should you test a sample before you buy? A sample is useful only if it was drawn at random across the whole dataset and labeled with how it was drawn. A hand-picked sample shows you the seller's best rows, not what you will receive. Sellers face a real constraint here. In a Hugging Face forum thread that ran from July to September 2026, data sellers described the deadlock: they cannot prove quality without exposing the data, and buyers will not pay without seeing enough of it. Contributors proposed fixes you can ask for directly. - A seeded, stratified random sample that is labeled as representative - A locked preview where you run fixed queries without downloading the full set - An evidence pack with a file manifest and hashes tied to an exact version - A paid pilot on an agreed experiment, with only aggregate results returned if the seller prefers - Acceptance tests at delivery, with replacement or refund for objective specification failures Run the sample through your own pipeline rather than reviewing it by eye. Check the schema against the manifest, then test for duplicates, label accuracy, residual personal data and fitness for your task. [How to evaluate a dataset sample](https://sellmydatatoai.net/guides/how-to-evaluate-dataset-samples) sets out the full test plan. ## How much does AI training data cost? AI training data costs anything from a few thousand dollars for a research license to tens of millions of dollars a year for a large platform deal. Price depends on exclusivity, uniqueness, rights clarity and how much preparation the seller has already done. | Deal type | Figure | Source | | --- | --- | --- | | Research license | $5,000 to $15,000 | Phyllo, reported market ranges, September 2026 | | Commercial training license | $25,000 to $500,000 and above | Phyllo, reported market ranges, September 2026 | | Company data, buyers' published offers | Median of $280,000 per company | AIData.Broker, via ddiy.co, October 2026 | | Company data, public sales | Median of $31,600 per company | AIData.Broker, via ddiy.co, October 2026 | | Video footage | About $3 per minute | Bloomberg on Adobe, April 2024 | | Platform content | About $60 million a year | Reuters on Reddit and Google, February 2024 | A research license is cheap for a reason. Phyllo notes that research licenses usually do not permit training a production model, and that evaluating on research terms and then shipping without re-licensing is a common and expensive mistake. Exclusivity is the largest single multiplier. AIData.Broker puts an exclusive deal at about 2.5 to 5 times a non-exclusive one. [AI training data pricing models](https://sellmydatatoai.net/guides/ai-training-data-pricing-models) explains flat fees, unit pricing, subscriptions and revenue shares. ## What should the license agreement cover? The license agreement should cover what you may do with the data, what you own of everything training creates from it and who carries the risk if the data is not as described. Nixon Peabody recommends layer-by-layer terms, because deleting source files does not remove their influence from a trained model. | Clause | What to agree | | --- | --- | | Permitted use | Training, evaluation, fine-tuning or product development, and which models and affiliates | | Derived assets | Ownership of annotations, embeddings, evaluation sets, fine-tuned models, weights and outputs | | Exclusivity, term and territory | Exclusive or non-exclusive, start and end dates, permitted regions | | Onward transfer | Whether data may pass to cloud vendors, subprocessors, model partners or research partners | | Representations and warranties | Authority, privacy notices, contract restrictions, IP, litigation holds, completeness of exclusions | | Indemnities and caps | Privacy claims, IP disputes, regulatory inquiries, cost of removal or retraining | | Acceptance and payment | Acceptance tests, inspection period, staged payments and holdbacks | | Rights failure | Deletion, retraining or model retirement if rights lapse or prohibited material surfaces | | Expiry | What happens to trained models and retained copies when the license ends | ## What legal obligations do you have as the buyer? As the buyer, you carry the regulatory obligations that attach to training and deploying your model, and they do not pass to the seller. Which ones apply depends on where your model is placed on the market and what the data contains. - EU AI Act: providers of general-purpose AI models must keep a copyright compliance policy and publish a summary of training content under Article 53(1)(c) and (d) of [Regulation (EU) 2024/1689](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng). From August 2, 2026 the Commission can fine providers up to 3% of worldwide annual turnover or €15 million, whichever is higher. - Personal data: Nixon Peabody notes that under the CCPA, data counts as de-identified only if it cannot reasonably be linked to a person, the holder publicly commits not to re-identify it and recipients are contractually bound. UK GDPR and EU GDPR apply where records relate to people in the UK or EU. - Health data: HIPAA allows de-identification by expert determination or by removing 18 types of identifier under the safe harbor method, according to the [US Department of Health and Human Services](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html). - Remedies: the same alert notes that the FTC has required companies to delete models and algorithms built on improperly obtained data. This is a summary of what the law and regulators say, not advice on your situation. Take independent legal advice before you sign. Our guide to [the EU AI Act and training data transparency](https://sellmydatatoai.net/guides/eu-ai-act-training-data-transparency) covers the Article 53 summary in detail. ## What are the steps to buy a dataset? 1. **Write the requirement**: Define the training use, modality, volume, languages, date range, exclusivity, budget and delivery date. 2. **Build a shortlist**: Compare sellers on rights evidence and fit for your task, not on volume alone. 3. **Sign a mutual NDA and review samples**: Ask for a seeded random sample labeled with its sampling method, plus the data manifest. 4. **Run rights diligence**: Collect privacy notices, IP assignments, contract reviews and the de-identification method before any pilot. 5. **Pilot**: Test the data on an agreed experiment in a controlled environment and expand access only as conditions are met. 6. **Negotiate the license**: Agree permitted use, derived assets, warranties, indemnities, acceptance tests and staged payments. 7. **Accept delivery**: Check the delivered files against the manifest and hashes within the agreed inspection period. 8. **Record provenance**: Keep the license, manifest, hashes and diligence file so you can complete regulatory summaries and answer challenges later. ## How does Sell My Data To AI help buyers? Sell My Data To AI helps buyers as an independent broker. We do not own an AI model and we do not buy data for our own use. Buyers pay nothing to submit a requirement or review a shortlist, and any fee is agreed in writing before an introduction. Every seller passes our eight-point vetting standard before any buyer sees their data, covering authority to sign, ownership, contract restrictions, privacy notices, third-party content, de-identification, sample quality and manifest accuracy. You see redacted samples under a mutual NDA and sign the license directly with the seller. [How we vet](https://sellmydatatoai.net/how-we-vet) explains each check. We are a broker, not a law firm, so take your own legal advice before signing. To start, [book a buyer consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer) or [submit a data requirement](https://sellmydatatoai.net/request-data). ## Frequently asked questions ### Can I buy AI training data outright rather than license it? Outright purchase of AI training data is rare. Most sellers grant a license because they want to keep the data or license it again elsewhere. An exclusive perpetual license comes closest to a purchase. Polyshares, for example, has published four company data deals between $110,000 and $450,000, all perpetual and exclusive, according to ddiy.co's October 2026 compilation of buyer terms. ### Is scraped web data safe to train a production model on? Scraped web data carries more rights uncertainty than any other source. Website terms, copyright and personal data law all apply, and in the EU, providers of general-purpose AI models must respect text and data mining opt-outs under Article 4(3) of Directive (EU) 2019/790. Licensed data with documented provenance is the safer basis for a model you plan to commercialize. ### How long does it take to buy a dataset? Buying a dataset usually takes weeks rather than days, and most of that time goes on rights diligence and legal review rather than delivery. Published seller-side timelines give a guide: Handshake AI states 7 to 14 business days for its process and Polyshares makes offers in 1 to 2 weeks, per ddiy.co. Allow longer for exclusive deals, personal data or a pilot. ### Do I need an exclusive license? You need an exclusive license only if your advantage depends on competitors not training on the same data, as with a private evaluation set or a unique domain archive. Exclusivity costs more: AIData.Broker puts exclusive deals at about 2.5 to 5 times a non-exclusive one. A time-limited or field-limited exclusive can be a cheaper middle ground. ### What if the data contains material the seller did not disclose? Your license should answer this before it happens. Agree warranties on the completeness of disclosed exclusions, a process for removing prohibited records, indemnities for resulting claims and whether deletion, retraining or model retirement applies once the data has been used. Nixon Peabody recommends settling these remedies at signing, because deleting source files does not remove their influence from a trained model. ### Does Sell My Data To AI hold the data I buy? No. Sell My Data To AI does not take custody of raw data. Buyers see redacted samples under a mutual NDA, and full data moves only after you sign a license directly with the seller. That keeps the contract, the delivery and the rights evidence between you and the data owner, with us helping both sides prepare. ## Sources - [Shopping for AI training data? Five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody LLP, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 - [Is selling datasets way harder than building them, or is it just me?](https://discuss.huggingface.co/t/is-selling-datasets-way-harder-than-building-them-or-is-it-just-me/177440), Hugging Face forums, accessed November 2026 - [A large-scale audit of dataset licensing and attribution in AI](https://www.media.mit.edu/publications/a-large-scale-audit-of-dataset-licensing-and-attribution-in-ai/), MIT Media Lab (Nature Machine Intelligence, 2024), accessed November 2026 - [Best public data providers for AI training and grounding](https://www.getphyllo.com/post/best-public-data-providers-ai-training-grounding), Phyllo, accessed November 2026 - [How much is company data worth to AI?](https://ddiy.co/how-much-is-company-data-worth-to-ai/), ddiy.co, accessed November 2026 - [Sell company data to AI](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 - [Adobe is buying videos for $3 per minute to build AI model](https://www.bloomberg.com/news/articles/2024-04-10/adobe-is-buying-video-clips-for-3-per-minute-to-build-ai-model), Bloomberg, accessed November 2026 - [Exclusive: Reddit in AI content licensing deal with Google, sources say](https://tech.yahoo.com/ai/articles/exclusive-reddit-ai-content-licensing-041012503.html), Reuters (syndicated by Yahoo), accessed November 2026 - [Regulation (EU) 2024/1689 (Artificial Intelligence Act)](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng), EUR-Lex, Publications Office of the European Union, accessed November 2026 - [Guidance regarding methods for de-identification of protected health information](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html), US Department of Health and Human Services, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # AI training data due diligence checklist URL: https://sellmydatatoai.net/guides/ai-training-data-due-diligence-checklist Reviewed: 2026-11-08 AI training data due diligence means confirming, before you pay or train, that the seller can sign, owns the data, faces no contract or privacy restrictions, has cleared third-party content, has de-identified personal data by a documented method and will deliver what the manifest promises. Ask for documents rather than assurances, and stage access so a sample always comes first. ## Key takeaways - Due diligence should test three rights separately: transfer to you, use for your AI purpose and commercialization of what training produces. - The privacy notice that matters is the one in force when the data was collected, not the seller's current notice. - De-identification needs a documented method, record-level exclusions and a re-identification test, not just deleted names. - Stage access: a limited sample in a controlled environment first, full delivery only after rights and quality checks pass. This checklist is for AI teams licensing data from companies, marketplaces or vendors. It follows the order in which problems usually surface, starting with who can sign and ending with the records you need to keep after delivery. It draws on Nixon Peabody's October 2026 alert on buying AI training data, guidance from Frankfurt Kurnit Klein and Selz on company data sales and the Commission's template for EU AI Act training data summaries. It is not legal advice, and you should take independent advice on any deal. ## What is AI training data due diligence? AI training data due diligence is the review a buyer runs to confirm that a dataset can lawfully be used for training and matches its description. It covers the seller's authority and title, restrictions inherited from contracts and privacy law, the handling of personal data and third-party material, and the quality of what will actually be delivered. The review matters more for training data than for most purchases because the use is hard to undo. Nixon Peabody notes that deleting source files may not remove their influence from a trained model, and that the FTC has required companies to delete models and algorithms built on improperly obtained data. ## Which documents should you ask a seller for? The documents to ask a seller for are the ones that turn each rights claim into evidence. Statements in an email or a license tag on a listing are a starting point, not proof. | Check | Ask for | Red flag | | --- | --- | --- | | Authority to sign | Board approval or signatory authority, liquidator or trustee appointment for a closed company | Signatory is a former employee or a shareholder acting alone | | Ownership and chain of title | Statement of how the records were created or acquired, IP assignments from staff and contractors | Data was bought, scraped or downloaded from elsewhere | | Contract restrictions | List of client contracts, NDAs, vendor terms and data processing agreements covering the records | No review done, or confidentiality clauses with no carve-out | | Privacy notice and lawful basis | The privacy notices in force when the data was collected, with dates | Only the current notice is available, or it was changed recently to mention AI | | Third-party content | Inventory of licensed images, open-source code, music and other owned material, with clearance or removal | Seller cannot say what third-party content is present | | De-identification method | Written method, list of fields removed, re-identification test results | Names deleted and nothing else | | Representative sample | Sample drawn at random, labeled with method and seed | Sample selected by the seller | | Manifest accuracy | Data manifest with record counts, date ranges, systems, languages and file hashes | Counts given as estimates with no version | ## How do you confirm the seller can sign? You confirm the seller can sign by checking that the named signatory can bind the legal entity that owns the records. That sounds basic, but data often sits in group companies, former subsidiaries or businesses that have closed. - Registered entity name and number for the company that holds the records - Evidence that the signatory has authority, such as a board resolution for a larger deal - For a closed or insolvent company, the appointment of the liquidator, administrator or trustee - For a US bankruptcy sale, whether the sale is consistent with the privacy policy or has court approval after a consumer privacy ombudsman review under Section 332 Nixon Peabody points out that Section 363(b)(1) of the US Bankruptcy Code restricts sales of personally identifiable information that conflict with the debtor's privacy policy, and that court approval does not cure every restriction inside the records. Our guide to [selling data from a closed company](https://sellmydatatoai.net/guides/selling-data-from-a-closed-company) covers the seller's side. ## How do you check ownership and chain of title? Ownership and chain of title are checked by tracing each part of the dataset to the person or system that created it and the agreement that passed the rights on. A company owns its own operational records in most cases, but staff, contractors and clients may hold rights in parts of them. - Employment and contractor agreements with IP assignment clauses, especially for code, designs and written content - Client contracts stating who owns deliverables and working files - Inbound licenses for any data the company bought or received from third parties - Open-source license terms for any code in a repository The Data Provenance Initiative found license omission rates above 70% on popular dataset hosting sites in its 2024 audit, which is why a license label is never enough on its own. [Data provenance explained](https://sellmydatatoai.net/guides/data-provenance-explained) shows how to build the trail. ## Which contracts can restrict a data sale? The contracts that can restrict a data sale are client agreements, NDAs, vendor terms, data processing agreements and employment arrangements. Frankfurt Kurnit Klein and Selz notes that email and chat archives often contain other companies' confidential information covered by NDAs. - Client confidentiality clauses that cover working documents and correspondence - Data processing agreements where the seller acted as a processor, which usually bar use for its own purposes - SaaS terms that limit how exported data from a platform may be used - Union collective bargaining agreements, which may restrict a sale of workplace communications, according to the same firm In the US, the firm also notes that employee communications about pay and working conditions may be protected concerted activity under Section 7 of the National Labor Relations Act. Ask whether staff representatives have been consulted where communications data is involved. ## How do you check privacy notices and lawful basis? Privacy notices are checked by comparing the notice in force at the time of collection with the proposed AI use. Frankfurt Kurnit Klein and Selz notes that companies must honor the statements made when the data was collected, and that the FTC treats quietly changing them as potentially deceptive. - Dated copies of each privacy notice that covered the records - The lawful basis relied on for disclosure where UK GDPR or EU GDPR applies - Whether the data counts as a sale under the CCPA, and whether it meets the CCPA de-identification standard - Whether any sector rules apply, such as HIPAA for health records or GLBA for financial records - Whether employee and applicant data is included, which Nixon Peabody notes is generally subject to the same CCPA sale restrictions as consumer data Our guide to [UK GDPR, EU GDPR and CCPA when selling data for AI training](https://sellmydatatoai.net/guides/gdpr-and-selling-data-for-ai-training) explains each regime from the seller's side. ## What should you check about de-identification? De-identification should be checked as a documented method with test results, not as a statement that names were removed. Nixon Peabody warns that job titles, dates and linked records can still identify people after names are gone. - A written method, ideally applied by an independent provider and informed by NIST guidance - Record-level exclusions for categories that should never transfer - Re-identification testing against external data - Certification before each transfer, not only at the start - Contractual bans on re-identification backed by technical controls Frankfurt Kurnit Klein and Selz goes further and recommends excluding high-risk data entirely rather than relying on de-identification. For US health data, the [Department of Health and Human Services](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html) recognizes two methods: expert determination, or safe harbor removal of 18 types of identifier. > **Rule of thumb** If a seller will not tell you where the data came from or how personal data was handled, treat the dataset as if the worst case were true and price or walk away accordingly. ## How do you check quality and the manifest? Quality and the manifest are checked together, because the manifest is the seller's promise about what you will receive. Every count, date range and language it lists should be testable in the sample and again at delivery. - Record counts, file counts and total size by source system - Date ranges and languages, with the share of each - File hashes tied to a specific version of the dataset - Duplicate detection method with disclosed thresholds - Label guidelines and reviewer agreement for any annotated data [How to evaluate a dataset sample](https://sellmydatatoai.net/guides/how-to-evaluate-dataset-samples) sets out the tests, and [how to build a data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest) shows what a good manifest contains. ## What security and delivery terms should you agree? Security and delivery terms should limit access to what each stage needs. Nixon Peabody recommends starting with a limited sample in a controlled environment and expanding access only as agreed conditions are met, and for sensitive archives keeping source data in a secure environment rather than handing over unrestricted copies. Frankfurt Kurnit Klein and Selz advises sellers to avoid giving buyer tools direct access to company systems and to use a secure transfer method. As a buyer, a fixed export delivered that way is also easier to hash, version and accept. ## Which contract protections should the license include? The license should include protections that match who controlled collection, who can test the data and who decides how it is used. Nixon Peabody lists the main tools. | Protection | What it covers | | --- | --- | | Representations and warranties | Authority, privacy notices, contract restrictions, IP rights, regulatory compliance, litigation holds, security, completeness of exclusions | | Indemnities and caps | Privacy claims, confidentiality breaches, IP disputes, regulatory inquiries, cost of removing data or retraining | | Holdbacks and insurance | Funds or cover available if a warranty fails | | Termination rights | Exit for rights failures or prohibited material | | Challenge process | How claims from former employees, customers, vendors or regulators are handled | | Derived asset terms | What happens to embeddings, evaluation sets, fine-tuned models and weights if rights lapse | The [AI data licensing agreement checklist](https://sellmydatatoai.net/guides/ai-data-licensing-agreement-checklist) covers the same clauses from the seller's side. ## What records do you need for the EU AI Act? For the EU AI Act you need enough information about each dataset to complete the Commission's public summary of training content if you provide a general-purpose AI model in the EU. Collect it during diligence, because it is much harder to reconstruct later. - Modality and approximate size of the data - Languages and any regional or demographic characteristics - The latest date the data was collected - Whether you licensed it directly from the rights holder or obtained it through an intermediary - For any scraped portion, how text and data mining opt-outs were respected - Measures taken to remove illegal content Our guide to [the EU AI Act and training data transparency](https://sellmydatatoai.net/guides/eu-ai-act-training-data-transparency) explains each section of the template. ## How does Sell My Data To AI handle seller diligence? Sell My Data To AI runs seller diligence before you see anything. Every seller passes our eight-point vetting standard before any buyer sees their data, and the eight checks match the core table above. Buyers pay nothing to submit a requirement or review a shortlist. We are a broker, not a law firm, so your own lawyers should still review the license. [Book a buyer consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer) to discuss a requirement. ## Frequently asked questions ### How long does AI training data due diligence take? AI training data due diligence takes as long as it takes the seller to produce documents. A well-prepared seller with a manifest, dated privacy notices and a de-identification report can answer most questions within days. A seller starting from an unreviewed email or chat archive may need weeks to map contracts and remove sensitive material before a sample is ready. ### Is a license tag on a public dataset enough? A license tag on a public dataset is not enough. The Data Provenance Initiative's 2024 audit of more than 1,800 text datasets found license omission rates above 70% and error rates above 50% on popular hosting sites. Trace the dataset to its original source and check the original terms before using it to train a commercial model. ### Who should run the de-identification, the buyer or the seller? The seller should run de-identification before data leaves its control, because the seller holds the original records and the legal duties attached to them. Nixon Peabody suggests an independent provider using documented methods, with re-identification testing and certification before each transfer. As the buyer, ask for the method and test results and agree a contractual ban on re-identification. ### What is the biggest red flag in a data sale? The biggest red flag in a data sale is a seller who cannot explain where the data came from. Data bought, scraped or downloaded from elsewhere rarely carries the rights to license it for AI training. A refusal to share a random sample or a dated privacy notice comes a close second. ### Do I need diligence for synthetic data too? Synthetic data still needs diligence, focused on the model that generated it and the data that model was trained on. The Commission's EU AI Act template asks providers to name the general-purpose AI models used to create synthetic training data, so record which model produced it, under what terms and whether those terms allow training a competing model. ## Sources - [Shopping for AI training data? Five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody LLP, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 - [A large-scale audit of dataset licensing and attribution in AI](https://www.media.mit.edu/publications/a-large-scale-audit-of-dataset-licensing-and-attribution-in-ai/), MIT Media Lab (Nature Machine Intelligence, 2024), accessed November 2026 - [Guidance regarding methods for de-identification of protected health information](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html), US Department of Health and Human Services, accessed November 2026 - [Explanatory notice and template for the public summary of training content for general-purpose AI models](https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models), European Commission, Shaping Europe's digital future, accessed November 2026 - [Regulation (EU) 2024/1689 (Artificial Intelligence Act)](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng), EUR-Lex, Publications Office of the European Union, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Data provenance explained for AI teams URL: https://sellmydatatoai.net/guides/data-provenance-explained Reviewed: 2026-11-08 Data provenance is the documented history of a dataset: where each part came from, who created and collected it, under what rights and how it was changed before it reached you. For AI teams it is the evidence that a dataset can lawfully be used for training, and the EU AI Act now requires model providers to document the provenance of their training data. ## Key takeaways - Provenance covers source, rights, collection method and every transformation, not just a license label. - The Data Provenance Initiative found license omission rates above 70% and error rates above 50% on popular dataset hosting sites. - Annex XI of the EU AI Act requires technical documentation on the type and provenance of training data and how it was obtained. - The Data and Trust Alliance's eight provenance standards give a practical checklist of fields to record for each dataset. Data provenance used to be a research concern. It is now a commercial and legal one, because AI teams have to show where their training data came from to regulators, rights holders and their own customers. ## What is data provenance? Data provenance is the record of a dataset's origin and history. It answers four questions: where the data came from, who created or collected it, under what rights it was obtained and what was done to it before it reached you. A provenance record is evidence, not a description. A sentence saying the data is licensed for commercial use is a claim. A copy of the license, the date it was signed and the list of files it covers is provenance. See the [glossary entry for data provenance](https://sellmydatatoai.net/glossary/data-provenance) for a short definition. ## Why does provenance matter for AI training data? Provenance matters for AI training data because training is hard to undo and the rules on what can be trained on are tightening. A dataset without provenance may be cheap to acquire and expensive to defend. The scale of the problem is documented. The Data Provenance Initiative, a team of legal and machine learning researchers, audited more than 1,800 text datasets and reported in Nature Machine Intelligence in 2024 that popular hosting sites had license omission rates above 70% and error rates above 50%. The authors describe this as a crisis of misattribution. Regulation has followed. Annex XI of the [EU AI Act](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng) requires providers of general-purpose AI models to document information on training data, including its type and provenance, curation methods and how it was obtained and selected. Article 53(1)(d) separately requires a public summary of training content. > "information on the data used for training, testing and validation, where applicable, including the type and provenance of data and curation methodologies" (Annex XI, point 2(c), Regulation (EU) 2024/1689) ## What does a provenance record contain? A provenance record contains fields that describe the source, the rights and the processing history of a dataset. The Data and Trust Alliance proposed eight standards in November 2023, built by experts from member companies including IBM, Mastercard, Pfizer, UPS and Walmart, and they make a practical checklist. | Standard | What it records | Example for licensed company data | | --- | --- | --- | | Source | Where the data originated | Support desk exports from the seller's ticketing system | | Legal rights | The rights under which it may be used | Signed license permitting training of named models | | Privacy and protection | Personal or sensitive data and how it was handled | De-identification method and re-identification test result | | Generation date | When the data was created or collected | Tickets from January 2019 to June 2026 | | Data type | The kind of data and its format | Text conversations in JSON lines | | Generation method | How it was produced | Written by human agents in live customer conversations | | Intended use and restrictions | What it may and may not be used for | Training and evaluation only, no resale | | Lineage | Transformations since creation | Deduplicated, identifiers removed, split by language | The standards also call for a unique provenance metadata ID for each dataset. At launch, the Alliance said only generation date was consistently surfaced in existing metadata. ## How is provenance different from lineage and chain of title? Provenance is the overall record, while lineage and chain of title are two parts of it. Lineage tracks technical changes to the data. Chain of title tracks the legal transfer of rights. | Term | Question it answers | Typical evidence | | --- | --- | --- | | Data provenance | Where did this come from and can we use it | Full record across source, rights, privacy and processing | | Data lineage | What has been done to this data | Processing logs, scripts, version history, hashes | | Chain of title | Who has held the rights and how they passed | IP assignments, licenses, purchase agreements | | Data manifest | What exactly is in this delivery | Record counts, date ranges, systems, languages, file hashes | The [chain of title](https://sellmydatatoai.net/glossary/chain-of-title) and [data manifest](https://sellmydatatoai.net/glossary/data-manifest) glossary entries go into each term in more detail. ## Which standards document dataset provenance? Dataset provenance is documented through a few widely used formats. None is mandatory, but buyers increasingly expect a seller to supply one of them. - Datasheets for datasets: a question set proposed by Timnit Gebru and co-authors on [arXiv](https://arxiv.org/abs/1803.09010) in March 2018, covering why a dataset was created, what it contains, how it was collected and what it is suited for. - Croissant: a metadata vocabulary from MLCommons built as an extension to schema.org, describing a dataset's attributes, provenance and usage restrictions. Its responsible AI extension adds a machine-readable summary of the attributes found in data cards. - Data and Trust Alliance standards: the eight cross-industry fields in the table above, aimed at enterprise data rather than research datasets. - Dataset cards: the documentation files that hosting platforms attach to each dataset, often based on the datasheet questions. MLCommons says any dataset with Croissant metadata is discoverable through Google Dataset Search, which covers more than 700,000 datasets across Hugging Face, Kaggle, OpenML and other sites. See our [dataset card](https://sellmydatatoai.net/glossary/dataset-card) entry for how cards relate to provenance. ## How do you verify a seller's provenance claims? You verify a seller's provenance claims by asking for the document behind each one and testing it against the data. A claim you cannot trace to a document should be treated as unverified. | Claim | Evidence to request | How to test it | | --- | --- | --- | | We created this data | Description of the source systems and the people who created the records | Check sample metadata for system names, authors and dates | | We own the rights | IP assignments, client contract terms, inbound licenses | Compare contract dates with record dates | | It is licensed for AI training | The license itself, with permitted uses | Read the permitted-use clause, not the summary | | Personal data has been removed | De-identification method and test results | Scan the sample for residual identifiers | | It covers 2015 to 2026 | Manifest with date ranges by source | Plot record dates in the sample | | It has not changed since you saw it | File hashes tied to a version | Recompute hashes at delivery | Phyllo's September 2026 guide adds two checks that are easy to miss: rights documentation that traces to individual assets rather than a blanket assurance for the whole corpus, and per-layer clearance for complex content, such as the music, images and trademarks inside video. ## What provenance should you keep after you buy? After you buy, keep the provenance file for as long as any model trained on the data is in use. You may need it to answer a rights holder, a regulator or a customer's procurement team years after the purchase. - The signed license and any amendments - The data manifest and file hashes for each delivered version - The seller's diligence file, including privacy notices and de-identification reports - Your own processing log from receipt to training - A list of the models and model versions trained on the data If you provide a general-purpose AI model in the EU, this file is what lets you complete the Commission's training data summary template. Licensed datasets fall into a separate section of that template, with lighter public disclosure than scraped or public data. Our guide to [the EU AI Act and training data transparency](https://sellmydatatoai.net/guides/eu-ai-act-training-data-transparency) explains how. ## How does provenance work for licensed company data? Provenance for licensed company data is usually simpler to establish than for public or scraped data, because there is one owner and one set of source systems. The work is in documenting it properly before a sale. Every seller we work with passes an eight-point vetting standard that includes ownership and chain of title, third-party content, de-identification method and manifest accuracy, before any buyer sees their data. You sign the license directly with the data owner, which keeps the rights trail short. Read the [due diligence checklist](https://sellmydatatoai.net/guides/ai-training-data-due-diligence-checklist) for the full document list, or [book a buyer consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer). ## Frequently asked questions ### Is data provenance a legal requirement? Data provenance is a legal requirement in part. The EU AI Act requires providers of general-purpose AI models to document the type and provenance of training data in technical documentation under Annex XI and to publish a summary of training content under Article 53(1)(d). Outside that, provenance is mainly how you show you had the rights you relied on. ### What is the difference between provenance and a license? A license is one document inside a provenance record. Provenance also covers where the data originated, how it was collected, what personal data it held and every transformation since. A valid license over data the licensor did not own gives you very little, which is why buyers trace the full history rather than relying on the license alone. ### Which provenance format should a seller use? A seller can use any format that answers the buyer's questions with evidence. Datasheets for datasets suits research-style documentation, Croissant suits machine-readable metadata on hosting platforms and the Data and Trust Alliance's eight standards suit enterprise data. Most buyers care more about the documents behind the fields than about the format chosen. ### Can provenance be fixed after a dataset has been used for training? Provenance can sometimes be reconstructed after training, but gaps are hard to close. You may be able to recover licenses or contracts, but you cannot easily show what a dataset contained if you did not hash and version it at the time. The Commission's template lets providers of older models state and justify information gaps, which shows how common they are. ## Sources - [A large-scale audit of dataset licensing and attribution in AI](https://www.media.mit.edu/publications/a-large-scale-audit-of-dataset-licensing-and-attribution-in-ai/), MIT Media Lab (Nature Machine Intelligence, 2024), accessed November 2026 - [Regulation (EU) 2024/1689 (Artificial Intelligence Act)](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng), EUR-Lex, Publications Office of the European Union, accessed November 2026 - [Leading corporations introduce data provenance standards](https://www.businesswire.com/news/home/20231130851266/en/Leading-Corporations-Introduce-Data-Provenance-Standards), Data and Trust Alliance via Business Wire, accessed November 2026 - [Datasheets for datasets](https://arxiv.org/abs/1803.09010), arXiv (Gebru et al.), accessed November 2026 - [Croissant metadata format](https://mlcommons.org/working-groups/data/croissant/), MLCommons, accessed November 2026 - [Best public data providers for AI training and grounding](https://www.getphyllo.com/post/best-public-data-providers-ai-training-grounding), Phyllo, accessed November 2026 - [Explanatory notice and template for the public summary of training content for general-purpose AI models](https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models), European Commission, Shaping Europe's digital future, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How to evaluate a dataset sample before you buy URL: https://sellmydatatoai.net/guides/how-to-evaluate-dataset-samples Reviewed: 2026-11-08 To evaluate a dataset sample, first confirm it was drawn at random across the whole dataset and labeled with its sampling method, then test it in your own pipeline: schema against the manifest, duplicates, label accuracy, residual personal data, third-party content and fitness for your task. Agree in advance which results would stop the deal or change the price. ## Key takeaways - A hand-picked sample tells you about the seller's best rows, not the dataset, so ask for a seeded random or stratified sample. - Test the sample in your own pipeline against the manifest, not by eye. - When a seller cannot share enough data, ask for a locked preview or a paid pilot instead of skipping the test. - Turn what you learn from the sample into written acceptance tests for the full delivery. The sample is the only part of a dataset most buyers see before signing. It has to do two jobs: show that the data is fit for your task, and give you a baseline to test the full delivery against. ## Why are dataset samples hard to trust? Dataset samples are hard to trust because the seller controls what you see, and the seller has reasons to show as little as possible. Full access would give the data away before payment. Sellers describe the same problem from the other side. In a Hugging Face forum thread that ran from July to September 2026, a dataset seller wrote that the seller cannot prove quality without exposing too much. A later contributor called this the main blocker in the market. > "The demand side is real. The plumbing is missing." (Talahakd, Hugging Face forums, September 2026) The fix is not to accept less evidence. It is to agree a sampling method and a test plan that give you enough signal without the seller handing over the dataset. ## What kind of sample should you ask for? The sample to ask for is a random sample drawn across the whole dataset, stratified by the segments that matter to you and drawn with a recorded seed. Each alternative tells you something different. | Sample type | How it is drawn | What it tells you | Main weakness | | --- | --- | --- | --- | | Hand-picked | The seller chooses the records | The best case | Overstates quality | | Simple random | Records drawn at random across the archive | What a typical record looks like | Can miss small segments | | Stratified random | Random draws within agreed segments such as source system, year or language | Quality in every segment | Needs agreed segments | | Seeded | Random draw with the seed recorded | A sample anyone can reproduce and audit | None beyond the base method | | Locked preview | You run fixed queries in the seller's environment | Fit with your pipeline without a download | Limited depth | | Paid pilot | An agreed training or evaluation run | Real performance signal | Cost and time | Ask the seller to label the sample with the method, the seed and the version of the dataset it came from. Contributors to the Hugging Face thread suggested a free representative tier that is clearly labeled as such, and a fixed-seed stratified human audit with reviewer disagreement reported. ## What should you test in a dataset sample? In a dataset sample you should test structure, content and fitness. Structure tests check the sample matches the manifest. Content tests check for problems that create legal or quality risk. Fitness tests check the data helps your model. | Test | Method | Pass condition | | --- | --- | --- | | Manifest match | Compare fields, formats, date ranges and languages with the manifest | No field or range missing | | Schema and format | Load into your pipeline with no manual fixes | Loads cleanly | | Duplicates | Exact and near-duplicate detection with stated thresholds | Rate within the agreed limit | | Label accuracy | Blind re-label a subset and compare | Agreement within the agreed limit | | Residual personal data | Automated identifier scan plus manual review of free text | No direct identifiers, indirect identifiers within the agreed method | | Third-party content | Search for licensed images, open-source license headers, branded material | Matches the seller's inventory | | Distribution | Compare the spread of dates, sources and languages with the manifest | Within the stated shares | | Contamination | Check overlap with public benchmarks you report on | No overlap for evaluation data | | Task fitness | Small fine-tune or evaluation run against a baseline | Measurable gain on your metric | Set the limits before you run the tests. If you agree them afterward, every result becomes a negotiation. Our guide to [building a data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest) shows what the manifest should state so these tests have something to check against. ## How do you check a sample for personal data? A sample is checked for personal data by scanning for direct identifiers and then reading free text for indirect ones. Nixon Peabody notes that context such as job titles, dates or linked records can identify people even after names are removed. - Run an automated scan for names, email addresses, phone numbers, account numbers and addresses - Read a slice of free-text fields by hand, because signatures, quoted emails and notes often carry identifiers - Check attachments, image metadata and audio transcripts separately - For US health records, check against the 18 safe harbor identifier types listed by the [Department of Health and Human Services](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html) If you find identifiers, ask how the seller's method missed them before you ask for a cleaner sample. Our guide to [de-identification for AI training data](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data) explains the methods. ## How do you test task fitness without the full dataset? Task fitness can be tested without the full dataset by running a small, controlled experiment on the sample and comparing it with a baseline. The aim is direction, not a final number. 1. **Fix a baseline**: Record your current model's score on the evaluation you care about. 2. **Train or evaluate on the sample**: Run a short fine-tune, or score your model on the sample if it is evaluation data. 3. **Compare like with like**: Use the same settings and evaluation set for both runs. 4. **Check for leakage**: Make sure the evaluation set shares no records with the sample. 5. **Decide what the result means**: A gain on a small sample suggests value, and no gain is a reason to ask for a larger pilot or walk away. If the seller will not release enough data for this, propose a paid pilot. The Hugging Face thread also describes a seller-side pilot, where the seller runs an agreed experiment on a fixed budget and returns only aggregate results. Nixon Peabody recommends a similar staged approach for sensitive archives. ## How do you turn a sample review into acceptance tests? A sample review becomes a set of acceptance tests when you write each finding down as a measurable condition for the full delivery. The sample is your baseline, and the delivery should match it. - File validation, record counts and hashes against the manifest - Duplicate rate no higher than in the sample - Label accuracy on a sampled subset no lower than in the sample - No residual identifiers beyond the agreed method - An inspection period, a correction deadline and replacement or refund for objective specification failures Contributors to the Hugging Face thread recommended separating remedies for specification failures from remedies for failing to reproduce a reference result or failing to improve your own model. Sellers will usually accept the first, and the last is often a matter for price rather than refund. ## What are the red flags in a dataset sample? The red flags in a dataset sample are signs that it is not representative or that the seller does not know what the dataset contains. - The seller will not say how the sample was drawn - Every record is clean and complete, with no edge cases - Date ranges or languages in the sample do not match the manifest - Personal data appears in a dataset described as de-identified - File names or metadata point to a different company or source - The sample comes from a different version than the one being offered ## How do samples work through Sell My Data To AI? Through Sell My Data To AI, every sample is drawn at random across the archive rather than hand-picked, so you can judge the whole package. That is one of the eight checks every seller passes before a buyer sees their data, alongside manifest accuracy and a re-identification test on a sample. Buyers see redacted samples under a mutual NDA, and full data moves only after you sign a license directly with the seller. [How we vet](https://sellmydatatoai.net/how-we-vet) describes each check. To discuss a requirement, [book a buyer consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer) or read [how to buy AI training data](https://sellmydatatoai.net/guides/how-to-buy-ai-training-data). ## Frequently asked questions ### How big should a dataset sample be? A dataset sample should be big enough to show every segment you care about and to run your fitness test, and no single size suits every dataset. Stratifying by source system, year and language matters more than raw size. If a useful test needs more data than the seller will release, move to a locked preview or a paid pilot. ### Should I pay for a sample? Paying for a basic sample is unusual, but paying for a pilot is common and often worth it. A pilot gives you a real training or evaluation signal and gives the seller a reason to release more data. Agree in writing what the pilot covers, who owns the results and whether the fee counts toward the final price. ### What is a locked preview? A locked preview is an arrangement where you run fixed queries or tests against a dataset inside the seller's environment, without downloading it. Contributors to a 2026 Hugging Face forum thread proposed it as a way to test pipeline fit while the seller keeps control of the data. It works best for structured data and automated checks. ### Can I keep a sample if I do not buy? Whether you can keep a sample depends on the NDA, and most require you to delete it if no license is signed. Ask for the deletion terms before you receive the sample, and do not use sample data in any training run outside the agreed evaluation. Keep your test results, which are your own work. ## Sources - [Is selling datasets way harder than building them, or is it just me?](https://discuss.huggingface.co/t/is-selling-datasets-way-harder-than-building-them-or-is-it-just-me/177440), Hugging Face forums, accessed November 2026 - [Shopping for AI training data? Five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody LLP, accessed November 2026 - [Guidance regarding methods for de-identification of protected health information](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html), US Department of Health and Human Services, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # AI training data pricing models URL: https://sellmydatatoai.net/guides/ai-training-data-pricing-models Reviewed: 2026-11-08 AI training data is priced in five main ways: a flat license fee, a price per unit such as per minute of video or per task, a recurring subscription, a revenue share or staged milestone payments. Exclusivity is the biggest multiplier, with AIData.Broker putting exclusive deals at about 2.5 to 5 times non-exclusive ones. Rights clarity and uniqueness drive the rest. ## Key takeaways - Flat fees suit one-off company archives, unit pricing suits media and tasks, and subscriptions suit data that is refreshed. - AIData.Broker reports a median of $280,000 in buyers' published offers per company, against a median of $31,600 in sales made public. - Exclusive deals cost about 2.5 to 5 times non-exclusive ones, according to AIData.Broker. - Research licenses are cheapest but usually do not permit training a production model, according to Phyllo. There is no list price for AI training data. Each deal is priced on the dataset, the rights granted and the buyer's alternatives. What is consistent is the set of structures used to express a price, and the factors that move it up or down. ## How is AI training data priced? AI training data is priced using five structures, and many deals combine two of them. The right structure depends on whether the data is a one-off archive, a stream of new records or a catalog sold to many buyers. | Model | How it works | Published example | Suits | | --- | --- | --- | --- | | Flat license fee | One price for a defined dataset and set of rights | Polyshares: four published deals of $110,000 to $450,000, all perpetual and exclusive (ddiy.co) | One-off company archives | | Per unit | A unit price multiplied by volume | Adobe: about $3 per minute of video (Bloomberg, April 2024) | Video, audio, images and tasks | | Subscription | An annual fee for ongoing access and updates | Reddit and Google: about $60 million a year (Reuters, February 2024) | Data that is refreshed | | Revenue share | The seller receives a share of each resale | Troveo: paid from revenue share, no fees (ddiy.co) | Catalogs sold many times | | Milestone payments | Payment in tranches as delivery and acceptance are met | Handshake AI: milestone payments (ddiy.co) | Large or staged deliveries | Hybrids are common. ddiy.co's October 2026 compilation lists Labelbox Prism as paying an upfront fee plus a revenue share, and many flat-fee deals are paid in stages tied to acceptance. ## What units are used in per-unit pricing? Per-unit pricing uses the unit the buyer actually consumes. AIData.Broker, as quoted by ddiy.co, lists the company, the code repository, the hour of audio or video and the task as the usual units, each multiplied by volume. | Data type | Typical unit | Reported figure | | --- | --- | --- | | Video footage | Per minute or per hour | About $3 per minute (Bloomberg on Adobe, April 2024) | | Video for OpenAI | Per minute | $1 to $4 per minute, according to a February 2026 Versos post that cites no source | | Company records | Per company | Median of $280,000 in buyers' published offers (AIData.Broker via ddiy.co) | | Source code | Per repository | About $25,000 for a shut-down startup's codebase (MIT Technology Review, October 2026) | | Agent tasks | Per task or per taskset | Rates set per listing on DataVendor, not published as a range | Buyers usually price a company archive as a whole rather than per document, according to ddiy.co. That is why ranges for company data are quoted per company. ## What do published price ranges look like? Published price ranges for AI training data run from a few thousand dollars for research use to over a million dollars for exclusive or custom data. Phyllo's September 2026 guide gives reported market bands, which it says should be treated as ranges rather than quotes. | License type | Reported range | Coverage | | --- | --- | --- | | Research | $5,000 to $15,000 | Academic and non-commercial use | | Commercial | $25,000 to $500,000 and above | Commercial training and deployment | | Professional annual | $25,000 to $50,000 a year | Ongoing access with periodic updates | | Enterprise annual | $50,000 to $500,000 a year | Real-time access and custom features | | Exclusive or custom | $1 million and above | Bespoke collection, exclusive rights, custom benchmarks | For company data, buyers publish their own ranges. Mercor's listing, as compiled by ddiy.co in October 2026, puts law firms at $50,000 to $1.2 million, construction companies at $115,000 to $1 million and call centers at $20,000 to $250,000 per company. These are offer ranges, not completed prices. Completed prices tend to sit lower. AIData.Broker reports a median of $280,000 across buyers' published offers but a median of $31,600 across the few sales made public. Fortune reported in September 2026 that payments are typically $10,000 to $100,000 per company, citing SimpleClosure, while Google bid $10 million for bankrupt Spirit Airlines' data archive in an August 2026 auction that was subject to court approval when reported. Our [AI data deals tracker](https://sellmydatatoai.net/ai-data-deals) lists reported transactions, and [how much is your data worth to AI companies](https://sellmydatatoai.net/guides/how-much-is-your-data-worth-to-ai) explains the same numbers from the seller's side. ## What makes one dataset worth more than another? One dataset is worth more than another when it is harder to replace and safer to use. Volume matters less than most sellers expect. - Exclusivity: AIData.Broker puts exclusive deals at about 2.5 to 5 times non-exclusive ones. - Rights clarity: a smaller dataset with clear provenance can be worth more than a larger one whose best parts must be excluded, according to Nixon Peabody. - Human authorship: DataVendor reports that labs pay more for code written mostly by people, and for complex products. - Language and difficulty: DataVendor names systems languages such as Rust, Go, C and C++ as attracting a premium. - Context: production code with pull requests, issue tickets, runbooks and incident notes explains why code changed, which DataVendor says buyers value. - Reproducibility: DataVendor notes that evaluation sets command more attention when independent reviewers can reproduce the scores. License scope moves price as well. A license limited to one model, one field of use or a fixed term should cost less than a perpetual license covering every model your company builds. ## How does exclusivity change the price? Exclusivity raises the price because the seller gives up every other sale of the same data. AIData.Broker's estimate of 2.5 to 5 times a non-exclusive deal reflects that lost income. You can narrow exclusivity to pay less for it. Common options are a time-limited exclusive, an exclusive within one field such as legal drafting, or an exclusive against named competitors only. Polyshares, by contrast, makes exclusive and perpetual the default in its published deals, according to ddiy.co. Our guide to [exclusive vs non-exclusive data licenses](https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses) covers the trade-offs. ## How should a buyer structure payment? A buyer should structure payment so that money follows evidence. Pay a first installment on signing, the main payment on acceptance of the delivered data and, for larger deals, hold back a portion until the warranty period ends. Published seller timelines show what the market expects. ddiy.co's compilation reports that Mercor pays 2 to 4 weeks after signing, micro1 takes up to 30 days to approve and pays 30 days after invoice, and Polyshares pays on net 30 to 60 day terms. Matching your schedule to these norms makes an offer easier to accept. > **Check the permitted use before the price** Phyllo warns that research licenses usually do not permit training a production model. A cheap license that does not cover your use is the most expensive option once you have to re-license or retrain. ## Which pricing model should you choose? The pricing model you choose should follow how the data will be used and how often it changes. | Situation | Model that usually fits | | --- | --- | | One company's historical archive | Flat fee, paid in stages against acceptance | | Media you will buy more of over time | Per unit with a volume tier | | Data that is refreshed monthly or quarterly | Annual subscription | | A seller who wants to keep upside | Lower upfront fee plus revenue share | | A custom collection or evaluation set | Milestone payments tied to delivery batches | ## What does Sell My Data To AI charge buyers? Sell My Data To AI charges buyers nothing to submit a requirement or review a shortlist. Any fee is agreed in writing before an introduction, and our [fees page](https://sellmydatatoai.net/fees) explains how this works. You negotiate the price and payment terms directly with the seller. To get a view of what a dataset should cost before you negotiate, [book a buyer consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer) or [submit a data requirement](https://sellmydatatoai.net/request-data). ## Frequently asked questions ### What is the cheapest way to buy AI training data? The cheapest way to buy AI training data is a non-exclusive license to an existing dataset, ideally one already prepared for sale. Research licenses cost less again, with Phyllo reporting $5,000 to $15,000, but they usually do not permit training a production model. Cheap data that you later have to re-license or retrain without costs more in the end. ### Why are published offers so much higher than completed sales? Published offers are ceilings set by buyers to attract sellers, while completed sales reflect what a specific dataset was worth after diligence. AIData.Broker reports a $280,000 median for buyers' published offers against $31,600 for the sales made public. Exclusions for personal data, client confidentiality and third-party content often reduce what is left to sell. ### Is revenue share a good deal for a buyer? Revenue share suits a buyer that resells data, such as a platform or aggregator, because payment tracks income. For an AI lab training its own models, a revenue share is harder to calculate and audit, so a flat fee or subscription is usually simpler. Labelbox Prism combines an upfront fee with a revenue share, according to ddiy.co. ### Should I pay per token for text data? Per-token pricing is rare for licensed company data, because buyers usually price an archive as a whole, according to ddiy.co. Tokens are useful for comparing offers and estimating training value, so ask sellers for token counts in the manifest. The Commission's EU AI Act template also asks providers to report text training data size in token ranges. ## Sources - [How much is company data worth to AI?](https://ddiy.co/how-much-is-company-data-worth-to-ai/), ddiy.co, accessed November 2026 - [Sell company data to AI](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 - [Best public data providers for AI training and grounding](https://www.getphyllo.com/post/best-public-data-providers-ai-training-grounding), Phyllo, accessed November 2026 - [Adobe is buying videos for $3 per minute to build AI model](https://www.bloomberg.com/news/articles/2024-04-10/adobe-is-buying-video-clips-for-3-per-minute-to-build-ai-model), Bloomberg, accessed November 2026 - [Exclusive: Reddit in AI content licensing deal with Google, sources say](https://tech.yahoo.com/ai/articles/exclusive-reddit-ai-content-licensing-041012503.html), Reuters (syndicated by Yahoo), accessed November 2026 - [Sell video content as AI training data](https://www.versos.ai/blog/sell-video-content-ai-training-data), Versos, accessed November 2026 - [AI labs are buying data from dead startups to train models](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models), MIT Technology Review, accessed November 2026 - [micro1 tries to beat Google's bid to buy bankrupt Spirit Airlines' data](https://fortune.com/2026/09/14/micro1-tries-to-beat-googles-bid-to-buy-bankrupt-spirit-airlines-data-for-12-5-million-highlighting-value-of-data-to-ai-companies/), Fortune, accessed November 2026 - [AI training data marketplaces](https://datavendor.ai/blog/ai-training-data-marketplaces), DataVendor, accessed November 2026 - [Shopping for AI training data? Five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody LLP, accessed November 2026 - [Explanatory notice and template for the public summary of training content for general-purpose AI models](https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models), European Commission, Shaping Europe's digital future, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # The EU AI Act and training data transparency URL: https://sellmydatatoai.net/guides/eu-ai-act-training-data-transparency Reviewed: 2026-11-08 Article 53(1)(d) of the EU AI Act requires every provider of a general-purpose AI model to publish a sufficiently detailed summary of its training content, using the template the Commission published on July 24, 2025. The duty has applied since August 2, 2025, Commission fines of up to 3% of worldwide turnover apply from August 2, 2026, and older models have until August 2, 2027. ## Key takeaways - Every general-purpose AI model provider must publish a training content summary, including providers of open-source models. - The summary lists large public datasets and the top 10% of scraped domains, but licensed datasets need only a yes or no answer and the modality. - Commission enforcement and fines began on August 2, 2026, and models placed on the market before August 2, 2025 must comply by August 2, 2027. - The Digital Omnibus on AI, Regulation (EU) 2026/1744, delayed high-risk rules but left the Article 53 training data duties unchanged. The EU AI Act is the first law to require AI model providers to say publicly what their models were trained on. For buyers of training data, it turns provenance records into a compliance input. For sellers, it determines what a buyer will disclose about their data. This guide is based on the text of Regulation (EU) 2024/1689 on EUR-Lex, the Commission's explanatory notice and template, its questions and answers on general-purpose AI and the Digital Omnibus on AI. It is a summary, not legal advice. ## What does Article 53 of the EU AI Act require? Article 53 of the EU AI Act sets four obligations for providers of general-purpose AI models. Two concern documentation for authorities and downstream users, and two concern training data. - Article 53(1)(a): keep technical documentation of the model, including its training and testing process, with at least the information in Annex XI - Article 53(1)(b): give information and documentation to providers who integrate the model into their own AI systems - Article 53(1)(c): put in place a policy to comply with EU copyright law, including identifying and complying with rights reservations under Article 4(3) of Directive (EU) 2019/790 - Article 53(1)(d): draw up and publish a summary of the content used for training, using the AI Office's template > "draw up and make publicly available a sufficiently detailed summary about the content used for training of the general-purpose AI model, according to a template provided by the AI Office" (Article 53(1)(d), Regulation (EU) 2024/1689) Annex XI adds that the technical documentation must describe the data used for training, testing and validation, including its type and provenance, curation methods, number of data points and how it was obtained and selected. That documentation goes to the AI Office on request, while the summary is public. ## Who has to publish a training data summary? Every provider of a general-purpose AI model placed on the EU market has to publish a training data summary. Article 3(63) defines such a model as one that displays significant generality and can competently perform a wide range of distinct tasks, excluding models used only for research, development or prototyping before they are placed on the market. - Compute guide: the [Commission's questions and answers](https://digital-strategy.ec.europa.eu/en/faqs/guidelines-obligations-gpai-providers) say models trained with more than 10^23 FLOP that can generate language, text-to-image or text-to-video are indicatively general-purpose, though this is not an absolute rule. - Open-source models: Article 53(2) exempts qualifying open-source models from points (a) and (b) only. The copyright policy and the training summary still apply, as recital 104 confirms. - Modified models: a downstream company that modifies a model becomes its provider only in exceptional cases, when the change uses more than one third of the original model's training compute, according to the same questions and answers. Its summary then covers only the data used for the modification. - Providers outside the EU: Article 54 requires providers established in third countries, including the UK and US, to appoint an authorized representative in the EU before placing a model on the market. Models trained with more than 10^25 FLOP are presumed to carry systemic risk and face extra obligations under Article 55. Those do not change the training summary itself. ## When do the training data obligations apply? The training data obligations have applied since August 2, 2025, and the Commission has been able to enforce them with fines since August 2, 2026. Article 113 sets the dates, and Article 111(3) gives older models a longer runway. | Date | What happened or happens | | --- | --- | | August 1, 2024 | The AI Act entered into force | | July 10, 2025 | The Commission published the General-Purpose AI Code of Practice | | July 24, 2025 | The Commission published the explanatory notice and training content summary template | | August 2, 2025 | Article 53 obligations began to apply | | December 5, 2025 | The notice and template were issued as Commission Communication C(2025) 8311 | | July 8, 2026 | The Digital Omnibus on AI was adopted, entering into force on July 27, 2026 | | August 2, 2026 | Commission enforcement powers and fines under Article 101 began to apply | | August 2, 2027 | Deadline for models placed on the market before August 2, 2025 | Under Article 101, the Commission may fine a provider up to 3% of its annual total worldwide turnover in the preceding financial year or €15 million, whichever is higher. The explanatory notice confirms that a non-compliant summary can be sanctioned at this level. ## What does the Commission's training data template ask for? The Commission's training data template asks for three sections: general information about the model and its data, a list of data sources and a description of data processing. The [template and explanatory notice](https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models) cover all stages of training, from pre-training to fine-tuning and alignment. Section 1 asks for the provider and model identity, the modalities in the training data and their size within broad ranges. Text is reported as less than 1 billion tokens, 1 billion to 10 trillion tokens or more than 10 trillion. Images are reported as less than 1 million, 1 million to 1 billion or more than 1 billion. Audio and video are reported as less than 10,000 hours, 10,000 to 1 million hours or more than 1 million hours. Providers also give the latest collection date and the languages covered. | Template section | Covers | What must be disclosed | | --- | --- | --- | | 2.1 Publicly available datasets | Free, downloadable third-party datasets such as Common Crawl snapshots | Name and link for each large dataset (over 3% of public data for a modality), general description of the rest | | 2.2.1 Commercially licensed datasets | Data licensed from rights holders or their representatives | Yes or no, and the modalities covered | | 2.2.2 Other private datasets | Data from intermediaries or private databases without a license from rights holders | Names if publicly known, otherwise a general description | | 2.3 Crawled and scraped data | Data the provider collected online itself | Crawler names, purpose, behavior, collection period and the top 10% of domains by content size | | 2.4 User data | Data from users of the provider's services and models | Whether used, and which services it came from | | 2.5 Synthetic data | Data generated by AI models for training, including distillation | The models used to generate it | | 2.6 Other sources | Offline sources, self-digitized media, commissioned human labeling, human reinforcement learning data | A narrative description | Small and medium-sized enterprises, including start-ups, may list the top 5% of scraped domains or 1,000 domains, whichever is lower. Section 3 then asks how the provider respected text and data mining opt-outs, whether it signed the Code of Practice and what it did to remove illegal content from the training data. ## How much detail does the summary need? The summary needs to be comprehensive in scope but not technically detailed. Recital 107 of the AI Act says it should help parties with legitimate interests, including copyright holders, to exercise their rights, for example by listing the main data collections and explaining other sources in narrative form. > "this summary should be generally comprehensive in its scope instead of technically detailed" (Recital 107, Regulation (EU) 2024/1689) The explanatory notice says the template does not require disclosure of the specific works used or the exact mix of data sources. Recital 108 adds that the AI Office will not carry out a work-by-work assessment of copyright compliance. The notice also recommends a voluntary mechanism for rights holders to ask whether content from a specific domain was used, where that domain is not listed. ## How often must the summary be updated? The summary must be updated when the provider trains its model further on data that changes the summary's content. The explanatory notice sets six-month intervals, or sooner if the new data requires a materially significant update. The summary must be public no later than the date the model is placed on the EU market. It should sit on the provider's official website and travel with the model across its public distribution channels. The same summary may cover several models or versions if the content is identical. ## What changed in 2026? In 2026 the training data rules moved from paper to enforcement, while the legislation around them was simplified. The core Article 53 duties did not change. - Enforcement: from August 2, 2026, the Commission can request information, require corrective measures and impose fines on providers of general-purpose AI models. - Digital Omnibus on AI: [Regulation (EU) 2026/1744](https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng), adopted on July 8, 2026 and published on July 24, 2026, moved the main high-risk obligations to December 2, 2027 for Annex III systems and August 2, 2028 for Annex I systems. - Content marking: the same regulation gives generative AI systems placed on the market before August 2, 2026 until December 2, 2026 to meet the Article 50(2) marking duty. - Codes of practice: the regulation removed the requirement for codes of practice to be approved by implementing act. Providers rely on codes the Commission assesses as adequate under Article 56(6). - Unchanged: the omnibus did not amend Article 53 or the Article 101 fines, so the training summary duty and its deadlines stand as originally set. The [Code of Practice page](https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai) lists signatories including Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI and OpenAI. It notes that xAI signed only the safety and security chapter, so it must show compliance with the transparency and copyright obligations by other adequate means. ## What does the template mean when you buy training data? When you buy training data, the template means you need the seller's information in a form you can report. Collect it at diligence, because reconstructing it after training is slow and sometimes impossible. - The modality and approximate size of the data, in the template's units - The languages covered and the latest collection date - Whether the license is with the rights holder or their representative (Section 2.2.1) or with an intermediary that holds no such license (Section 2.2.2) - For any scraped component, how text and data mining opt-outs were respected - Any synthetic data and the models that generated it - Measures taken to remove illegal content Licensed data carries the lightest public disclosure. The explanatory notice explains that limited detail is required for licensed data because the rights holders are already parties to the license. Data from intermediaries that lack a license from rights holders must be named if publicly known. Our [due diligence checklist](https://sellmydatatoai.net/guides/ai-training-data-due-diligence-checklist) and [data provenance guide](https://sellmydatatoai.net/guides/data-provenance-explained) cover how to gather this. ## What does the template mean for data sellers? For data sellers, the template means a buyer that provides a general-purpose AI model will report your data in its public summary, but usually without naming you. A license signed directly with the rights holder falls under Section 2.2.1, which asks only whether such data was used and in which modalities. Expect buyers to ask for the information in the list above, and to want warranties that it is accurate. If your data includes material you scraped or bought from elsewhere, expect closer questions, because that data can push a buyer into the more detailed sections. Our guide to [whether you can legally sell company data for AI](https://sellmydatatoai.net/guides/can-you-legally-sell-company-data-for-ai) covers the seller's side. ## How does Sell My Data To AI fit with the EU AI Act? Sell My Data To AI fits with the EU AI Act as a broker that helps both sides document a deal properly. Sellers sign the license directly with the buyer, and we do not take custody of raw data. Every seller passes our eight-point vetting standard, which covers ownership, third-party content and manifest accuracy. We are not a law firm, and the provider remains responsible for its own summary. Take independent legal advice on your obligations. To discuss a requirement, [book a buyer consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer) or read [how to buy AI training data](https://sellmydatatoai.net/guides/how-to-buy-ai-training-data). ## Frequently asked questions ### Does the EU AI Act training summary apply to UK and US companies? The EU AI Act training summary applies to any provider placing a general-purpose AI model on the EU market, wherever the provider is based. Article 54 requires providers established outside the EU to appoint an authorized representative in the EU before doing so. A UK or US model provider with EU customers is therefore covered. ### Do open-source model providers have to publish a training summary? Open-source model providers do have to publish a training summary. Article 53(2) exempts qualifying open-source models from the technical documentation duties in points (a) and (b), but not from the copyright policy or the public summary. Recital 104 explains that releasing weights does not reveal enough about the training data or copyright compliance. ### Will a buyer name my company in its training data summary? A buyer is unlikely to name your company if it licensed the data directly from you. Section 2.2.1 of the Commission's template asks only whether commercially licensed data was used and in which modalities. Providers may disclose more voluntarily, so agree in the license whether the buyer may name you. ### What happens if a provider's summary is incomplete? An incomplete summary can be treated as non-compliance with Article 53(1)(d). From August 2, 2026, the AI Office can request corrective measures, and the Commission can fine providers up to 3% of worldwide annual turnover or €15 million, whichever is higher. Providers of models placed on the market before August 2, 2025 may state and justify information gaps. ### Does fine-tuning a model make you responsible for a training summary? Fine-tuning a model makes you responsible for a training summary only in exceptional cases. The Commission's questions and answers say a downstream modifier becomes a provider when the modification uses more than one third of the original model's training compute. Your summary then covers only the data used for the modification, with a link to the original summary. ## Sources - [Regulation (EU) 2024/1689 (Artificial Intelligence Act)](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng), EUR-Lex, Publications Office of the European Union, accessed November 2026 - [Explanatory notice and template for the public summary of training content for general-purpose AI models](https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models), European Commission, Shaping Europe's digital future, accessed November 2026 - [Guidelines on obligations for general-purpose AI providers: questions and answers](https://digital-strategy.ec.europa.eu/en/faqs/guidelines-obligations-gpai-providers), European Commission, Shaping Europe's digital future, accessed November 2026 - [The General-Purpose AI Code of Practice](https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai), European Commission, Shaping Europe's digital future, accessed November 2026 - [Regulation (EU) 2026/1744 (Digital Omnibus on AI)](https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng), EUR-Lex, Publications Office of the European Union, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Agent trajectories and RL environments explained URL: https://sellmydatatoai.net/guides/agent-trajectories-and-rl-environments Reviewed: 2026-11-08 An agent trajectory is the step-by-step record of an AI agent or a person completing a task in software: each observation, click, tool call and result. An RL environment is the replayable sandbox where those tasks run and are scored. AI labs buy both to train agents with reinforcement learning, and they value realistic tasks whose outcomes can be checked automatically. ## Key takeaways - A trajectory records the steps of a task, while an environment lets a buyer rerun the task and generate new trajectories. - TechCrunch reported in September 2025 that Anthropic leaders had discussed spending more than $1 billion on RL environments over the following year, citing The Information. - Benchmarks such as SWE-bench, OSWorld and tau-bench show what good tasks look like: real software, a clear goal and automatic scoring. - Buyers must check personal data in logs and screenshots, the rights to the software inside the environment and whether the scoring can be gamed. AI labs are moving from models that answer questions to agents that complete tasks in software. Training those agents needs a different kind of data: records of tasks being done, and sandboxes where tasks can be done again and scored. ## What is an agent trajectory? An agent trajectory is the record of an agent working through a task, step by step. DataVendor describes it as each click, tool call, observation and final result in a workflow. The agent may be an AI model or a person whose actions are logged. | Field | What it holds | | --- | --- | | Task | The goal, such as resolving a support ticket or fixing a failing test | | Observation | What the agent could see, such as a screenshot, page structure or tool output | | Action | What the agent did, such as a click, keystroke, API call or code edit | | Reasoning | Notes or model output explaining the action, where captured | | Result | The new state after the action | | Outcome | Whether the task succeeded, and how that was judged | Trajectories that fail are useful too, provided the outcome is recorded honestly. See the [agent trajectory](https://sellmydatatoai.net/glossary/agent-trajectory) glossary entry for a short definition. ## What is an RL environment? An RL environment is a controlled copy of a piece of software where an agent can attempt a task, be scored and start again. TechCrunch describes environments as training grounds that simulate what an AI agent would be doing in a real software application, such as a simulated browser where an agent is graded on completing a purchase. - An initial state that can be reset, such as a configured desktop, repository or database - A task description - The actions and tools available to the agent - A verifier or scoring script that judges the outcome DataVendor explains the difference for buyers: a replayable environment adds the software or sandbox needed to rerun a task, so buyers can generate fresh trajectories and train models against verified outcomes. See the [RL environment](https://sellmydatatoai.net/glossary/rl-environment) glossary entry. ## Why are AI labs buying trajectories and environments? AI labs are buying trajectories and environments because reinforcement learning on checkable tasks has become a main route to better agents. The Allen Institute for AI's Tulu 3 paper, submitted to [arXiv](https://arxiv.org/abs/2411.15124) in November 2024, introduced the term Reinforcement Learning with Verifiable Rewards for training against outcomes that can be checked automatically. Spending has followed. TechCrunch reported in September 2025 that Anthropic leaders had discussed spending more than $1 billion on RL environments over the following year, citing The Information. The same article named Scale AI, Surge and Mercor among data companies moving into environments, alongside start-ups such as Mechanize and Prime Intellect. > "All the big AI labs are building RL environments in-house" (Jennifer Li, Andreessen Horowitz, quoted by TechCrunch, September 2025) Not everyone is convinced. Ross Taylor of General Reasoning told TechCrunch that people are underestimating how difficult it is to scale environments. For buyers, that difficulty is the reason well-built environments and real-world trajectories carry value. ## What do benchmarks show about good agent tasks? Benchmarks show that good agent tasks use real software, have a clear goal and are scored automatically. Three widely cited benchmarks illustrate the pattern. | Benchmark | Tasks | Environment | Scoring | Result at publication | | --- | --- | --- | --- | --- | | SWE-bench (2023) | 2,294 real GitHub issues from 12 popular Python repositories | The repository codebase and the issue text | Whether the edit resolves the issue | Best model resolved 1.96% of issues | | OSWorld (2024) | 369 computer tasks | Real Ubuntu, Windows and macOS environments | A custom evaluation script per task | Humans 72.36%, best model 12.24% | | Tau-bench (2024) | Customer service tasks with a simulated user | Domain API tools and policy rules | Final database state against the goal state | GPT-4o under 50%, pass^8 under 25% in retail | Each one is built from the same parts a buyer should look for: a realistic starting state, a task drawn from real work and an objective check. [SWE-bench](https://arxiv.org/abs/2310.06770) uses real issues and pull requests. [OSWorld](https://arxiv.org/abs/2404.07972) gives each task an initial-state configuration and an evaluation script. [Tau-bench](https://arxiv.org/abs/2406.12045) adds a pass^k metric to measure reliability over repeated trials. ## What makes trajectory data valuable? Trajectory data is valuable when it is complete, real and checkable. A log of clicks with no task or outcome attached is hard to train on. - Every step is captured, including what the agent saw before acting - The task and the outcome are recorded, including failures - The work happened in software people actually use, not a toy simulation - The operators were skilled at the task, which matters for expert domains - An environment exists to replay the task or generate new attempts - Personal data and credentials have been removed from screenshots and logs DataVendor notes that on its marketplace, raw trajectories, internal support workflows and software-use logs can be packaged as raw data ZIPs, while environments cannot be listed on their own and are sold with tasksets. Our [agent trajectory data](https://sellmydatatoai.net/buy-ai-training-data/agent-trajectory-data) page describes formats buyers ask for. ## Where does trajectory data come from? Trajectory data comes from three main places: companies that log how staff use software, AI product companies that record agent runs and specialist vendors that commission experts to perform tasks. - Support and operations teams, whose ticket workflows show multi-step problem solving - Software companies, whose product event data and internal tooling logs show real use - AI product companies, whose agent run traces, user feedback and ratings show where agents succeed and fail - Data vendors, who pay experts to complete tasks in prepared environments If you hold this kind of data, our pages for [software companies](https://sellmydatatoai.net/sell-your-data/software-companies) and [call centers](https://sellmydatatoai.net/sell-your-data/call-centers) explain what buyers ask for. ## What should buyers check before licensing trajectories or environments? Before licensing trajectories or environments, buyers should check rights, privacy and scoring quality. Each creates a different kind of risk. | Risk | What to check | | --- | --- | | Personal data | Screenshots, page captures and logs often show names, emails and account details | | Credentials | Tokens, passwords and API keys captured in tool calls or terminal output | | Third-party software | Whether the licenses of applications inside the environment allow this use | | Employer and client rights | Whether staff activity logs and client systems can be licensed, and under which privacy notice | | Scoring quality | Whether the verifier can be passed without completing the task | | Reproducibility | Whether the environment resets to the same state and gives the same score | | Contamination | Whether tasks overlap with public benchmarks you report on | | Derived rights | Whether you may generate new trajectories, modify the environment and keep trained policies | Test a sample of tasks yourself before buying a full set. [How to evaluate a dataset sample](https://sellmydatatoai.net/guides/how-to-evaluate-dataset-samples) covers the method, and [evaluation datasets](https://sellmydatatoai.net/buy-ai-training-data/evaluation-datasets) covers task sets used for scoring rather than training. ## How are trajectories and environments priced? Trajectories and environments are usually priced per task or per task set rather than per record. AIData.Broker lists the task among the standard pricing units, as quoted by ddiy.co, and DataVendor sets task and environment rates across each listing's tasksets. Public price ranges for this data are scarce. Treat any quote as specific to the task difficulty, the expertise needed and whether an environment is included. [AI training data pricing models](https://sellmydatatoai.net/guides/ai-training-data-pricing-models) explains the structures in use. ## How does the EU AI Act treat this data? The EU AI Act treats human trajectory and reinforcement learning data as a training source that providers must describe. The Commission's training content template names datasets labeled by humans commissioned by the provider and human-generated reinforcement learning data in Section 2.6, while AI feedback used in reinforcement learning falls under synthetic data in Section 2.5. Keep a record of where each task set and trajectory came from so you can complete those sections. Our guide to [the EU AI Act and training data transparency](https://sellmydatatoai.net/guides/eu-ai-act-training-data-transparency) explains the full template. To source trajectories or environments, [book a buyer consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer). ## Frequently asked questions ### What is the difference between a trajectory and an environment? A trajectory is a recording of one attempt at a task, while an environment is the sandbox where tasks can be attempted again and scored. Trajectories suit supervised training on demonstrated behavior. Environments suit reinforcement learning, because a model can generate many new attempts and learn from the scores. ### Can a company sell logs of how its staff use software? A company can sometimes sell logs of how its staff use software, but it needs to check its privacy notices, employment terms and the licenses of the software being used. Screenshots and logs often contain personal data, client information and credentials, which must be removed or the records excluded. Take independent legal advice before licensing staff activity data. ### Why do verifiable outcomes matter? Verifiable outcomes matter because reinforcement learning needs a reliable score for each attempt. The Tulu 3 paper from November 2024 named the approach Reinforcement Learning with Verifiable Rewards. If the score can be passed without doing the task, the model learns to pass the check rather than do the work, so buyers test verifiers before buying. ### Are public benchmarks like SWE-bench available to train on? Public benchmarks like SWE-bench are designed for evaluation, and training on them makes your reported scores unreliable. They are useful as a model of what good tasks look like: real issues, a realistic codebase and an objective check. Buyers looking for training data usually want private tasks built the same way but kept out of public benchmarks. ## Sources - [AI training data marketplaces](https://datavendor.ai/blog/ai-training-data-marketplaces), DataVendor, accessed November 2026 - [Silicon Valley bets big on environments to train AI agents](https://techcrunch.com/2025/09/21/silicon-valley-bets-big-on-environments-to-train-ai-agents/), TechCrunch, accessed November 2026 - [Tulu 3: Pushing frontiers in open language model post-training](https://arxiv.org/abs/2411.15124), arXiv (Lambert et al.), accessed November 2026 - [SWE-bench: Can language models resolve real-world GitHub issues?](https://arxiv.org/abs/2310.06770), arXiv (Jimenez et al., ICLR 2024), accessed November 2026 - [OSWorld: Benchmarking multimodal agents for open-ended tasks in real computer environments](https://arxiv.org/abs/2404.07972), arXiv (Xie et al.), accessed November 2026 - [Tau-bench: A benchmark for tool-agent-user interaction in real-world domains](https://arxiv.org/abs/2406.12045), arXiv (Yao et al.), accessed November 2026 - [How much is company data worth to AI?](https://ddiy.co/how-much-is-company-data-worth-to-ai/), ddiy.co, accessed November 2026 - [Explanatory notice and template for the public summary of training content for general-purpose AI models](https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models), European Commission, Shaping Europe's digital future, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How to run an AI training data procurement URL: https://sellmydatatoai.net/guides/ai-training-data-procurement Reviewed: 2026-11-08 An AI training data procurement runs in six gated stages: requirements, a brief or RFP, a scored shortlist, rights and quality due diligence, a contract with acceptance terms, and acceptance testing at delivery. Each stage needs a named owner and a written pass condition, because rights failures found after training are far more expensive to fix than failures found before signing. ## Key takeaways - Treat rights evidence as a pass or fail gate before you score sellers on price or volume. - Use a short brief for owners licensing existing records and a full RFP only for commissioned human data across several vendors. - Meta's Llama 2 team reported that different annotation vendors produced markedly different downstream model performance, so test vendors on your own task. - US fair use is unsettled for training: a June 2025 ruling in Bartz v. Anthropic favored training, while the Third Circuit rejected fair use in Thomson Reuters v. Ross in September 2026. - Write acceptance tests into the contract before delivery, with a holdback tied to passing them. Buying training data looks like ordinary vendor procurement until something goes wrong. A software license can be terminated. Data that has already shaped a model's weights cannot practically be returned, so the checks have to happen before training starts. This guide sets out a procurement process an AI team, a legal team and a procurement lead can run together. It complements our guide on [how to buy AI training data](https://sellmydatatoai.net/guides/how-to-buy-ai-training-data), which covers sourcing routes and pricing, by focusing on stages, owners, gates and the paperwork each stage should produce. ## What are the stages of an AI training data procurement? An AI training data procurement has six stages, and each one ends at a gate with a written pass condition. Moving to the next stage without passing the gate is how teams end up training on data they cannot defend. | Stage | Lead owner | Output | Gate to pass | | --- | --- | --- | --- | | 1. Requirements | ML lead | Written requirement with the target evaluation | Budget holder signs off the use and the budget model | | 2. Brief or RFP | Procurement lead | Request sent to sellers under a mutual NDA | At least two responses that answer every required field | | 3. Shortlist | ML lead and procurement | Scored comparison of sellers | Every shortlisted seller passes the rights gate | | 4. Due diligence | Legal and privacy counsel | Diligence file and sample test results | No open rights issue and a sample that meets the spec | | 5. Contract | Legal counsel | Signed license with acceptance schedule | Acceptance tests, holdback and remedies agreed in writing | | 6. Acceptance | ML lead and security | Acceptance report and provenance record | Delivered data matches the manifest and passes the tests | ## Who should own a training data procurement? A training data procurement needs one accountable owner, usually the ML lead whose model will use the data, plus named owners for legal, privacy, security and budget. Data deals stall when nobody is sure who can say no. - ML lead: writes the requirement, designs the sample tests and signs the acceptance report - Procurement lead: runs the brief or RFP, keeps the scoring record and manages timelines - Legal counsel: reviews rights evidence, negotiates the license and tracks copyright risk - Privacy counsel: checks privacy notices, de-identification and state privacy law exposure - Security lead: approves the storage environment, access controls and deletion process - Budget holder: approves the pricing model, the payment schedule and any holdback ## How do you set requirements before going to market? Requirements should start from the evaluation the data is meant to move, not from a volume target. If you cannot name the score you expect to improve, you cannot judge a sample or accept a delivery. Volume is often the wrong lever. Meta's Llama 2 paper reports that its team set aside millions of third-party fine-tuning examples, collected 27,540 higher-quality annotations from its own vendor-based effort and saw results notably improve. The same paper notes that different annotation platforms and vendors produced markedly different downstream model performance. Write down the training use, domain, volume range, time range, formats, rights, exclusivity, de-identification standard, delivery method and budget model. Our guide on [how to write a training data request](https://sellmydatatoai.net/guides/how-to-write-a-training-data-request) explains each field with examples. ## Should you issue an RFP or a short brief? Issue a full RFP when you are commissioning new human data from several established vendors, and a short brief when you want to license records a company already holds. Data owners such as law firms or manufacturers rarely answer long RFPs, because selling data is not their business. | Factor | Full RFP | Short brief | | --- | --- | --- | | Typical seller | Training-data companies and annotation vendors | Companies licensing their own records | | What you are buying | Work to be done, such as demonstrations or preference labels | Rights to data that already exists | | Length | Detailed scope, quality metrics, staffing and security questions | One to two pages of required fields | | Price basis | Per task, per hour or per project | Flat license fee, per unit or recurring access | | Main risk | Quality drift once work scales up | Rights gaps in the seller's records | For commissioned work, our guide to [sourcing expert and human-generated data](https://sellmydatatoai.net/guides/sourcing-expert-and-human-data) covers vendors, pricing models and quality control in detail. ## How do you build and score a shortlist? Build the shortlist in two passes: a pass or fail rights gate first, then a scored comparison of the sellers who pass. Scoring rights alongside price lets a cheap seller with weak evidence outscore a sound one. 1. Rights gate: the seller can show who may sign, how it owns the data and which contracts and privacy notices apply 2. Fit: the data matches the domain, time range, format and languages in your requirement 3. Sample result: a seeded random sample performs on your test plan 4. Documentation: a manifest, a datasheet and hashes tied to an exact version 5. Commercial terms: price, exclusivity, payment schedule and willingness to accept a holdback ## What due diligence should run before the contract? Due diligence should confirm what is in the dataset, whether the seller can transfer it for AI use and who carries the risk if it is not as described. Nixon Peabody's October 2026 alert frames this as five questions. 1. What is actually in the dataset? 2. Does the seller have the right to transfer data for AI use? 3. Can sensitive material be removed without destroying the data's value? 4. What can the buyer do with the data and with what it creates? 5. Who bears the risk if the dataset is not what the parties expected? Our [AI training data due diligence checklist](https://sellmydatatoai.net/guides/ai-training-data-due-diligence-checklist) lists the documents to request for each question. Two US points shape how hard you push. First, the Federal Trade Commission warned in a February 13, 2024 post that it may be unfair or deceptive for a company to adopt more permissive data practices, such as using consumer data for AI training, through a quiet or retroactive change to its terms or privacy policy. Ask for the privacy notice in force when the data was collected. Nixon Peabody also notes that the FTC has required companies to delete models and algorithms built with improperly obtained data. Second, US fair use law on training is unsettled. On June 23, 2025, Judge William Alsup held in Bartz v. Anthropic that training on lawfully acquired books was fair use but that building a library from pirated copies was not. In September 2026 the Third Circuit rejected Ross Intelligence's fair use defense for training a non-generative legal search tool on Westlaw headnotes, according to Nieman Lab. Our [AI training data lawsuits tracker](https://sellmydatatoai.net/ai-training-data-lawsuits) follows the open cases. Take independent legal advice on your own use. ## What should the contract settle? The contract should settle permitted use, ownership of derived assets, warranties, indemnities and what happens if rights fail after training. For procurement, the key addition is an acceptance schedule that turns your sample tests into contract terms. - Acceptance schedule: the tests, thresholds and inspection period that define a conforming delivery - Holdback: part of the fee retained until acceptance, which Nixon Peabody lists among the risk tools alongside indemnities, liability caps, insurance and termination rights - Staged access: a limited sample in a controlled environment first, then wider access as agreed conditions are met - Change control: how the seller notifies you of changes to the dataset or its sources - Deletion certificate: written confirmation of deletion at expiry or on a rights failure The [AI data licensing agreement checklist](https://sellmydatatoai.net/guides/ai-data-licensing-agreement-checklist) covers the full clause list from both sides of the deal. ## How do you accept a training data delivery? Accept a delivery only after the files match the manifest and pass the tests in the acceptance schedule within the inspection period. Check record counts, date ranges, formats and hashes first, then run quality and personal data checks on a random sample. If you plan a pilot before the full license, our guide to [training data pilot design](https://sellmydatatoai.net/guides/training-data-pilot-design) explains sample sizes, held-out evaluation and staged payment. ## AI training data procurement checklist Copy this checklist into your procurement tracker. Each row names the evidence that closes the item. | Stage | Item | Evidence that closes it | Owner | | --- | --- | --- | --- | | Requirements | Training use and target evaluation named | Signed requirement document | ML lead | | Requirements | Budget model approved | Budget holder sign-off | Budget holder | | Brief or RFP | Mutual NDA in place with each seller | Countersigned NDA | Legal counsel | | Brief or RFP | Every required field answered | Response log | Procurement lead | | Shortlist | Rights gate passed | Seller's signatory, ownership and contract evidence | Legal counsel | | Shortlist | Seeded random sample tested | Sample test report | ML lead | | Due diligence | Privacy notice at collection reviewed | Dated copy of the notice | Privacy counsel | | Due diligence | De-identification method checked | Written method and re-identification test | Privacy counsel | | Due diligence | Third-party and open-source content cleared | Clearance log | Legal counsel | | Due diligence | Storage environment approved | Security review | Security lead | | Contract | Acceptance schedule and holdback agreed | Signed license | Legal counsel | | Contract | Derived assets and rights failure remedies agreed | Signed license | Legal counsel | | Acceptance | Delivery matches manifest and hashes | Acceptance report | ML lead | | Acceptance | Provenance record filed | License, manifest, hashes and diligence file stored together | Procurement lead | ### If you place models on the EU market Providers of general-purpose AI models placed on the EU market must keep a copyright policy and publish a summary of training content under Article 53 of the EU AI Act. The provenance record from stage 6 is what you will draw on. Our guide to [the EU AI Act and training data transparency](https://sellmydatatoai.net/guides/eu-ai-act-training-data-transparency) covers the summary in detail. ## How does Sell My Data To AI fit into a procurement? Sell My Data To AI can run stages 2 and 3 for you as an independent broker. Buyers pay nothing to submit a requirement or review a shortlist, and any fee is agreed in writing before an introduction. Every seller passes our eight-point vetting standard before any buyer sees their data, covering authority to sign, ownership, contract restrictions, privacy notices, third-party content, de-identification, sample quality and manifest accuracy. We do not take custody of raw data. You review redacted samples under a mutual NDA and sign the license directly with the seller. We are a broker, not a law firm, so take your own legal advice before signing. To start, [submit a data requirement](https://sellmydatatoai.net/request-data) or [book a buyer consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer). ## Frequently asked questions ### How long does an AI training data procurement take? An AI training data procurement usually takes weeks rather than days, and most of the time goes on rights diligence and contract review. Commissioned human data adds vendor ramp-up and calibration. Licensing existing company records adds time for the owner to gather privacy notices, contracts and IP assignments. Plan the gates first and the timeline follows from how quickly each owner can close them. ### Do we need an RFP to buy training data? You need a full RFP only when you are commissioning new work, such as demonstrations or preference labels, from several established vendors. To license data a company already holds, a one to two page brief works better, because most data owners are not set up to answer long RFPs. Both should state the training use, rights needed and budget model. ### Is training on licensed data protected by fair use? A license does not depend on fair use. It gives you contractual permission for the uses it lists. Fair use matters when you rely on data without a license, and US courts are split so far: Bartz v. Anthropic favored training on lawfully acquired books in June 2025, while the Third Circuit rejected fair use in Thomson Reuters v. Ross in September 2026. Take independent legal advice. ### What is a holdback in a data license? A holdback is part of the license fee that the buyer keeps back until the delivered data passes acceptance tests. It gives the seller a reason to fix nonconforming records quickly. Nixon Peabody lists holdbacks alongside indemnities, liability caps, insurance and termination rights as tools to allocate risk when a dataset is not what the parties expected. ## Sources - [Shopping for AI training data? Five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody LLP, accessed November 2026 - [AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive](https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/02/ai-other-companies-quietly-changing-your-terms-service-could-be-unfair-or-deceptive), Federal Trade Commission, accessed November 2026 - [District court rules AI training can be fair use in Bartz v. Anthropic](https://www.akingump.com/en/insights/ai-law-and-regulation-tracker/district-court-rules-ai-training-can-be-fair-use-in-bartz-v-anthropic), Akin Gump Strauss Hauer and Feld, accessed November 2026 - [Federal appeals court upholds Thomson Reuters' landmark AI copyright win](https://www.niemanlab.org/2026/09/federal-appeals-court-upholds-thomson-reuters-landmark-ai-copyright-win/), Nieman Lab, accessed November 2026 - [Llama 2: Open foundation and fine-tuned chat models](https://arxiv.org/abs/2307.09288), arXiv (Touvron et al., Meta, 2023), accessed November 2026 - [Regulation (EU) 2024/1689 (Artificial Intelligence Act)](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng), EUR-Lex, Publications Office of the European Union, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How to write a training data request sellers can answer URL: https://sellmydatatoai.net/guides/how-to-write-a-training-data-request Reviewed: 2026-11-08 A training data request sellers can answer states the training use and then fixes ten fields: domain, volume, time range, formats, languages, rights needed, exclusivity, de-identification standard, delivery method and budget model. Give a range or a firm rule for each field. Sellers quote quickly against specific fields and slowly, or not at all, against open questions. ## Key takeaways - Open with the training use, because rights, price and format all depend on it. - Give every field a range or a rule a seller can check against its own records. - Name the de-identification standard, such as HIPAA safe harbor or expert determination for health data. - State the budget model as well as the amount: Reddit's S-1 shows its data licenses were priced on a fixed fee or usage-based fee. - Leave out anything you would not want a competitor to read, and send the full request only under a mutual NDA. Most data requests fail before any seller reads the second paragraph. A request that says 'we need high-quality legal data' gets either silence or a sales call. A request that names the record types, the date range and the rights needed gets a quote. This guide lists the fields a request should cover, what to write in each and why sellers need it. It is the requirement stage of a [training data procurement](https://sellmydatatoai.net/guides/ai-training-data-procurement). ## What makes a training data request answerable? A training data request is answerable when a seller can check each field against its own records and reply yes, no or partly. Every field should be a range, a list or a rule, never an adjective. Start with the training use. Pre-training, supervised fine-tuning, preference tuning, evaluation and reinforcement learning need different data shapes and different rights, so the same archive can carry very different prices depending on the use. ## Which fields should a training data request include? A training data request should include ten fields after the training use. The table shows what to write in each and why the seller needs it. | Field | What to write | Example | Why sellers need it | | --- | --- | --- | --- | | Domain | Industry, record types and source systems | Closed commercial contract files from US law firms | Lets the seller find the right archive | | Volume | Minimum and target, in the unit you will price on | 5,000 to 20,000 complete matter files | Sets scope and price | | Time range | Earliest and latest creation dates | Created 2015 to 2025 | Rules out stale or out-of-scope records | | Formats | File types and structure | Native DOCX and PDF with a JSON manifest | Shows the conversion work needed | | Languages | Languages and regional variants | US English only | Narrows the archive | | Rights needed | Uses, models, term and territory | Training and evaluation of internal models for 3 years, worldwide | Drives legal review and price | | Exclusivity | Exclusive, non-exclusive or a limited exclusive | Exclusive for evaluation use for 12 months | Changes price and who else may buy | | De-identification | Standard and who applies it | Client and counterparty names removed, method documented | Shows the preparation cost | | Delivery | Method, environment and schedule | Secure transfer in two tranches with hashes | Sets the logistics and security terms | | Budget model | Pricing basis and range | Flat license fee, with a holdback until acceptance | Tells the seller whether a deal is viable | ## How do you specify domain and volume? Specify domain as record types from named source systems, and volume as a minimum and a target in the unit you will price on. Sellers search their archives by system and record type, not by topic. Pick the unit carefully. Matter files, support tickets, hours of audio, repositories and comparisons each count work differently. If you will pay per unit, the unit in the request becomes the unit in the contract. Our [training data catalog](https://sellmydatatoai.net/buy-ai-training-data) lists the usual unit for each data type, such as support tickets or expert domain records. A smaller, cleaner set can beat a large one. The LIMA paper from 2023 fine-tuned a 65 billion parameter model on only 1,000 curated prompts and responses, and human raters judged its answers equivalent or preferred to GPT-4 in 43% of cases. State a minimum you can use rather than the largest number you can imagine. ## How do you specify time range and formats? Give the earliest and latest creation dates you will accept, and the file types and structure you need on delivery. Dates matter for relevance, for privacy notices in force at collection and for whether records predate widespread AI-generated content. For formats, say whether you want native files, a converted format such as JSONL, or both. Ask for documentation as a datasheet, a format proposed by Timnit Gebru and colleagues that records a dataset's motivation, composition, collection process and recommended uses, or as machine-readable metadata in the MLCommons Croissant format. ## What rights should a training data request ask for? Ask for the narrowest rights that cover your real use: which uses, which models, which affiliates, what term and what territory. Broad rights cost more and take longer to clear, and a seller's lawyer will ask why you need them. - Uses: training, fine-tuning, evaluation, or all three - Models: named internal models, future versions, or models offered to customers - Derived assets: ownership of annotations, embeddings, evaluation sets and fine-tuned weights - Onward transfer: whether cloud vendors, subprocessors or model partners may touch the data - Term and expiry: what happens to trained models and retained copies at the end Do not treat US fair use as a substitute for the rights you need. US courts have reached different results on training so far, and a license gives you certainty a defense cannot. Our [data rights](https://sellmydatatoai.net/data-rights) page covers the background. ## Should a training data request ask for exclusivity? Ask for exclusivity only when your advantage depends on competitors not training on the same data, as with a private evaluation set. Otherwise ask for a non-exclusive license or a limited exclusive by field, term or territory. Our guide to [exclusive vs non-exclusive data licenses](https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses) sets out the options and how they change price. ## How do you state de-identification requirements? State the legal standard the data must meet, the identifiers to remove and who will do the work. Sellers price de-identification as real effort, and they cannot quote it against 'anonymized'. - Health data: HIPAA allows de-identification by expert determination or by the safe harbor method of removing 18 types of identifier, according to the US Department of Health and Human Services. - California consumer data: Nixon Peabody notes that under the CCPA, data counts as de-identified only if it cannot reasonably be linked to an individual, the holder publicly commits not to re-identify it and recipients are contractually bound. - Business records: list the identifiers to remove, such as client, counterparty and employee names, and ask for a documented method and a re-identification test on a sample. Our guide to [de-identification for AI training data](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data) explains the methods sellers use. ## How should you describe delivery? Describe the delivery method, the environment the data will live in and the schedule. Nixon Peabody notes that for especially sensitive archives, source data can stay in a secure environment where defined training or evaluation takes place, without transferring unrestricted copies. If you can accept that model, say so in the request. It widens the pool of sellers willing to talk. Also ask for a manifest with record counts and file hashes, and for delivery in tranches if you plan to pilot first. ## How do you state a budget model? State the pricing basis as well as the range: flat fee, per unit, recurring access or revenue share. Sellers can judge viability from a model faster than from a number. Public filings show how large sellers structure this. Reddit's 2024 S-1 states that the transaction price in its data licensing arrangements is generally a fixed fee or usage-based fee, recognized over the license period, with payment usually due within 30 days. Our guide to [AI training data pricing models](https://sellmydatatoai.net/guides/ai-training-data-pricing-models) compares the structures, and the [AI data price index](https://sellmydatatoai.net/ai-data-price-index) lists published ranges. ## What should you leave out of a training data request? Leave out model architecture, roadmap detail and anything you would not want a competitor to read. A first request can go to many sellers, so keep it to the fields above and send detail only under a mutual NDA. > **Rule of thumb** If a seller cannot answer a field from its own records, rewrite the field. If a competitor could learn your strategy from it, cut it. ### If you hold or buy EU or UK personal data Add the lawful basis you expect the seller to rely on and whether data will leave the EU or UK. Under the EU AI Act, providers of general-purpose AI models also publish a training content summary, so ask sellers for the source descriptions you will need for it. ## How do you send a request through Sell My Data To AI? Our [data requirement form](https://sellmydatatoai.net/request-data) is built around the fields in this guide. Buyers pay nothing to submit a requirement or review a shortlist, and any fee is agreed in writing before an introduction. Every seller we shortlist has passed our [eight-point vetting standard](https://sellmydatatoai.net/how-we-vet). ## Frequently asked questions ### How detailed should a training data request be? A training data request should be detailed enough that a seller can answer each field from its own records, which usually means one to two pages. Give a range or rule for domain, volume, time range, formats, languages, rights, exclusivity, de-identification, delivery and budget model. Keep model and roadmap detail for later, under a mutual NDA. ### Should I put a budget in a data request? Yes, put at least a budget model and a range in a data request. Sellers spend real effort preparing samples and rights evidence, and without a budget they cannot tell whether a deal is viable. Reddit's 2024 S-1, for example, describes its data licenses as priced on a fixed fee or usage-based fee, which is the kind of basis sellers expect to see. ### What does 'rights needed' mean in a data request? Rights needed are the specific uses the license must allow: training, fine-tuning or evaluation, which models and affiliates, the term, the territory and ownership of derived assets such as embeddings and fine-tuned weights. Ask for the narrowest rights that cover your real use, because broad rights cost more and take longer for the seller's lawyers to clear. ### Can I ask a seller to de-identify data for me? You can ask a seller to de-identify data, and many will, but name the standard and ask for the method in writing. For US health data, HHS describes two HIPAA methods: expert determination and the safe harbor removal of 18 identifier types. For other records, list the identifiers to remove and ask for a re-identification test on a sample. ### Who should I send a training data request to? Send a short version to sellers who plausibly hold the records, such as companies in the domain, training-data companies and marketplaces, and the full version only under a mutual NDA. Our directory of [data buyers and platforms](https://sellmydatatoai.net/data-buyers) lists the companies active in this market, and a broker can circulate the request to vetted sellers for you. ## Sources - [Guidance regarding methods for de-identification of protected health information](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html), US Department of Health and Human Services, accessed November 2026 - [Shopping for AI training data? Five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody LLP, accessed November 2026 - [Reddit, Inc. Form S-1 registration statement](https://www.sec.gov/Archives/edgar/data/1713445/000162828024006294/reddits-1q423.htm), US Securities and Exchange Commission (EDGAR), accessed November 2026 - [LIMA: Less is more for alignment](https://arxiv.org/abs/2305.11206), arXiv (Zhou et al., 2023), accessed November 2026 - [Datasheets for datasets](https://arxiv.org/abs/1803.09010), arXiv (Gebru et al.), accessed November 2026 - [Croissant metadata format](https://mlcommons.org/working-groups/data/croissant/), MLCommons, accessed November 2026 - [Regulation (EU) 2024/1689 (Artificial Intelligence Act)](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng), EUR-Lex, Publications Office of the European Union, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How to design a training data pilot before a full license URL: https://sellmydatatoai.net/guides/training-data-pilot-design Reviewed: 2026-11-08 A training data pilot tests a limited slice of a dataset on an agreed experiment before you sign the full license. Design it with two samples, a quality audit of a few hundred random records and a training slice, a held-out evaluation the seller never sees, written acceptance criteria and staged payment, so the full license is only signed once the data has proved itself. ## Key takeaways - Use two samples: a random audit sample for defects and a training slice to measure model effect. - If an audit of n random records finds no defects, the upper 95% confidence limit on the defect rate is about 3 divided by n, so 300 clean records puts it below about 1%. - Keep the evaluation set held out from the seller and check it for paraphrased overlap, not only exact matches. - Write acceptance criteria before the pilot starts and attach them to the full license. - Stage payment as a pilot fee, a license fee on conversion and a holdback until delivery passes acceptance. A pilot is the cheapest point at which to find out that a dataset does not help your model. After a full license, the money is spent. After training, the data is in the weights. This guide explains how to size a pilot, set up the evaluation, write acceptance criteria and structure payment. It builds on our guide to [evaluating a dataset sample](https://sellmydatatoai.net/guides/how-to-evaluate-dataset-samples), which covers checking the records themselves. ## What is a training data pilot? A training data pilot is a limited, paid test of a dataset on an experiment both sides agree in advance. It sits between a sample review and a full license. Nixon Peabody's October 2026 diligence alert describes the same staging from the legal side: the buyer starts with a limited sample in a controlled environment, validates rights and quality, and expands access only as agreed conditions are met. | Step | Question it answers | Data involved | Rights needed | | --- | --- | --- | --- | | Sample review | Is the data what the manifest says? | A seeded random sample | Evaluation under a mutual NDA | | Pilot | Does the data improve our model on our task? | An audit sample plus a training slice | Pilot license for testing only | | Full license | Can we train and ship on the whole dataset? | The full delivery in tranches | Training rights for named models and uses | ## How big should a training data pilot be? A pilot needs two samples of different sizes: an audit sample to estimate defect rates and a training slice to measure model effect. Size the audit sample with simple statistics and the training slice by the training use. ### Sizing the audit sample The audit sample is a random draw of records that your team reviews by hand or by script for errors, personal data, duplicates and off-spec records. Its size sets how precisely you can estimate a defect rate. | Records audited | Margin of error | | --- | --- | | 100 | About 9.8 percentage points | | 400 | About 4.9 percentage points | | 1,000 | About 3.1 percentage points | | 2,500 | About 2.0 percentage points | The figures follow from the standard formula for a proportion, 1.96 times the square root of p times 1 minus p over n. Margins shrink when the true rate is far from 50%, so they are an upper bound. For rare defects such as residual personal data, use the rule of three. As Thomas Newman summarized in JAMA in 1995, citing Hanley and Lippman-Hand, if no events are observed in n trials, the upper limit of the 95% confidence interval for their frequency is about 3 divided by n. An audit of 300 random records with no personal data found supports a rate below about 1%, and 3,000 clean records supports below about 0.1%. ### Sizing the training slice The training slice should be big enough to move your evaluation if the data is useful. For fine-tuning data, published work suggests that can be small: the LIMA paper fine-tuned a 65 billion parameter model on 1,000 curated examples, and Meta's Llama 2 team stopped collecting supervised fine-tuning data at 27,540 annotations, finding that tens of thousands were enough for a high-quality result. Pre-training data is harder to pilot, because small slices rarely move a large model. A common approach is to train a smaller proxy model on mixtures with and without the candidate data. Microsoft's Phi-4 report describes choosing its data mixture through ablations at 7 billion parameters on a shorter 1 trillion token run, after observing a high rank correlation between 7 billion and 14 billion parameter results across mixtures. ## How do you set up a held-out evaluation for a pilot? A held-out evaluation is a fixed test set that neither the seller nor the pilot data has touched. Record your current model's score on it before the pilot so you can measure the change. 1. **Fix the evaluation set**: Choose or build tasks that reflect the deployment you care about and freeze them before you see the pilot data. 2. **Record a baseline**: Score your current model, and record the settings, seed and compute used. 3. **Run a control**: Train on the same compute budget with data you already hold, so you separate the effect of new data from the effect of extra training. 4. **Run the pilot**: Train with the pilot slice under the same settings, and repeat with more than one seed if budget allows. 5. **Check for contamination**: Search the pilot slice for overlap with the evaluation set, including paraphrases, before you trust any gain. Exact matching is not enough. Shuo Yang and colleagues showed in 2023 that paraphrased or translated test items slip past n-gram decontamination, and found that 8% to 18% of the HumanEval coding benchmark overlapped with the RedPajama-Data-1T and StarCoder-Data pre-training sets. Our guide to [buying evaluation datasets](https://sellmydatatoai.net/guides/evaluation-datasets-for-ai) covers contamination in more depth. ## What acceptance criteria should a pilot use? Acceptance criteria should be written before the pilot starts and should cover both the data and the model effect. The thresholds below are examples to adapt, not market standards. | Criterion | How to measure it | Example threshold to adapt | | --- | --- | --- | | Spec conformance | Audit sample checked against the manifest fields | Every record matches the agreed schema and date range | | Defect rate | Errors found in the random audit sample | Upper confidence limit below the rate you can tolerate | | Residual personal data | Automated and manual scan of the audit sample | No direct identifiers found in a sample sized by the rule of three | | Duplication | Near-duplicate detection within the slice and against public sets | Below an agreed share of records | | Model effect | Held-out evaluation score against the baseline and control | A gain larger than the spread across seeds | | Documentation | Datasheet, manifest and hashes for the slice | Complete and tied to an exact version | Attach the same criteria to the full license as its acceptance schedule. A pilot that passes on criteria you never wrote into the contract gives you no remedy if the full delivery differs from the slice. ## How should payment be staged around a pilot? Payment around a pilot is usually staged in three parts: a pilot fee, a license fee on conversion and a holdback released when the full delivery passes acceptance. Nixon Peabody lists holdbacks among the tools for allocating risk, alongside indemnities, liability caps, insurance and termination rights. | Stage | Trigger | What the seller receives | What the buyer receives | | --- | --- | --- | --- | | Pilot fee | Pilot license signed | Payment for preparing the slice | Testing rights on the slice | | Conversion | Pilot passes and full license signed | License fee, often with the pilot fee credited | Training rights and the first tranche | | Tranche payments | Each tranche delivered | Payment per tranche | Further data, checked against the manifest | | Holdback release | Full delivery passes acceptance | The retained balance | A conforming dataset and a deletion process for rejects | Whether the pilot fee is credited, and how large the holdback is, are commercial points to negotiate. Our guide to [AI training data pricing models](https://sellmydatatoai.net/guides/ai-training-data-pricing-models) explains how flat fees, unit prices and recurring access interact with staged payment. ## What rights does a pilot license need? A pilot license needs testing rights only, with clear rules for what happens to anything trained on the slice if you do not convert. Sellers will ask for this, and it protects you from a dispute later. - Use limited to the agreed experiment on named models - Deletion of the slice and of pilot-trained weights if you do not convert, with a written certificate - No onward transfer outside the approved environment - Confidentiality of results, or aggregate results only if the seller prefers - A conversion window during which the pilot price and terms hold In the US, the license rather than fair use is what governs your pilot. Our [AI training data lawsuits tracker](https://sellmydatatoai.net/ai-training-data-lawsuits) shows why buyers prefer contract certainty while the case law develops. ## What are the common mistakes in pilot design? - Letting the seller choose the pilot slice instead of drawing it at random - Measuring against no baseline or control, so extra training gets credited to the data - Writing acceptance criteria after seeing the results - Using an evaluation set the seller has seen or that overlaps public benchmarks - Signing a full license whose acceptance schedule does not match the pilot criteria ## How does Sell My Data To AI support pilots? Sell My Data To AI shortlists vetted sellers and helps both sides prepare for a pilot. Every seller passes our eight-point vetting standard, which includes a representative sample drawn at random across the archive and a check that the manifest matches what is delivered. Buyers see redacted samples under a mutual NDA, and you sign any pilot or full license directly with the seller. To scope a pilot, [book a buyer consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer) or [submit a data requirement](https://sellmydatatoai.net/request-data). Our [procurement guide](https://sellmydatatoai.net/guides/ai-training-data-procurement) shows where the pilot fits in the wider process. ## Frequently asked questions ### How many records should I audit in a data pilot? Audit enough random records to estimate defects at the precision you need. At 95% confidence, 400 records gives a margin of about 4.9 percentage points in the worst case and 1,000 records about 3.1 points. For rare problems such as residual personal data, the rule of three applies: no defects in 300 records puts the upper limit near 1%. ### Should I pay for a training data pilot? Paying for a pilot is normal when the seller has to prepare, de-identify or convert a slice of data. A paid pilot also gives you a signed pilot license with testing rights and deletion terms, which an informal trial does not. Many buyers negotiate for the pilot fee to be credited against the full license fee on conversion. ### What happens to a model trained during a pilot if I do not buy? What happens to a pilot-trained model depends on the pilot license, so agree it before training. The usual approach is deletion of the slice and of any weights trained on it, confirmed by a written certificate. Nixon Peabody notes that deleting source files does not remove their influence from a trained model, which is why weights need covering too. ### How do I know a pilot gain is real? A pilot gain is real when it beats both the baseline and a control run trained on the same compute with data you already hold, and when it exceeds the spread across random seeds. Check the pilot slice for overlap with your evaluation set, including paraphrases, because Yang and colleagues showed n-gram matching misses rephrased test items. ## Sources - [Shopping for AI training data? Five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody LLP, accessed November 2026 - [If almost nothing goes wrong, is almost everything all right? Interpreting small numerators](https://jamanetwork.com/journals/jama/article-abstract/389714), JAMA (Newman, 1995), accessed November 2026 - [LIMA: Less is more for alignment](https://arxiv.org/abs/2305.11206), arXiv (Zhou et al., 2023), accessed November 2026 - [Llama 2: Open foundation and fine-tuned chat models](https://arxiv.org/abs/2307.09288), arXiv (Touvron et al., Meta, 2023), accessed November 2026 - [Phi-4 technical report](https://arxiv.org/abs/2412.08905), arXiv (Abdin et al., Microsoft, 2024), accessed November 2026 - [Rethinking benchmark and contamination for language models with rephrased samples](https://arxiv.org/abs/2311.04850), arXiv (Yang et al., 2023), accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Sourcing expert and human-generated data: vendors, costs and quality URL: https://sellmydatatoai.net/guides/sourcing-expert-and-human-data Reviewed: 2026-11-08 Expert and human-generated data is commissioned work, such as demonstrations, preference comparisons and red-team attacks, bought from vendors including Scale AI, Surge AI, Mercor, Turing, micro1, Invisible and Handshake AI. None of them publishes a buyer rate card, so price is set per hour, task or project. Quality depends on rater screening, calibration and overlap checks you can audit. ## Key takeaways - Human data covers demonstrations for fine-tuning, preference comparisons for RLHF, red teaming and evaluation tasks. - None of the seven vendors in this guide publishes buyer pricing, but contributor pay gives a floor: Mercor shows a $90 an hour average contracted rate. - OpenAI's InstructGPT paper reported agreement of 72.6% between its training labelers, which is a useful reference point for preference data. - Ask every vendor for screening tests, gold tasks, overlap rates and versioned guidelines, and audit a random sample yourself. - Your contract needs IP assignments from every contributor and rules on reuse of your prompts with other clients. Licensed archives tell a model what people have already written. Expert and human-generated data is made to order: people write the ideal answer, compare two model outputs or try to make a model fail. Labs commission it because it targets the exact behavior they want to change. This guide covers the main types, the vendors that sell them, how pricing works and how to control quality. Every vendor detail below comes from the vendor's own site, checked in November 2026. ## What types of expert and human data do AI teams buy? AI teams buy four main types of human data: demonstrations, preference comparisons, red-team attacks and evaluation tasks. The InstructGPT paper by Long Ouyang and colleagues at OpenAI set the template in 2022, using about 40 contractors hired on Upwork and through Scale AI to write demonstrations and rank model outputs. | Type | What people do | Used for | Usual pricing basis | | --- | --- | --- | --- | | Demonstrations | Write the ideal response to a prompt | Supervised fine-tuning | Per task or per hour | | Preference data | Compare or rank model outputs | Reward models and RLHF | Per comparison or per hour | | Red teaming | Try to make a model produce harmful output | Safety tuning and testing | Per hour or per project | | Evaluation tasks | Write tasks with reference answers and rubrics | Measuring capability | Per task or per set | | Agent trajectories and environments | Complete or design multi-step tasks | Agent training and grading | Per environment or per project | Our pages on [RLHF and preference data](https://sellmydatatoai.net/buy-ai-training-data/rlhf-preference-data) and [agent trajectories and RL environments](https://sellmydatatoai.net/guides/agent-trajectories-and-rl-environments) cover two of these in depth. ## Which vendors supply expert human data? This guide covers seven vendors that sell expert human data to AI labs: Scale AI, Surge AI, Mercor, Turing, micro1, Invisible and Handshake AI. The table records what each states on its own site, which is marketing copy, not verified performance. | Vendor | What it says it offers | Published figures | Pricing published | | --- | --- | --- | --- | | Scale AI | Generative AI Data Engine with datasets curated by vetted subject matter experts, targeted red teaming and an Ops Center for quality control | None on the product page | No | | Surge AI | Off-the-shelf data, a training lab and an expert workforce, plus its own benchmarks such as DAYJOB: Tax | None on the homepage | No | | Mercor | Human evaluation, expert data, benchmarks and RL environments built on professional work | 5M+ experts, $90 an hour average contracted rate | No | | Turing | Datasets, RL environments and expert talent for frontier labs in software engineering, enterprise knowledge work and frontier STEM | None | No | | micro1 | Expert data and evaluation environments, with Zara, an AI recruiter agent that sources and vets experts | 100+ fields | No | | Invisible | Training data and RL environments, with Meridial, a network of vetted experts | Says it has trained over 80% of the world's leading AI models | No | | Handshake AI | Expert data on demand and RL environments for frontier labs and enterprises | 30M+ experts in its network, 200+ in-house researchers and engineers | No | Several of these companies also buy existing company data through separate programs. Our profiles of [Scale AI](https://sellmydatatoai.net/data-buyers/scale-ai), [Mercor](https://sellmydatatoai.net/data-buyers/mercor), [micro1](https://sellmydatatoai.net/data-buyers/micro1) and [Handshake AI](https://sellmydatatoai.net/data-buyers/handshake-ai) cover those programs from the seller's side. ## How much does expert human data cost? Expert human data is priced per hour, per task or per project, and none of the seven vendors above publishes a buyer rate card. The pay rates vendors publish to contributors give a floor, because the buyer price must also cover recruiting, review, tooling and margin. | Vendor | Published contributor pay | Context | | --- | --- | --- | | Mercor | $90 an hour average contracted rate | Homepage figure across its expert network | | Mercor | $60 to $70 an hour | Sales and marketing expert role listing | | Mercor | $65 to $75 | Chemicals safety for red teaming role listing | | Handshake AI | Up to $17 an hour | AI evaluation specialist in the Handshake AI Fellowship | | Handshake AI | Up to $150 an hour | Software engineer in the Handshake AI Fellowship | Volume is large. Mercor's homepage reports more than $4 million in daily payouts, and Handshake's fellowship page reports more than $500 million in payouts and more than 188,000 fellows, who must be based in the US. Scale AI's Generative AI Data Engine page quotes investor Nat Friedman on why labs pay these rates. > "Experts are the new GPUs." (Nat Friedman, quoted on Scale AI's Generative AI Data Engine page) | Model | How it works | Works best for | Watch for | | --- | --- | --- | --- | | Per hour | You pay for contributor time | Exploratory work and red teaming | Slow throughput with no cap | | Per task | You pay per accepted demonstration, comparison or eval item | Stable, well-specified tasks | Vendors rushing tasks to hit volume | | Per project | Fixed fee for a defined deliverable | Benchmarks and RL environments | Scope changes priced as extras | | Managed team | A dedicated pool billed monthly | Long programs with changing guidelines | Paying for idle capacity | Pay only for accepted tasks where you can, and define acceptance before work starts. Our guide to [AI training data pricing models](https://sellmydatatoai.net/guides/ai-training-data-pricing-models) compares these with licensing structures. ## How do you control quality in human-generated data? Quality in human-generated data comes from screening raters, calibrating them against a reference and measuring agreement throughout the project. Ask vendors to show each step, then audit a random sample yourself. The InstructGPT paper is a useful benchmark. OpenAI selected contractors using a screening test and their agreement with researchers on a labeling task, onboarded them with detailed instructions and kept a shared chat room for questions. Training labelers agreed with each other 72.6% of the time, plus or minus 1.5%, and held-out labelers 77.3%, plus or minus 1.3%. The same paper admits a cost shortcut: most comparisons were labeled by only one contractor. Vendors are often less specific. Surge AI's case study on its RLHF work for Anthropic credits proprietary quality control technology but does not explain how raters are vetted or calibrated, so ask. Meta's Llama 2 team found that different annotation platforms and vendors produced markedly different downstream model performance, and checked quality by examining 180 examples by hand. - Screening: a test of task skill and agreement with your reference answers before anyone works on live tasks - Gold tasks: items with known answers mixed into the queue to score each rater continuously - Overlap: a share of tasks labeled by two or more raters so agreement can be measured - Calibration sessions: regular reviews where raters and your researchers discuss disagreements - Versioned guidelines: a dated guideline document, because a change in guidance changes what labels mean - Credential checks: proof of the qualifications claimed for expert raters - Your own audit: a random sample reviewed by your team before each payment ## How should you source red teaming data? Source red teaming data from raters trained on a written attack taxonomy, with content warnings and wellbeing support, because the work exposes people to harmful material. Ask for the instructions raters followed along with the attacks themselves. Anthropic's 2022 red teaming paper by Deep Ganguli and colleagues released 38,961 red team attacks and found that RLHF models became harder to red team as they scaled. It also documented the instructions, processes and statistical methods used, which is the level of detail to ask a vendor for. Mercor's homepage lists a chemicals safety role for red teaming at $65 to $75, which shows red teaming being staffed by domain experts. ## What should an expert data contract include? An expert data contract should give you clear ownership of the outputs, protect your prompts and guidelines, and set quality and acceptance terms. Most contributors are contractors, so ownership depends on the vendor's agreements with them. - Written IP assignments from every contributor to the vendor, passed on to you - No reuse of your prompts, guidelines or outputs for other clients - Exclusivity of the delivered data where you need it, especially for evaluation sets - Rules on which models generated any outputs being rated, and that their terms allow the use - Acceptance tests, rework at the vendor's cost and the right to audit quality records - Contributor consent, pay and location records for regulatory and reputational checks The model terms point matters. OpenAI's terms of use, for example, bar using output from its services to develop models that compete with OpenAI. Take independent legal advice on your own contract. ## Should you build an in-house rater pool or use a vendor? Use a vendor when you need volume or rare expertise quickly, and build in-house when the guidelines are a competitive secret or change daily. Many labs do both. | Factor | In-house pool | Vendor | | --- | --- | --- | | Speed to start | Slow: recruiting and onboarding | Fast for common domains | | Control of guidelines | Full | Shared with the vendor | | Rare expertise | Hard to recruit | Easier through large networks | | Confidentiality | Highest | Depends on contract and vendor controls | | Cost visibility | Direct pay plus management time | Bundled into the vendor's price | ## When is licensed company data a better fit? Licensed company data is a better fit when the expertise you need already exists in completed work, such as closed legal matters, audit files or resolved support tickets. Each record pairs a real problem with a verified outcome, which commissioned raters would have to recreate. Sell My Data To AI shortlists vetted sellers of that kind of data. Every seller passes our eight-point vetting standard before any buyer sees their data, and buyers pay nothing to submit a requirement or review a shortlist. [Submit a data requirement](https://sellmydatatoai.net/request-data) or [book a buyer consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer). ## Frequently asked questions ### How much does RLHF data cost? RLHF data is priced per comparison, per hour or per project, and none of Scale AI, Surge AI, Mercor, Turing, micro1, Invisible or Handshake AI publishes a buyer rate card. Published contributor pay sets a floor: Mercor shows a $90 an hour average contracted rate, and Handshake AI lists fellowship roles from up to $17 to up to $150 an hour. ### What is a good inter-annotator agreement rate for preference data? There is no single standard, but published work gives reference points. OpenAI's InstructGPT paper reported that training labelers agreed with each other 72.6% of the time and held-out labelers 77.3%, and cited researcher-to-researcher agreement of 73% in earlier summarization work. Contested prompts lower agreement, so ask vendors to flag them. ### Who owns data created by contract raters? Ownership of rater output depends on the contracts between the raters, the vendor and you. Ask the vendor to confirm it holds written IP assignments from every contributor and to assign the delivered data to you, or license it on terms you accept. Without that chain, a contributor could later dispute rights in the work. Take independent legal advice. ### Can I use outputs from another company's model in my human data? Using another company's model outputs in human data depends on that model's terms. OpenAI's terms of use, for example, bar using output from its services to develop models that compete with OpenAI. Ask your vendor which models generated any outputs raters compare or edit, and check those terms before you train on the result. ### Should I use one human data vendor or several? Several vendors reduce dependence on one pool and let you compare quality on the same tasks. Meta's Llama 2 team found that different annotation platforms and vendors produced markedly different downstream model performance, which argues for running a small paid test with two or three vendors before you commit volume to one. ## Sources - [Generative AI Data Engine](https://scale.com/generative-ai-data-engine), Scale AI, accessed November 2026 - [Surge AI: Human intelligence for AGI](https://www.surgehq.ai/), Surge AI, accessed November 2026 - [How Anthropic uses Surge AI's RLHF platform to train its LLM assistant](https://www.surgehq.ai/blog/anthropic-surge-ai-rlhf-platform-train-llm-assistant-human-feedback), Surge AI, accessed November 2026 - [Mercor: Organizing human intelligence to power the AI economy](https://www.mercor.com/), Mercor, accessed November 2026 - [Services](https://www.turing.com/services), Turing, accessed November 2026 - [Frontier evaluations and RL environments for AI labs](https://www.micro1.ai/human-data), micro1, accessed November 2026 - [Applied AI research for enterprise](https://invisibletech.ai/), Invisible Technologies, accessed November 2026 - [Handshake Enterprise AI](https://joinhandshake.com/ai/enterprise/), Handshake, accessed November 2026 - [Handshake AI program](https://joinhandshake.com/ai/), Handshake, accessed November 2026 - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155), arXiv (Ouyang et al., OpenAI, 2022), accessed November 2026 - [Llama 2: Open foundation and fine-tuned chat models](https://arxiv.org/abs/2307.09288), arXiv (Touvron et al., Meta, 2023), accessed November 2026 - [Red teaming language models to reduce harms: methods, scaling behaviors, and lessons learned](https://arxiv.org/abs/2209.07858), arXiv (Ganguli et al., Anthropic, 2022), accessed November 2026 - [Terms of use](https://openai.com/policies/terms-of-use/), OpenAI, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Buying evaluation datasets: what makes a good AI eval set URL: https://sellmydatatoai.net/guides/evaluation-datasets-for-ai Reviewed: 2026-11-08 A good AI evaluation set is private, expert-checked and uncontaminated: its tasks have never appeared in any model's training data, each answer has been verified by a second expert and the set is refreshed or retired before it leaks. Buy it with exclusivity, a ban on reuse by the seller and a remedy if tasks surface in public, because a leaked eval set measures memory, not ability. ## Key takeaways - Contamination is the main risk: Zhang and colleagues found accuracy drops of up to 8% when leading models met fresh problems mirroring the public GSM8k benchmark. - Exact-match checks miss paraphrases, and Yang and colleagues found 8% to 18% of HumanEval overlapping two public pre-training sets. - Refresh or retire sets on a schedule, as LiveBench does with monthly question updates. - Public benchmarks such as Humanity's Last Exam keep a private held-out split to detect overfitting, which is the model to copy. - License eval sets exclusively, for evaluation only, with leak notification and replacement terms. An evaluation set is the one dataset you buy hoping no model ever learns from it. Its value is the gap between what a model has memorized and what it can actually do, and that gap closes the moment the tasks leak. This guide is for teams buying private evals for model development, vendor selection or regression testing. Our [evaluation datasets](https://sellmydatatoai.net/buy-ai-training-data/evaluation-datasets) page covers formats and delivery, so this guide focuses on contamination, exclusivity, refresh cycles and private benchmarks. ## What makes a good AI evaluation set? A good AI evaluation set has tasks that match your deployment, answers checked by experts, scoring rules you can reproduce and a documented history showing nobody else has trained on it. Difficulty matters too: a set every model passes tells you nothing. Public benchmarks show the effort involved. OpenAI's SWE-bench Verified, published in August 2024, used 93 software developers experienced in Python to screen tasks and kept 500 samples verified as non-problematic. Humanity's Last Exam accepted only questions that made frontier models fail, then refined each one with between one and three expert peer reviews. | Element | What good looks like | | --- | --- | | Task | Self-contained, with everything needed to answer it | | Reference answer | Verified by a second expert, or a rubric where answers vary | | Scoring rule | Automatic check or written rubric that two graders apply consistently | | Metadata | Stable ID, author, creation date, difficulty and domain tags | | Provenance | Who wrote it, from what source and who has seen it since | ## Why does contamination ruin an evaluation set? Contamination ruins an evaluation set because a model that has seen the tasks during training scores well by recall. The score then measures exposure, and you can no longer compare models or track progress. The effect is measurable. Hugh Zhang and colleagues built GSM1k in 2024 to mirror the public GSM8k math benchmark and observed accuracy drops of up to 8% for leading models, with several model families showing evidence of systematic overfitting. Contamination also hides well. Shuo Yang and colleagues showed in 2023 that paraphrased or translated test items pass n-gram decontamination, that a 13 billion parameter model trained on such variants could reach scores on par with GPT-4, and that synthetic datasets generated by GPT-3.5 and GPT-4 also contained benchmark overlap. ## How do you check a purchased eval set for contamination? Check a purchased eval set for contamination by tracing each task's origin, then searching for it and its paraphrases in public benchmarks, the open web and your own training corpora. Do this on a sample before you sign and on the full set at delivery. 1. **Trace provenance**: Ask for each task's author, creation date and source, and whether the seller has shown it to any other buyer. 2. **Search public benchmarks**: Compare tasks against well-known public sets in the same domain. 3. **Search the web**: Run sampled tasks through web search to catch tasks copied from forums, textbooks or exams. 4. **Check paraphrases**: Use a model-based similarity check, since Yang and colleagues showed exact matching misses rephrased items. 5. **Check your own corpora**: Search your training data, including synthetic data, for overlap before the first evaluation run. 6. **Use time cutoffs**: Prefer tasks created after your models' training cutoff, the approach LiveCodeBench takes by collecting new contest problems over time. ## Should you buy an evaluation set exclusively? You should usually buy an evaluation set exclusively, at least for a fixed term, because every additional licensee is another route for tasks to reach training data. Exclusivity costs more, so match its length to how long you expect to use the set. - Exclusive use for a stated term, with no sale of the same tasks or close variants to anyone else - No reuse by the seller of tasks, answers or rubrics in its own training or other products - No publication of the tasks, including in papers or marketing - Leak notification within a set period if any task appears in public - Replacement of leaked tasks at the seller's cost, or a partial refund Our guide to [exclusive vs non-exclusive data licenses](https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses) explains term, field and territory limits that can lower the price of exclusivity. ## How often should an evaluation set be refreshed? Refresh an evaluation set on a schedule set by how quickly your domain changes and how widely the set is used. Published benchmarks give three models: rolling updates, a private held-out split and versioned retirement. | Model | Example | How it works | Fits when | | --- | --- | --- | --- | | Rolling updates | LiveBench | Questions added and updated monthly from recent sources such as math competitions, arXiv papers and news | The domain changes quickly | | Time-windowed collection | LiveCodeBench | New problems collected over time from LeetCode, AtCoder and CodeForces, so models can be scored only on problems after their cutoff | Tasks carry a natural release date | | Private held-out split | Humanity's Last Exam | Public questions released while a private held-out set is kept to assess overfitting | You need a public score and a check on gaming | | Unpublished set | FrontierMath | New, unpublished problems with automated verification | Tasks are expensive to write and must stay secret | For a purchased set, write the refresh into the contract: how many new tasks per period, how retired tasks are handled and whether retired tasks may then be used for training. ## What are private benchmarks and who runs them? Private benchmarks are evaluation sets whose tasks are withheld from the public so that model builders cannot train on them. Several US data vendors now run their own and sell evaluation work alongside training data. | Publisher | Benchmark | What it publishes about the set | | --- | --- | --- | | Scale AI | SEAL leaderboards | Curated evaluation sets combining private datasets to prevent overfitting with open-source datasets | | Mercor | Harvey LAB, with Harvey | Attorney-built benchmark of 1,749 real legal tasks across 24 practice areas, graded on finished work products | | Mercor | APEX-Accounting | Long-horizon accounting tasks across accounting software, spreadsheets and PDFs | | Surge AI | DAYJOB: Tax and others | Lists benchmarks including DAYJOB: Tax, a benchmark for AI tax agents | | Epoch AI | FrontierMath | Several hundred unpublished, expert-written mathematics problems | Ask who funded a benchmark and who can see its tasks. The FrontierMath paper, for example, acknowledges OpenAI for its support in creating the benchmark. Funding does not imply access, but a buyer comparing vendors' models should know the relationships. ## What does a private evaluation set cost? Private evaluation sets are usually priced per task or per set, with exclusivity priced higher. We found no published per-task buyer prices from the vendors named in this guide in November 2026, so expect quotes rather than list prices. The cost drivers are the expert time to write and verify each task, the rubric work for open-ended answers and the length of exclusivity. Our guide to [sourcing expert and human-generated data](https://sellmydatatoai.net/guides/sourcing-expert-and-human-data) lists the contributor pay rates vendors publish, which set a floor on task cost. ## What should an evaluation set license include? An evaluation set license should limit use to evaluation, define exclusivity and set out leak remedies and refresh deliveries. It should also stop your own teams from training on the set by mistake. - Use limited to evaluation, with training on the tasks prohibited for both parties - Access controls: named teams, logged access and storage away from training pipelines - Ownership of scoring code and any rubric changes you make - Refresh deliveries and retirement rules - Leak notification, replacement and refund terms - Provenance records for each task, kept for audit Company records can also become evaluation tasks. A resolved support ticket or a closed legal matter pairs a problem with a verified outcome. If you license records for that purpose, check that the seller has not licensed the same records to anyone for training, a point our [due diligence checklist](https://sellmydatatoai.net/guides/ai-training-data-due-diligence-checklist) covers. ## How can Sell My Data To AI help you source eval sets? Sell My Data To AI shortlists vetted sellers whose records can be turned into private evaluation tasks, such as professional firms and software companies. Every seller passes our eight-point vetting standard, and you sign the license, including exclusivity, directly with the seller. To start, [submit a data requirement](https://sellmydatatoai.net/request-data) or [book a buyer consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer). If you plan to use the eval set to judge a pilot, read our guide to [training data pilot design](https://sellmydatatoai.net/guides/training-data-pilot-design). ## Frequently asked questions ### What is evaluation data contamination? Evaluation data contamination is when tasks from a test set, or close paraphrases of them, appear in a model's training data. The model then scores well by recall rather than ability. Zhang and colleagues' 2024 GSM1k study found accuracy drops of up to 8% when leading models met fresh problems built to mirror the public GSM8k benchmark. ### Is n-gram matching enough to decontaminate an eval set? N-gram matching is not enough on its own. Yang and colleagues showed in 2023 that paraphrased or translated test items pass n-gram checks, and they found 8% to 18% of the HumanEval benchmark overlapping the RedPajama-Data-1T and StarCoder-Data pre-training sets. Add a model-based similarity check and trace each task's origin. ### How long does a private eval set stay useful? A private eval set stays useful until its tasks leak or models saturate it. Use and sharing speed up both. Published benchmarks manage this with refresh cycles, such as LiveBench's monthly question updates, or with a private held-out split, as Humanity's Last Exam keeps. Write a refresh schedule into any license you sign. ### Can I use company records as evaluation data? Company records can make strong evaluation data when each one pairs a problem with a verified outcome, such as a resolved support ticket or a closed legal matter. Each task must be rewritten to stand alone and given a scoring rule. Check that the records have not been licensed to anyone else for training, or the set may already be contaminated. ### Should evaluation data be exclusive? Evaluation data should usually be exclusive, at least for a fixed term, because each extra licensee raises the chance that tasks reach someone's training data. Agree that the seller will not reuse or publish the tasks, will notify you of any leak and will replace leaked tasks at its own cost. ## Sources - [A careful examination of large language model performance on grade school arithmetic](https://arxiv.org/abs/2405.00332), arXiv (Zhang et al., 2024), accessed November 2026 - [Rethinking benchmark and contamination for language models with rephrased samples](https://arxiv.org/abs/2311.04850), arXiv (Yang et al., 2023), accessed November 2026 - [LiveBench: A challenging, contamination-limited LLM benchmark](https://arxiv.org/abs/2406.19314), arXiv (White et al., 2024), accessed November 2026 - [LiveCodeBench: Holistic and contamination free evaluation of large language models for code](https://arxiv.org/abs/2403.07974), arXiv (Jain et al., 2024), accessed November 2026 - [Humanity's Last Exam](https://arxiv.org/abs/2501.14249), arXiv (Phan et al., 2025), accessed November 2026 - [FrontierMath: A benchmark for evaluating advanced mathematical reasoning in AI](https://arxiv.org/abs/2411.04872), arXiv (Glazer et al., Epoch AI, 2024), accessed November 2026 - [Introducing SWE-bench Verified](https://openai.com/index/introducing-swe-bench-verified/), OpenAI, accessed November 2026 - [SEAL leaderboards](https://labs.scale.com/leaderboard), Scale AI, accessed November 2026 - [Mercor: Organizing human intelligence to power the AI economy](https://www.mercor.com/), Mercor, accessed November 2026 - [Surge AI: Human intelligence for AGI](https://www.surgehq.ai/), Surge AI, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Synthetic data vs licensed real data for AI training URL: https://sellmydatatoai.net/guides/synthetic-vs-licensed-training-data Reviewed: 2026-11-08 Synthetic data works best as a supplement to real data, not a replacement. Shumailov and colleagues reported in Nature in 2024 that training indiscriminately on model-generated data makes the tails of the original distribution disappear, while Gerstgrasser and colleagues found that accumulating synthetic data alongside the original real data avoided collapse. Licensed real data anchors the mix and covers domains no generator knows. ## Key takeaways - Replacing real data with each generation's synthetic output leads toward model collapse, but keeping and adding to the real data avoided it in Gerstgrasser and colleagues' experiments. - Microsoft's Phi-4 report allocates 40% of pre-training tokens to synthetic data, and Nvidia reports over 98% of Nemotron-4 340B's alignment data was synthetic. - Theory papers disagree on safe ratios: Dohmatob and colleagues show even 1% synthetic data can cause collapse in a regression setting. - Synthetic data carries its own rights questions, including the generator model's terms of use. - Buy licensed real data for rare cases, new domains, human behavior and evaluation. Synthetic data is cheap to produce, has no rights holder to negotiate with and can be generated to order. That makes it tempting to treat as a substitute for licensed data. The research says the two do different jobs. This guide summarizes what peer-reviewed and preprint research reports on model collapse and mixing, what model builders say about their own data mixtures and how to decide which to buy. ## What is the difference between synthetic and licensed real data? Synthetic data is generated by a model or a simulation, while licensed real data was created by people or systems doing real work and is used under a contract with its owner. Synthetic data inherits the strengths and blind spots of whatever generated it. | Factor | Synthetic data | Licensed real data | | --- | --- | --- | | Source | A generator model, simulator or rules | People and systems doing real work | | Cost driver | Compute and verification | License fee and preparation | | Coverage of rare cases | Limited to what the generator can produce | Includes the real tail of cases | | Rights | Generator terms, seed data rights and privacy of any source records | The license and the seller's chain of title | | Main risk | Collapse, errors and contamination | Rights gaps and personal data | ## What is model collapse? Model collapse is the degradation that occurs when models are trained on data generated by earlier models. Ilia Shumailov and colleagues, writing in Nature in July 2024, found that indiscriminate use of model-generated content in training causes irreversible defects in which tails of the original content distribution disappear. The paper distinguishes early collapse, where the model loses information about rare events, from late collapse, where it converges to a distribution with little resemblance to the original. It showed the effect in language models, variational autoencoders and Gaussian mixture models. > "The value of data collected about genuine human interactions with systems will be increasingly valuable in the presence of LLM-generated content in data crawled from the Internet." (Shumailov et al., Nature, July 2024) ## Does mixing real and synthetic data prevent model collapse? Mixing real and synthetic data reduces or prevents collapse in several studies, but the safe proportion depends on the setting and the papers disagree. The key variable is whether real data is kept and added to, or replaced. | Paper | Setting | Reported finding | | --- | --- | --- | | Shumailov et al., Nature, 2024 | Fine-tuning a language model over generations | Keeping a random 10% of the original data each generation led to only minor degradation | | Gerstgrasser et al., 2024 | Pre-training sequences of language models | Replacing real data with synthetic data tends toward collapse, while accumulating synthetic data alongside the original real data avoids it | | Seddik et al., 2024 | Statistical model of language model collapse | Collapse cannot be avoided when training solely on synthetic data, and there is a maximal amount of synthetic data below which it can be avoided | | Alemohammad et al., 2023 | Image generation in self-consuming loops | Without enough fresh real data in each generation, quality or diversity progressively decreases | | Dohmatob et al., 2024 (Strong model collapse) | Supervised regression under scaling laws | Even 1% synthetic data can lead to collapse, where larger training sets stop improving performance | | Dohmatob et al., ICML 2024 (A tale of tails) | Scaling laws with synthesized data | Mixing human and synthesized data changes scaling, including the un-learning of skills | Read the settings column before applying a ratio. Dohmatob and colleagues' 1% result comes from a regression setting with theory and experiments, while Gerstgrasser's result comes from training sequences of language models. Neither tells you the right mix for your model, which is why you should test it. ## What mixing ratios do model builders report? Model builders who publish their data mixtures report heavy use of synthetic data for reasoning and alignment, alongside web, code and acquired sources. These are each company's reports about its own models. | Model | Reported by | What the report says | | --- | --- | --- | | Phi-4 (14 billion parameters) | Microsoft, December 2024 | Pre-training mix of 40% synthetic, 15% web, 15% web rewrites, 20% code and 10% acquired sources | | Nemotron-4 340B | Nvidia, June 2024 | Over 98% of data used in model alignment was synthetically generated | | phi-1 (1.3 billion parameters) | Microsoft, June 2023 | 6 billion tokens of textbook-quality web data and 1 billion tokens of synthetic textbooks and exercises from GPT-3.5 | Note the 10% of acquired sources in Phi-4's mix. Even a synthetic-first recipe kept a share of licensed and organic data, which the report describes as targeted acquisitions and organic data such as academic data, books and forums. ## When does synthetic data fit? Synthetic data fits when you have a strong generator and a way to verify outputs. Code with unit tests, math with checkable answers and format-following tasks are the clearest cases. - Tasks with automatic verifiers, such as code that must pass tests - Rewrites and reformatting of data you already hold the rights to - Alignment and instruction data, as Nvidia reports for Nemotron-4 340B - Rare scenarios that are dangerous or expensive to collect, such as simulated edge cases - Privacy-sensitive domains, where synthetic records replace direct use of personal data, subject to a re-identification check ## When does licensed real data fit? Licensed real data fits wherever the tails of the distribution matter, the domain is new to generators or the task depends on how real people and businesses behave. Shumailov and colleagues note that access to real human-produced data is needed in tasks where the tails of the underlying distribution matter. - Professional work such as legal, accounting and engineering files, where generators lack the domain - Real customer conversations and support tickets, which carry the errors and edge cases users produce - Recent records created after a generator's training cutoff - Evaluation sets, where synthetic tasks risk sharing the generator's blind spots - An anchor set of real data to keep in every training mix Our [training data catalog](https://sellmydatatoai.net/buy-ai-training-data) lists licensed data by type, including [expert domain data](https://sellmydatatoai.net/buy-ai-training-data/expert-domain-data) and [support ticket data](https://sellmydatatoai.net/buy-ai-training-data/support-ticket-data). ## What rights issues does synthetic data carry? Synthetic data is not rights-free. Its rights depend on the generator's terms, the rights in any seed data and whether personal data shaped the output. - Generator terms: OpenAI's terms of use, for example, bar using output from its services to develop models that compete with OpenAI. - Seed data: synthetic records derived from licensed or scraped seed data inherit questions about those seeds. - Personal data: synthetic records generated from personal data should be tested for re-identification before you treat them as anonymous. - Contamination: Yang and colleagues found benchmark overlap in synthetic datasets generated by GPT-3.5 and GPT-4. For licensed real data, US copyright questions sit in the license rather than in fair use arguments. Our [AI training data lawsuits tracker](https://sellmydatatoai.net/ai-training-data-lawsuits) follows the cases that shape unlicensed use. Take independent legal advice on your own sources. ## How should you decide between synthetic and licensed data? Decide by the task, not by price. If a verifier can check the output and the generator knows the domain, start with synthetic data. If neither is true, license real data, and keep a real anchor set in any mix. | Your situation | Better starting point | | --- | --- | | Outputs can be checked automatically | Synthetic, verified | | The domain is specialist or proprietary | Licensed real data | | You need rare and realistic edge cases | Licensed real data, with synthetic augmentation | | You need more of a format you already have | Synthetic rewrites of data you hold rights to | | You are building an evaluation set | Licensed real data or expert-written tasks | | You train repeatedly on your own outputs | Keep and add to the real data each round | Test the mix on a held-out evaluation before committing, as our guide to [training data pilot design](https://sellmydatatoai.net/guides/training-data-pilot-design) explains. ## How do you buy real data to anchor a synthetic pipeline? Sell My Data To AI shortlists vetted sellers of rights-cleared real data. Every seller passes our eight-point vetting standard before any buyer sees their data, and buyers pay nothing to submit a requirement or review a shortlist. [Submit a data requirement](https://sellmydatatoai.net/request-data) or [book a buyer consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer). ## Frequently asked questions ### Can you train an AI model only on synthetic data? Training only on synthetic data carries a documented risk of model collapse. Seddik and colleagues' 2024 analysis concludes collapse cannot be avoided when training solely on synthetic data, and Shumailov and colleagues reported in Nature that indiscriminate training on model-generated content makes rare cases disappear. Published recipes such as Microsoft's Phi-4 still keep web, code and acquired data in the mix. ### What is a safe ratio of synthetic to real data? No single safe ratio of synthetic to real data has been established. Gerstgrasser and colleagues found that accumulating synthetic data alongside the original real data avoided collapse, while Dohmatob and colleagues show even 1% synthetic data can cause collapse in a regression setting. Microsoft reports 40% synthetic tokens in Phi-4's pre-training. Test your own mix on a held-out evaluation. ### Is synthetic data free of copyright and privacy issues? Synthetic data is not automatically free of rights issues. The generator's terms may restrict use, as OpenAI's terms of use bar using output to develop competing models. Synthetic records derived from personal data should be tested for re-identification, and seed data carries its own rights. Take independent legal advice on your sources. ### Why is human-created data becoming more valuable? Human-created data is becoming more valuable because the web increasingly contains model-generated text. Shumailov and colleagues wrote in Nature in 2024 that data about genuine human interactions with systems will be increasingly valuable as LLM-generated content spreads in data crawled from the internet. Licensed archives created before and outside that loop carry real-world tails. ## Sources - [AI models collapse when trained on recursively generated data](https://www.nature.com/articles/s41586-024-07566-y), Nature (Shumailov et al., July 2024), accessed November 2026 - [Is model collapse inevitable? Breaking the curse of recursion by accumulating real and synthetic data](https://arxiv.org/abs/2404.01413), arXiv (Gerstgrasser et al., 2024), accessed November 2026 - [How bad is training on synthetic data? A statistical analysis of language model collapse](https://arxiv.org/abs/2404.05090), arXiv (Seddik et al., 2024), accessed November 2026 - [Self-consuming generative models go MAD](https://arxiv.org/abs/2307.01850), arXiv (Alemohammad et al., 2023), accessed November 2026 - [Strong model collapse](https://arxiv.org/abs/2410.04840), arXiv (Dohmatob et al., 2024), accessed November 2026 - [A tale of tails: Model collapse as a change of scaling laws](https://arxiv.org/abs/2402.07043), arXiv (Dohmatob et al., ICML 2024), accessed November 2026 - [Phi-4 technical report](https://arxiv.org/abs/2412.08905), arXiv (Abdin et al., Microsoft, 2024), accessed November 2026 - [Nemotron-4 340B technical report](https://arxiv.org/abs/2406.11704), arXiv (Nvidia, 2024), accessed November 2026 - [Textbooks are all you need](https://arxiv.org/abs/2306.11644), arXiv (Gunasekar et al., Microsoft, 2023), accessed November 2026 - [Terms of use](https://openai.com/policies/terms-of-use/), OpenAI, accessed November 2026 - [Rethinking benchmark and contamination for language models with rephrased samples](https://arxiv.org/abs/2311.04850), arXiv (Yang et al., 2023), accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How big is the AI training data market? URL: https://sellmydatatoai.net/guides/ai-training-data-market-size Reviewed: 2026-11-08 Published estimates put the AI training dataset market at $3.2 billion to $8.74 billion for 2025, depending on the research firm: Grand View Research estimates $3.2 billion and Mordor Intelligence $8.74 billion. Forecasts run from $9.58 billion by 2029 to $49.82 billion by 2031. The figures differ because firms define the market, base year and method differently. ## Key takeaways - 2025 estimates range from $3.2 billion (Grand View Research, June 2026) to $8.74 billion (Mordor Intelligence, May 2026). - Forecasts diverge more: $16.3 billion by 2033 from Grand View Research against $49.82 billion by 2031 from Mordor Intelligence. - Firms segment the market differently, by data type, offering, deployment and buyer industry, and their public summaries do not list the deals they count. - Filings give harder numbers: Reddit's 2024 S-1 reports data licensing arrangements with an aggregate contract value of $203.0 million. - Of the 21 deals in our tracker, 11 have undisclosed values, so any total built from deals understates the market. Market size figures for AI training data appear in pitch decks, press releases and news stories, often without the firm or year that produced them. They are estimates from commercial research firms, and they disagree by a factor of nearly three for the same year. This guide compares the published estimates, explains why they differ and sets them beside harder numbers from US company filings and the reported deal values in our [AI data deals tracker](https://sellmydatatoai.net/ai-data-deals). ## How big is the AI training data market? The AI training data market was worth between $3.2 billion and $8.74 billion in 2025, according to the research firms that publish estimates. Each figure below is the firm's own estimate, taken from its report page or press release in November 2026. | Research firm | Published | Base estimate | Forecast | Growth rate stated | | --- | --- | --- | --- | --- | | Grand View Research | June 2026 | $3.2 billion in 2025, $3.9 billion in 2026 | $16.3 billion by 2033 | 22.6% a year, 2026 to 2033 | | Precedence Research | Updated September 21, 2026 | $3.35 billion in 2025, $3.92 billion in 2026 | $14.94 billion by 2035 | 16.13% a year, 2026 to 2035 | | Global Market Insights | May 2025 | $3.2 billion in 2024 | $16.3 billion by 2034 | 20.5% a year, 2025 to 2034 | | MarketsandMarkets | November 26, 2024 | $2.82 billion in 2024 | $9.58 billion by 2029 | 27.7% a year | | Mordor Intelligence | Updated May 26, 2026 | $8.74 billion in 2025, $11.91 billion in 2026 | $49.82 billion by 2031 | 33.14% a year, 2026 to 2031 | | Meticulous Research | December 1, 2022 | Not stated in the release | $9.35 billion by 2029 | 24.7% a year, 2022 to 2029 | > **How to cite these figures** Always name the firm and the year, and call the figure an estimate. Write "Grand View Research estimated the market at $3.2 billion in 2025", not "the market is worth $3.2 billion". ## Why do AI training data market estimates differ? AI training data market estimates differ because each firm draws the market boundary in a different place, starts from a different base year and uses its own unpublished model. Small differences in definition become large differences in a forecast compounded over six to nine years. | Firm | Segments named | Example share it reports | | --- | --- | --- | | Grand View Research | Type (text, image or video, audio) and vertical | Image and video held 41.9% of revenue in 2025 | | Mordor Intelligence | Data modality, dataset offering (off-the-shelf or custom creation), deployment and end-user industry | Off-the-shelf datasets held 46.84% in 2025 | | Global Market Insights | Data modality and deployment mode | Cloud held 73% in 2024, and the US generated about $1.23 billion | | MarketsandMarkets | Offering, dataset creation, data modality and others | Its release says dataset creation software is expected to lead the dataset creation segment | Mordor Intelligence's definition includes custom dataset creation, which it projects as the fastest-growing offering at 33.74% a year through 2031. A firm that counts the services work of building datasets will report a larger market than one that counts dataset products alone. Method matters too. Grand View Research says its findings rest on proprietary technology databases, executive interviews and regulatory analysis, with internal peer review. None of the public summaries we read lists the company revenues or licensing deals included, so you cannot reconcile one firm's number with another's. ## What do US company filings show about data licensing revenue? US company filings give audited or sworn figures for individual sellers, which are narrower than market estimates but more reliable. Two filings stand out. | Company | Document | Figure as stated | | --- | --- | --- | | Reddit | Form S-1, filed February 2024 | Data licensing arrangements entered in January 2024 with an aggregate contract value of $203.0 million and terms of two to three years, with at least $66.4 million expected as revenue in 2024 | | Shutterstock | Form 10-K for 2023 | Data, Distribution, and Services revenue up 256% to $137.3 million in 2023, with its data offering accounting for $84.9 million of the growth | | Scale AI | Press release, June 12, 2025 | A significant new investment from Meta that values Scale at over $29 billion | Reddit's S-1 also describes how the money is earned: customers receive the right to access data over the contract period, the price is generally a fixed fee or usage-based fee and revenue is recognized over the license period. It warns that substantially all of the contract value came from one partner at the time of filing. Shutterstock's 10-K figure includes Giphy and Shutterstock Studios as well as data licensing, so it is not a pure AI data number. Scale AI's valuation reflects what investors pay for a training-data company, not the size of the market it serves. ## What do reported AI data deals show? Reported AI data deals show a market with a few very large publisher and platform licenses and a long tail of small company sales. The table lists every deal in our tracker with a reported value, exactly as the cited source reported it. | Date | Licensor | Licensee | Value as reported | | --- | --- | --- | --- | | October 2026 | A shut-down crypto startup (founder Keelan Miskell) | Not named (sale brokered by SimpleClosure) | About $25,000 for the codebase, with about $15,000 likely for workspace data (reported) | | August 2026 | Spirit Airlines (in bankruptcy) | Google | $10 million winning auction bid (reported) | | May 2025 | The New York Times | Amazon | $20 million to $25 million a year (reported by The Wall Street Journal) | | November 2024 | HarperCollins | Microsoft (reported by Bloomberg, not confirmed by HarperCollins) | $5,000 per title, split equally between author and publisher (reported) | | November 2024 | Dotdash Meredith | OpenAI | At least $16 million a year (reported) | | August 2024 | Wiley | Undisclosed technology companies | $44 million expected from AI rights deals (reported) | | May 2024 | News Corp | OpenAI | About $250 million over five years in cash and credits (reported by The Wall Street Journal) | | May 2024 | Informa (Taylor and Francis) | Microsoft | Initial data access fee of about $10 million, plus recurring payments over three years (reported) | | February 2024 | Reddit | Google | About $60 million a year (reported by Reuters, citing one source) | | December 2023 | Axel Springer | OpenAI | Tens of millions of euros (reported by Bloomberg) | The largest reported value is News Corp's license to OpenAI, at about $250 million over five years in cash and credits according to The Wall Street Journal. At the other end, MIT Technology Review reported about $25,000 for the codebase of a shut-down crypto startup in October 2026. Many deals never publish a price. Of the 21 deals in our tracker, 11 are listed as undisclosed, including OpenAI's licenses with the Financial Times, Condé Nast and Stack Overflow. Adding up reported values would understate the market, so we do not publish a total. ## Do market estimates include payments to data owners? It is not clear whether published market estimates include licensing payments to data owners such as publishers, platforms and companies. The firms' public summaries do not say which payments they count. Mordor Intelligence's report cites the New York Times licensing agreement with Amazon in May 2025 as an example of rights-cleared content in the custom dataset segment, which suggests it considers such licenses part of the picture. Whether their value is in the $8.74 billion figure is not stated. If your question is what AI companies pay data owners, filings and deal reports answer it better than market sizing does. ## Which market size figure should you use? Use the range rather than one firm's number, and name each source with its year. For a single figure, pick the firm whose definition matches your question, such as Mordor Intelligence if you include custom dataset creation services. - For a pitch or report: quote the range of 2025 estimates, $3.2 billion to $8.74 billion, with both firms named - For a forecast: show at least two forecasts with their horizons, since 2029, 2031, 2033 and 2035 are not comparable - For pricing a deal: ignore market size and use published price ranges and comparable deals - For a seller's valuation: use buyer offers and deal values in your data type, not a share of the market For prices, our [AI data price index](https://sellmydatatoai.net/ai-data-price-index) compiles published payout ranges and reported deal values in one table, and our guide to [AI training data pricing models](https://sellmydatatoai.net/guides/ai-training-data-pricing-models) explains how deals are structured. ## What does market growth mean for AI data buyers? Market growth means more competition for scarce, rights-cleared data, especially proprietary company records and expert work. In the 2026 auction for bankrupt Spirit Airlines' de-identified business data, Google's reported $10 million bid won, micro1 filed a competing bid and Mercor was named as alternate bidder. Court approval was pending at the time of reporting. For buyers, that argues for starting procurement early and paying attention to exclusivity. Our guide to [running an AI training data procurement](https://sellmydatatoai.net/guides/ai-training-data-procurement) sets out the process, and our guide to [synthetic vs licensed training data](https://sellmydatatoai.net/guides/synthetic-vs-licensed-training-data) covers when real data is worth the cost. ## How does Sell My Data To AI fit into this market? Sell My Data To AI is an independent broker between data owners and AI teams. We do not own an AI model and we do not buy data for our own use. Buyers pay nothing to submit a requirement or review a shortlist, and sellers pay nothing to be assessed. To source data, [submit a data requirement](https://sellmydatatoai.net/request-data). To license your own records, start with a [data valuation](https://sellmydatatoai.net/data-valuation). ## Frequently asked questions ### What is the AI training dataset market worth in 2026? Estimates for 2026 range from $3.9 billion to $11.91 billion. Grand View Research estimates $3.9 billion in 2026, Precedence Research $3.92 billion and Mordor Intelligence $11.91 billion, according to their report pages checked in November 2026. These are research firm estimates using different definitions of the market, so always cite the firm and year. ### How fast is the AI training data market growing? Research firms forecast annual growth from 16.13% to 33.14%. Precedence Research forecasts 16.13% a year from 2026 to 2035, Grand View Research 22.6% from 2026 to 2033, MarketsandMarkets 27.7% to 2029 and Mordor Intelligence 33.14% from 2026 to 2031. Different base years and horizons mean these rates are not directly comparable. ### How much has Reddit earned from licensing data to AI companies? Reddit's February 2024 Form S-1 states that in January 2024 it entered data licensing arrangements with an aggregate contract value of $203.0 million and terms of two to three years. It expected at least $66.4 million of that as revenue in 2024. Reuters separately reported Google's deal at about $60 million a year, citing one source. ### What is the biggest reported AI data licensing deal? The largest reported value in our tracker is News Corp's deal with OpenAI, which The Wall Street Journal reported at about $250 million over five years in cash and credits. Neither company disclosed terms. Reddit's deal with Google was reported by Reuters at about $60 million a year, and many other large deals have no reported value. ### Why do market research firms disagree about AI data? Market research firms disagree because they define the AI training data market differently, start from different base years and use unpublished models. Mordor Intelligence includes custom dataset creation services and reports $8.74 billion for 2025, while Grand View Research segments by data type and vertical and reports $3.2 billion. Neither lists the deals it counts. ## Sources - [AI training dataset market size and share report, 2026 to 2033](https://www.grandviewresearch.com/industry-analysis/ai-training-dataset-market), Grand View Research, accessed November 2026 - [AI training dataset market companies, size and trends 2026 to 2035](https://www.precedenceresearch.com/ai-training-dataset-market), Precedence Research, accessed November 2026 - [AI training dataset market size and share, forecasts report 2034](https://www.gminsights.com/industry-analysis/ai-training-dataset-market), Global Market Insights, accessed November 2026 - [The rise of AI training dataset market: a $9.58 billion industry (MarketsandMarkets)](https://www.globenewswire.com/news-release/2024/11/26/2987491/0/en/The-Rise-of-AI-Training-Dataset-Market-A-9-58-billion-Industry-Dominated-by-Tech-Giants-Scale-AI-US-Appen-Australia-AWS-US-Sama-US-MarketsandMarkets.html), MarketsandMarkets via GlobeNewswire, accessed November 2026 - [AI training dataset market size, share and 2031 growth trends report](https://www.mordorintelligence.com/industry-reports/ai-training-dataset-market), Mordor Intelligence, accessed November 2026 - [AI training dataset market worth $9.35 billion by 2029](https://www.meticulousresearch.com/index.php/pressrelease/628/ai-training-dataset-market-2029), Meticulous Research, accessed November 2026 - [Reddit, Inc. Form S-1 registration statement](https://www.sec.gov/Archives/edgar/data/1713445/000162828024006294/reddits-1q423.htm), US Securities and Exchange Commission (EDGAR), accessed November 2026 - [Shutterstock, Inc. Form 10-K for fiscal year 2023](https://www.sec.gov/Archives/edgar/data/1549346/000154934624000007/sstk-20231231.htm), US Securities and Exchange Commission (EDGAR), accessed November 2026 - [Scale AI announces next phase of company's evolution](https://seekingalpha.com/pr/20135874-scale-ai-announces-next-phase-of-company-s-evolution), Scale AI press release via Seeking Alpha, accessed November 2026 - [MIT Technology Review, October 2026](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models), MIT Technology Review, accessed November 2026 - [The Next Web, October 2026](https://thenextweb.com/news/google-spirit-airlines-data-ombudsman-approval), The Next Web, accessed November 2026 - [Engadget, citing The Wall Street Journal, July 2025](https://www.engadget.com/ai/the-new-york-times-and-amazons-ai-licensing-deal-is-reportedly-worth-up-to-25-million-per-year-135523853/), Engadget, citing The Wall Street Journal, accessed November 2026 - [Publishers Weekly, November 2024](https://www.publishersweekly.com/pw/print/20241125/96533-agents-authors-question-harpercollins-ai-deal.html), Publishers Weekly, accessed November 2026 - [Engadget, citing Adweek, November 2024](https://www.engadget.com/ai/openai-will-pay-dotdash-meredith-at-least-16-million-per-year-to-license-its-content-191756662.html), Engadget, citing Adweek, accessed November 2026 - [The Bookseller, August 2024](https://thebookseller.com/news/wiley-set-to-earn-44m-from-ai-rights-deals-confirms-no-opt-out-for-authors), The Bookseller, accessed November 2026 - [Engadget, citing The Wall Street Journal, May 2024](https://www.engadget.com/openai-will-reportedly-pay-250-million-to-put-news-corps-journalism-in-chatgpt-214615249.html?src=rss), Engadget, citing The Wall Street Journal, accessed November 2026 - [interactive investor, May 2024](https://www.ii.co.uk/analysis-commentary/ii-view-confident-informa-does-ai-data-deal-microsoft-ii531624), interactive investor, accessed November 2026 - [Reuters via Nieman Lab, February 2024](https://www.niemanlab.org/reading/reddit-in-ai-content-licensing-deal-with-google-sources-say/), Reuters via Nieman Lab, accessed November 2026 - [Bloomberg Law, December 2023](https://news.bloomberglaw.com/tech-and-telecom-law/openai-to-pay-axel-springer-tens-of-millions-to-use-news-content), Bloomberg Law, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Sell accounting and tax firm data to AI companies URL: https://sellmydatatoai.net/sell-your-data/accounting-firms Accounting and tax firms can license working papers, tax computations, review notes and technical manuals to AI companies, and Mercor publishes a range of $25,000 to $865,000 per firm. Client information is the constraint: the ICAEW Code of Ethics bars using it for the firm's advantage without authority, and US preparers also need taxpayer consent under Section 7216. ## Key takeaways - AI buyers most want these records from accounting and tax firms: (1) working papers with reviewer notes and sign-offs, (2) tax computations showing each adjustment and its reasoning and (3) client query logs and technical helpdesk answers. - Published buyer range for accounting and tax firms: $25,000 to $865,000 per company. - Source: Mercor's published referral listing for accounting and tax firms, as reported by ddiy.co, checked November 2026. - Size buyers look for: 5 to 200 employees (Mercor's listing). - The first rights issue to clear: ICAEW Code R114.2: no use of client information for the firm's advantage, including former clients. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Working papers with reviewer notes and sign-offs; Tax computations showing each adjustment and its reasoning; Client query logs and technical helpdesk answers; Advisory reports and planning memos, de-identified; Firm audit and tax manuals, checklists and training material - Why buyers want them: Finance and tax models need examples of professional judgment applied to real numbers. Review notes, corrections and the reasoning behind adjustments show that judgment in a way finished accounts do not. - Published buyer range: $25,000 to $865,000 per company (Mercor's published referral listing for accounting and tax firms, as reported by ddiy.co, checked November 2026) - Size buyers look for: 5 to 200 employees (Mercor's listing) - Rights issues: ICAEW Code R114.2: no use of client information for the firm's advantage, including former clients; IRS Section 7216 consent rules for US tax return information; Money laundering report records and tipping off under POCA section 333A; Engagement letter terms on confidentiality and ownership of papers; Client payroll and identity documents containing personal data ## What accounting firm data do AI companies buy? AI companies buy accounting firm data that shows a qualified person applying judgment to real figures. Finished statutory accounts are filed at Companies House for most UK companies, so they are already public. The work behind them is not. - Working papers that keep the preparer's schedules and the reviewer's notes - Tax computations where each adjustment is explained - Client query logs, with the question, the research and the answer given - Internal technical helpdesk threads between staff and the tax or audit technical team - Audit and tax manuals, checklists and staff training material The review layer is what buyers value most. A schedule prepared by a junior and corrected by a manager is a labeled example of a mistake and its fix, which is exactly what [expert domain data](https://sellmydatatoai.net/buy-ai-training-data/expert-domain-data) buyers look for. ## How much do AI companies pay for accounting firm data? AI companies publish a range of $25,000 to $865,000 for accounting and tax firm data. That is Mercor's referral listing for firms of 5 to 200 employees, as reported by ddiy.co in October 2026. Mercor does not publish what places a firm at either end of that range. The factors buyers price are the same as in other sectors: years of records, how much reviewed work survives the rights checks, and whether the license is exclusive. Our [valuation guide](https://sellmydatatoai.net/guides/how-much-is-your-data-worth-to-ai) explains each one. ## Does the ICAEW Code of Ethics allow you to license client data? The ICAEW Code of Ethics allows disclosure of client information only where there is a legal or professional duty or right, or where the client authorizes it and the law does not prohibit it. That is R114.3 of the Code, summarized in ICAEW's Technical Advisory Service helpsheet on disclosure of confidential information, reviewed in October 2025. Two other points from section 114 bear directly on a data sale. R114.2 says a member must not use confidential information for the advantage of the accountant, the firm or a third party, which is what licensing it to an AI buyer would do. The duty also covers former clients and continues after the relationship ends. So the practical routes are client authorization, or material that no longer contains confidential client information. Whether well de-identified working papers meet that test is a question for your own adviser, and ICAEW's Technical Advisory Service can help members with ethics queries. ## What does Section 7216 mean for US tax preparers? Section 7216 means a US tax return preparer cannot disclose or use tax return information without taxpayer consent, except where the regulations allow it. The IRS Section 7216 information center lists consent requirements set out in Rev. Proc. 2013-14 and refers to criminal and civil penalties under sections 7216 and 6713. The IRS also notes that the 2010 rules allow statistical compilations, but only if they are anonymous as to taxpayer identity and contain data from at least 10 tax returns, and other prohibitions still apply. A UK firm with US clients needs advice on this before any return data goes into a package. ## What should an accounting firm leave out? An accounting firm should leave out records where the law restricts disclosure regardless of consent, and records with personal data but little training value. | Record | Issue | Usual approach | | --- | --- | --- | | Money laundering reports and related notes | POCA section 333A makes disclosure likely to prejudice an investigation an offense in the regulated sector | Exclude | | US tax returns and return information | IRS Section 7216 consent rules | Exclude unless consent and advice are in place | | Client payroll and employee records | Personal data of the client's staff | Exclude | | Identity documents and bank statements | High-risk personal and financial data | Exclude | | Reviewed working papers and computations | Client confidential information under ICAEW R114 | Include only with authorization or after de-identification and advice | | Firm manuals, checklists and training | Firm-owned | Include | ## Which accounting records are easiest to clear? The accounting records easiest to clear are the ones the firm wrote for itself. Audit and tax manuals, review checklists, template engagement letters and internal training courses contain no client information, and the firm owns the copyright in work its employees created under section 11(2) of the Copyright, Designs and Patents Act 1988. Internal technical helpdesk threads come next. A question from a manager and an answer from the technical partner can often be de-identified by removing the client name and figures, while keeping the reasoning. Check anything bought in, such as third-party manuals or subscription content, because it is licensed to you under the supplier's terms. ## How does selling accounting firm data work with Sell My Data To AI? Selling accounting firm data with Sell My Data To AI starts with a consultation about your practice management system, document store and the years covered. We tell you if we think buyers will not be interested. If there is interest, we help you separate firm-owned material from client material, record the split in a [data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest) and run our [eight vetting checks](https://sellmydatatoai.net/how-we-vet). Buyers you approve see a redacted sample under a mutual NDA, and you sign the license directly with them. [Book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller) to start. ## Frequently asked questions ### Can an accountant sell anonymized client data to AI companies? An accountant can only disclose client information where a legal or professional right exists or the client authorizes it, under R114.3 of the ICAEW Code of Ethics. Whether anonymized working papers still count as confidential client information is a question for your own adviser. Many firms start with firm-owned manuals and training material, which need no client authority. ### Do I need taxpayer consent to include US tax returns? In most cases yes. The IRS says Section 7216 governs disclosure and use of tax return information by preparers without taxpayer consent, with consent rules in Rev. Proc. 2013-14. Statistical compilations are allowed only if anonymous and drawn from at least 10 returns. Take US tax advice before including any return information. ### Are audit working papers worth more than tax files? Neither type is published as worth more. Mercor's referral listing, reported by ddiy.co, gives one range of $25,000 to $865,000 for accounting and tax firms combined. What lifts value is reviewer notes, years of history and a clear rights position, whichever service line the papers come from. ### Will selling data affect my firm's AML obligations? Selling data does not change your anti-money laundering obligations, but it can create a tipping off risk. Section 333A of the Proceeds of Crime Act 2002 makes it an offense in the regulated sector to disclose a report in a way likely to prejudice an investigation. Exclude all money laundering reporting records from any package. ## Sources - [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), ddiy.co (figures taken from each buyer's own site), accessed November 2026 - [AI labs are buying data from dead startups to train their models](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models), MIT Technology Review, accessed November 2026 - [Disclosure of confidential information (for members in practice)](https://www.icaew.com/technical/tas-helpsheets/ethics/disclosure-of-confidential-information-for-members-in-practice), ICAEW Technical Advisory Service, accessed November 2026 - [Section 7216 information center](https://www.irs.gov/node/7488), Internal Revenue Service, accessed November 2026 - [Proceeds of Crime Act 2002, section 333A: tipping off in the regulated sector](https://www.legislation.gov.uk/ukpga/2002/29/section/333A), legislation.gov.uk, accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 11](https://www.legislation.gov.uk/ukpga/1988/48/section/11), legislation.gov.uk, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Sell call center data to AI companies URL: https://sellmydatatoai.net/sell-your-data/call-centers Call centers and BPOs can license call recordings, transcripts, QA scorecards and agent knowledge bases to AI companies, and Mercor publishes a range of $20,000 to $250,000 per company. The first question is ownership: an outsourcer usually holds recordings as its clients' processor. Card codes, voiceprints and health details in calls must also come out. ## Key takeaways - AI buyers most want these records from call centers and bpos: (1) call recordings with transcripts, (2) chat and email transcripts and (3) QA scorecards linked to calls. - Published buyer range for call centers and bpos: $20,000 to $250,000 per company. - Source: Mercor's published referral listing for call centers, as reported by ddiy.co, checked November 2026. - Size buyers look for: 50 to 5,000 agents (Mercor's listing). - The first rights issue to clear: Outsourcers hold client recordings as processor under UK GDPR Article 28. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Call recordings with transcripts; Chat and email transcripts; QA scorecards linked to calls; Agent scripts and knowledge base articles; Escalation notes and disposition codes; Calls in languages other than English - Why buyers want them: Voice agents and customer service models need real conversations with real outcomes. Recordings linked to QA scores and resolutions show what good and bad handling sounds like. - Published buyer range: $20,000 to $250,000 per company (Mercor's published referral listing for call centers, as reported by ddiy.co, checked November 2026) - Size buyers look for: 50 to 5,000 agents (Mercor's listing) - Rights issues: Outsourcers hold client recordings as processor under UK GDPR Article 28; PCI DSS Requirement 3.3.1: no card validation codes in recordings; Voiceprints under Illinois BIPA section 15(c); Health and other special category details disclosed by callers; Recording notices and call-recording consent laws ## What call center data do AI companies buy? AI companies buy call center data that pairs real conversations with what happened next. A recording on its own is useful. The same recording with a transcript, the QA score, the disposition code and whether the customer called back is far more useful. - Call recordings with transcripts and speaker labels - Chat and email transcripts from the same queues - QA scorecards linked to the calls they assess - Agent scripts, call flows and knowledge base articles - Escalation notes, disposition codes and resolution outcomes - Calls in languages other than English, which are scarcer Buyers training voice agents want the audio. Buyers training text models may want transcripts only, which are easier to de-identify. See [call recording data](https://sellmydatatoai.net/buy-ai-training-data/call-recording-data) and [speech and audio data](https://sellmydatatoai.net/buy-ai-training-data/speech-and-audio-data) for how each is used. If your platform records the agent and the caller on separate channels, keep them separate in the export. Split channels make speaker labeling and redaction much easier. Record the codec, sample rate, queue and date range for each batch in your [data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest). ## How much do AI companies pay for call center data? AI companies publish a range of $20,000 to $250,000 for call center data. That is Mercor's referral listing for call centers of 50 to 5,000 agents, as reported by ddiy.co in October 2026. The $250,000 ceiling is the lowest of the 11 Mercor listings ddiy.co reports. Call data is plentiful, so buyers pay more for linked outcomes, specialist queues such as technical support or claims handling, and languages that are hard to source. Transcripts from regulated queues, such as debt collection or insurance claims, need more careful redaction. A [free data valuation](https://sellmydatatoai.net/data-valuation) tells you whether yours stands out. ## Who owns the recordings: you or your clients? The recordings usually belong to your clients if you are an outsourcer, and to you if you run an in-house contact center. This is the first question to settle, because it decides whether there is anything to sell. A BPO handling calls for a client normally processes the callers' personal data as the client's processor. Article 28(3)(a) of the UK GDPR says a processor processes personal data only on documented instructions from the controller, and Article 28(10) treats a processor that decides its own purposes as a controller. Licensing client recordings without the client's written agreement would breach that. What an outsourcer usually owns outright is its own material: agent training content, QA frameworks, scripts it wrote, and internal operations data. In-house contact centers face a simpler question, because the company that took the calls usually controls the recordings, but the privacy notice and recording message in force at the time still apply. Read each outsourcing contract for any rights it gives you over de-identified or aggregated data. ## What must be removed from call recordings? Card details, voiceprints and sensitive personal details must be removed from call recordings before any sale. Each is governed by its own rule. | Content | Rule | What to do | | --- | --- | --- | | Card validation codes spoken on calls | PCI DSS Requirement 3.3.1 prohibits storing sensitive authentication data after authorization, including in audio | Confirm suppression was in place, exclude any calls where it was not | | Card numbers and payment details | PCI DSS and the client's card scheme obligations | Redact from audio and transcripts | | Voiceprints from voice authentication | Illinois BIPA section 15(c) bars selling or profiting from biometric identifiers | Exclude | | Health, religion and other special category details | UK GDPR Article 9 | Exclude calls or redact, depending on queue | | Names, addresses, account numbers | Personal data | Redact from audio and transcripts | The PCI Security Standards Council's FAQ 1210 says storing card validation codes in any digital audio recording after authorization breaches Requirement 3.3.1. Under BIPA, which lists voiceprints as biometric identifiers, a negligent violation carries liquidated damages of $1,000 and an intentional or reckless one $5,000, per violation. ## Do callers' recording notices cover AI training? Callers' recording notices often do not cover AI training by a third party. A message saying calls are recorded for training and quality purposes was written with your own staff in mind, and may not extend to licensing the recording to another company. Frankfurt Kurnit Klein and Selz notes that state call-recording consent laws apply in the US, and that companies must comply with the privacy statements in force when data was collected. Agents' voices are personal data too. These points make de-identified transcripts and your own content the easier place to start, see [de-identification for AI training data](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data). ## How does selling call center data work with Sell My Data To AI? Selling call center data with Sell My Data To AI starts with a consultation about your platforms, queues, languages and client contracts. We work out which data you own, which needs client agreement and which must be excluded, before anyone else is involved. We then run our [eight vetting checks](https://sellmydatatoai.net/how-we-vet), including a re-identification test on a sample, and share a redacted sample only with buyers you approve. [Book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller) to start. ## Frequently asked questions ### Can a BPO sell recordings of calls it handled for clients? Not without the client's written agreement in most cases. A BPO normally processes callers' personal data as its client's processor, and Article 28(3)(a) of the UK GDPR limits a processor to the controller's documented instructions. Check each outsourcing contract for any rights over de-identified data, and take independent advice. ### Are transcripts easier to sell than audio? Transcripts are usually easier to clear, because names, numbers and other identifiers can be redacted and checked in text. Audio carries the speaker's voice, which is personal data, and voice authentication systems can create voiceprints, which Illinois BIPA treats as biometric identifiers. Buyers training voice agents still want audio, so the choice depends on the buyer. ### What if card numbers were read out on recorded calls? Calls where card validation codes were captured should be excluded. The PCI Security Standards Council says storing card validation codes in any digital audio recording after authorization breaches PCI DSS Requirement 3.3.1. Card numbers and other payment details should be redacted from both audio and transcripts before a sample is drawn. ### How large does a call center need to be? Mercor's referral listing, reported by ddiy.co, covers call centers of 50 to 5,000 agents, with a published range of $20,000 to $250,000. Linked outcomes, specialist queues and languages other than English matter more to buyers than raw call volume, because general customer service calls are widely available. ## Sources - [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), ddiy.co (figures taken from each buyer's own site), accessed November 2026 - [UK GDPR, Article 28: processor](https://www.legislation.gov.uk/eur/2016/679/article/28), legislation.gov.uk, accessed November 2026 - [UK GDPR, Article 9: processing of special categories of personal data](https://www.legislation.gov.uk/eur/2016/679/article/9), legislation.gov.uk, accessed November 2026 - [Are audio/voice recordings permitted to contain sensitive authentication data? (FAQ 1210)](https://www.pcisecuritystandards.org/faqs/1210/), PCI Security Standards Council, accessed November 2026 - [Biometric Information Privacy Act (740 ILCS 14)](https://ilga.gov/legislation/ilcs/ilcs3.asp?ActID=3004), Illinois General Assembly, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Sell construction company data to AI companies URL: https://sellmydatatoai.net/sell-your-data/construction-companies Construction companies can license RFIs, change orders, site diaries, defect reports and project programs to AI companies, and Mercor publishes a range of $115,000 to $1 million per company. Before a sale you need to check who owns the design files, what client and subcontractor contracts allow, and what personal data sits in site records. ## Key takeaways - AI buyers most want these records from construction companies: (1) RFIs and the responses that closed them, (2) change orders, variations and early warning notices and (3) site diaries and daily reports. - Published buyer range for construction companies: $115,000 to $1 million per company. - Source: Mercor's published referral listing for construction companies, as reported by ddiy.co, checked November 2026. - Size buyers look for: 50 to 2,000 employees (Mercor's listing). - The first rights issue to clear: Copyright in drawings and models stays with architects and engineers unless assigned. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: RFIs and the responses that closed them; Change orders, variations and early warning notices; Site diaries and daily reports; Snagging and defect reports with photos; Method statements, risk assessments and project programs; Tender submissions and bills of quantities - Why buyers want them: Project records document real problems and how they were resolved, in sequence and with costs and dates attached. That is the material AI tools for estimating, scheduling and project control need and cannot generate. - Published buyer range: $115,000 to $1 million per company (Mercor's published referral listing for construction companies, as reported by ddiy.co, checked November 2026) - Size buyers look for: 50 to 2,000 employees (Mercor's listing) - Rights issues: Copyright in drawings and models stays with architects and engineers unless assigned; Confidentiality clauses in main contracts and frameworks; Subcontractor and supplier pricing shared in confidence; Accident reports and site photos containing personal and health data; Security restrictions on sensitive sites ## What construction data do AI companies buy? AI companies buy construction data that records decisions on real projects in the order they happened. An RFI, the designer's answer, the change order that followed and the program update a week later form a complete example of a problem and its resolution. | Record | What it teaches a model | Typical system | | --- | --- | --- | | RFIs and responses | How design gaps are found and resolved | Project management platform or email | | Change orders and variations | Cost and time effects of scope changes | Commercial system or spreadsheets | | Site diaries and daily reports | Day-to-day progress, weather, labor and delays | Field app or scanned paper | | Snagging and defect reports | Defect types, causes and fixes, with photos | Snagging app | | Method statements and risk assessments | How work is planned safely | Document store | | Tenders and bills of quantities | How projects are priced | Estimating software | | Programs | Sequencing and the effect of delays | Primavera P6 or Microsoft Project | Linked records are worth more than isolated ones. If your RFIs carry the project, drawing reference and date, a buyer can connect them to the change orders and program updates that followed. Photos from defect reports are also [image data](https://sellmydatatoai.net/buy-ai-training-data/image-datasets) in their own right. ## How much do AI companies pay for construction company data? AI companies publish a range of $115,000 to $1 million for construction company data. That is Mercor's referral listing for construction companies of 50 to 2,000 employees, as reported by ddiy.co in October 2026. The $115,000 starting point is one of the highest in the 11 Mercor listings ddiy.co reports. Only consulting firms, at $150,000, start higher. The ceiling reflects what a large contractor with many projects and many years of linked records can offer. Our [free data valuation](https://sellmydatatoai.net/data-valuation) tells you where your archive is likely to sit. ## Who owns the drawings and models in your project files? The drawings and models in your project files are usually owned by whoever created them, which is often not you. Under section 11 of the Copyright, Designs and Patents Act 1988 the author is the first owner of copyright, and an employer owns work its employees create in the course of employment. That means drawings and BIM models produced by an independent architect or engineer belong to that consultant unless their appointment assigned the copyright. Many appointments grant a license to use the designs for the project instead. A project license will not usually cover licensing the files to an AI company, so check each appointment or leave consultant design files out. Your own records are a different matter. Site diaries, RFIs your staff wrote, method statements, programs and estimates produced by your employees belong to your company. See [chain of title](https://sellmydatatoai.net/glossary/chain-of-title) for how buyers check this. ## What do client and subcontractor contracts allow? Client and subcontractor contracts often restrict what you can do with project information, so read the confidentiality clauses before any project goes into a package. Public sector frameworks and private clients both use them. - Main contract confidentiality clauses, which may survive completion - Framework agreements with their own information rules - Subcontractor quotes and rates shared in confidence during tendering - Security requirements on sensitive sites such as prisons, defense estates and energy infrastructure Projects for clients with strict security requirements should be excluded outright. For others, removing client names, addresses and subcontractor rates is often enough to protect commercial confidentiality, but the contract decides, not the redaction. ## What personal data sits in construction records? Construction records hold more personal data than most contractors expect. Site photos show workers and members of the public, timesheets and induction records name individuals, and copies of competence cards carry photos and card numbers. Accident and incident reports usually describe injuries. Data concerning health is special category data under Article 9 of the UK GDPR, so those reports should be excluded rather than redacted. Photos can be kept if faces and identifying details are removed by a documented method and checked, as our [vetting standard](https://sellmydatatoai.net/how-we-vet) requires. ## Which construction records are easiest to clear? The easiest construction records to clear are the ones your own staff wrote about how your company works. Method statements, risk assessment templates, site diaries, lessons-learned reviews and estimating rates you built yourself contain little client information, and your company owns them under section 11(2) of the CDPA. RFIs and change orders come next, once client names, site addresses and subcontractor rates are removed. Consultant design files and anything from security-sensitive sites come last, or not at all. ## How does selling construction data work with Sell My Data To AI? Selling construction data with Sell My Data To AI starts with a consultation about which projects and systems you hold. We then help you build a project-by-project [data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest), marking which projects are cleared, which need client consent and which are excluded. Verified buyers you approve see a redacted sample drawn at random across projects, not your best job. You sign the license directly with the buyer. [Book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller) to find out what your project archive could be worth. ## Frequently asked questions ### Can a contractor sell drawings produced by its architects? A contractor can only license drawings if it owns the copyright or holds rights that allow it. Under section 11 of the Copyright, Designs and Patents Act 1988 the author is the first owner, so an independent architect owns their drawings unless the appointment assigned copyright. A license to use designs for the project does not usually extend to AI training. ### How many projects do AI buyers want to see? No buyer publishes a minimum number of projects. Mercor's listing, reported by ddiy.co, uses company size instead: 50 to 2,000 employees. More projects with linked records across several years give buyers more complete examples, which tends to support a higher price within the published range. ### Are scanned paper site diaries worth anything? Scanned site diaries can be worth including, especially if they cover many years and are legible. Buyers will run text recognition on them, so scan quality matters. Make sure the scans are indexed by project and date in your data manifest, and remove names of individuals that appear in the entries. ### Do we need client permission to include a finished project? You need client permission if the contract's confidentiality clause covers the information and survives completion, which many do. If the clause does not apply, removing the client's name and site address can still be a sensible protection. Check each contract and take independent legal advice on any you are unsure of. ## Sources - [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), ddiy.co (figures taken from each buyer's own site), accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 11](https://www.legislation.gov.uk/ukpga/1988/48/section/11), legislation.gov.uk, accessed November 2026 - [UK GDPR, Article 9: processing of special categories of personal data](https://www.legislation.gov.uk/eur/2016/679/article/9), legislation.gov.uk, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How consulting and advisory firms can sell data to AI companies URL: https://sellmydatatoai.net/sell-your-data/consulting-firms Consulting and advisory firms can license their methodologies, frameworks, proposal libraries, templates and anonymized engagement records to AI companies. Mercor's published listing puts consulting firm payouts at $150,000 to $300,000 for firms with 10 to 500 staff. Engagement contracts often give clients the deliverables, so the firm's own know-how is usually the licensable core. ## Key takeaways - AI buyers most want these records from consulting and advisory firms: (1) methodologies, frameworks and toolkits your staff developed, (2) proposal and pitch libraries with scoping and pricing rationale and (3) templates, models and analysis workbooks. - Published buyer range for consulting and advisory firms: $150,000 to $300,000 per company. - Source: Mercor's published listing for consulting firms, recorded by ddiy.co, checked October 2026. - Size buyers look for: 10 to 500 staff (Mercor). - The first rights issue to clear: Engagement contracts often assign deliverable IP to the client. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Methodologies, frameworks and toolkits your staff developed; Proposal and pitch libraries with scoping and pricing rationale; Templates, models and analysis workbooks; Internal knowledge base and training materials; Engagement plans, workstream trackers and lessons-learned reviews; Anonymized deliverables where the contract leaves rights with you - Why buyers want them: Consulting records show structured reasoning applied to business problems. micro1's brief asks for SOPs, knowledge bases, templates and project histories, and a consulting firm's library contains all four. - Published buyer range: $150,000 to $300,000 per company (Mercor's published listing for consulting firms, recorded by ddiy.co, checked October 2026) - Size buyers look for: 10 to 500 staff (Mercor) - Rights issues: Engagement contracts often assign deliverable IP to the client; Client facts and data are confidential under engagement letters and NDAs; Transaction work may contain inside information under UK MAR; Expert interview transcripts hold third parties' views given in confidence; Associates' and subcontractors' work needs a written IP assignment Mercor's published range for consulting firms is the narrowest of the 11 Mercor industry listings ddiy.co recorded in October 2026: $150,000 to $300,000 per company, for firms with 10 to 500 staff. It also has the highest floor of the 11. ## How much do AI companies pay for consulting firm data? AI companies pay consulting firms $150,000 to $300,000 per company, according to Mercor's published listing. The $150,000 floor compares with $50,000 for law firms and software companies and $115,000 for construction companies in the same set of listings. Other buyers publish general terms. micro1's tiers start at $100,000 to $500,000 for a qualified partnership, and it asks for 30 or more full-time staff. Our guide to [what your data is worth](https://sellmydatatoai.net/guides/how-much-is-your-data-worth-to-ai) explains the factors that move a price within a range. ## What consulting data do AI companies buy? AI companies buy consulting data that shows how a problem is framed, analyzed and answered. Methodologies, proposal libraries and engagement reviews do that without exposing a client, which makes them the natural core of a consulting license. | Record | Usually owned by | Licensing position | | --- | --- | --- | | Methodologies and frameworks | The firm | Licensable if staff and associates assigned IP | | Proposal and pitch library | The firm | Licensable once client names and prices are removed | | Templates and models | The firm | Licensable, check for embedded client data | | Final deliverables | Often the client, by contract | Check each engagement contract | | Client data and interview notes | The client or the interviewee | Exclude | | Deal and transaction files | The client | Exclude, may hold inside information | ## Do you own your client deliverables? You own client deliverables only if the engagement contract leaves the rights with you. Section 11 of the Copyright, Designs and Patents Act 1988 makes the author, or the employer of an employee author, the first owner, but many consulting contracts assign deliverable IP to the client and leave only pre-existing materials and know-how with the firm. Even where you keep the copyright, the facts inside a deliverable are usually confidential to the client. Daniel Goldberg of Frankfurt Kurnit Klein and Selz notes that email and chat contain other companies' confidential information, often covered by NDAs and commercial agreements, and holding data does not give you the right to sell it. ## Can consulting firms license transaction and deal files? Consulting firms should leave transaction and capital markets files out of any data license. The [FCA explains](https://www.fca.org.uk/markets/market-abuse/inside-information-how-to-identify-control-disclose) that Article 10 of UK MAR sets out what amounts to unlawful disclosure of inside information, and that disclosure is permitted only where necessary in the normal exercise of employment, a profession or duties. Older deal files may no longer hold inside information, but checking that file by file costs more than excluding them. Take independent advice if transaction work is a large part of your archive. ## What else should consulting firms check? - Associates and subcontractors: confirm their contracts assign IP in methodologies and templates they helped build - Expert interviews: transcripts carry the views of third parties who spoke in confidence - Client logos and case studies: marketing permissions are unlikely to cover data licensing - Staff names and CVs in proposals: remove or de-identify them > **Know-how versus client facts** Your method is yours to license. Your client's numbers, people and plans are not. Separate the two before you assemble a [representative sample](https://sellmydatatoai.net/glossary/representative-sample). ## How do you sell consulting data through Sell My Data To AI? We start with a consultation, then a rights pre-check against your engagement contracts. Every seller passes our eight-point vetting standard before a buyer sees anything, including ownership and chain of title and contract restrictions. Sellers pay nothing to be assessed, and our success fee is agreed in writing before any introduction. Read the [licensing agreement checklist](https://sellmydatatoai.net/guides/ai-data-licensing-agreement-checklist), or [book a consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### How much does Mercor pay consulting firms for data? Mercor's published listing puts consulting firm payouts at $150,000 to $300,000 per company, for firms with 10 to 500 staff, as recorded by ddiy.co in October 2026. Its $150,000 floor is the highest of the 11 Mercor industry ranges ddiy.co reported. Rights clarity and the depth of your methodology library affect where you land. ### Can we license anonymized client reports? You can license anonymized client reports only where the engagement contract leaves you the rights and the client's confidential facts are removed. Anonymizing a report does not override a confidentiality clause. Check each contract, and expect to rely more on your methodologies and templates than on finished deliverables. ### Will clients find out we licensed our data? Clients may find out, and your engagement contracts may require you to tell them about new uses of material connected to their work. Planning the conversation early avoids surprises. Our guide to [telling staff and clients about a data license](https://sellmydatatoai.net/guides/telling-staff-and-clients-about-a-data-sale) covers what to say and when. ### Do we need our associates' permission? You need your associates' permission, or an existing written IP assignment, for any material they created. Copyright belongs to the author unless they are an employee working in the course of employment or have assigned it in writing. Our vetting checks chain of title, including assignments from staff and contractors. ## Sources - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 - [Monetize your company's operational data](https://www.micro1.ai/data), micro1, accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 11: first ownership of copyright](https://www.legislation.gov.uk/ukpga/1988/48/section/11), legislation.gov.uk, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz (Daniel Goldberg), accessed November 2026 - [Inside information: how to identify, control and disclose](https://www.fca.org.uk/markets/market-abuse/inside-information-how-to-identify-control-disclose), Financial Conduct Authority, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How cybersecurity companies can sell data to AI companies URL: https://sellmydatatoai.net/sell-your-data/cybersecurity-companies Cybersecurity companies can license detection rules, incident response playbooks, SOC triage notes, anonymized penetration test findings and training lab environments to AI companies. Mercor's published listing puts cybersecurity company payouts at $20,000 to $750,000 for firms with 20 to 2,000 staff. Threat intelligence shared under TLP restrictions, client findings and export-controlled tooling need clearing first. ## Key takeaways - AI buyers most want these records from cybersecurity companies: (1) detection rules, signatures and queries your team wrote, (2) incident response playbooks and runbooks and (3) SOC triage notes and case histories, de-identified. - Published buyer range for cybersecurity companies: $20,000 to $750,000 per company. - Source: Mercor's published listing for cybersecurity companies, recorded by ddiy.co, checked October 2026. - Size buyers look for: 20 to 2,000 staff (Mercor). - The first rights issue to clear: Intelligence received under TLP:RED, TLP:AMBER+STRICT or TLP:AMBER cannot be passed to a buyer. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Detection rules, signatures and queries your team wrote; Incident response playbooks and runbooks; SOC triage notes and case histories, de-identified; Penetration test methodologies and anonymized findings; Threat intelligence reports you authored; Training labs, capture-the-flag challenges and test environments - Why buyers want them: Security work is a sequence of observations, tool calls and decisions, the same structure as an agent trajectory. AI labs also buy replayable environments, and training labs and challenge sets can become RL environments. - Published buyer range: $20,000 to $750,000 per company (Mercor's published listing for cybersecurity companies, recorded by ddiy.co, checked October 2026) - Size buyers look for: 20 to 2,000 staff (Mercor) - Rights issues: Intelligence received under TLP:RED, TLP:AMBER+STRICT or TLP:AMBER cannot be passed to a buyer; Client penetration test and incident reports are confidential to the client; Logs carry IP addresses and usernames, which are online identifiers; Exploit tooling may fall under UK export control entries 4D004 and 4E001.c; Malware samples and third-party tools are not yours to license Cybersecurity firms produce records that match two formats AI labs buy. [Agent trajectories](https://sellmydatatoai.net/buy-ai-training-data/agent-trajectory-data) log observations, tool calls and results, and replayable environments let a model attempt a task and be scored. A SOC analyst's case history and a training lab are those formats in working form. ## How much do AI companies pay for cybersecurity data? AI companies pay cybersecurity firms $20,000 to $750,000 per company, according to Mercor's published listing, for firms with 20 to 2,000 staff. ddiy.co recorded the range from Mercor's referral page in October 2026. The floor matches the $20,000 starting point Mercor gives for agencies, call centers and AI product companies. Where a firm lands depends on volume, how much of the material is its own work rather than client material, and how cleanly it can be de-identified. ## What cybersecurity data do AI companies buy? AI companies buy cybersecurity data that captures analyst reasoning and repeatable tasks. The table sets out the common record types and the issue each raises. | Record | Format a buyer sees | Main issue | | --- | --- | --- | | Detection rules and queries | Code with intent and test cases | Low if written in-house | | IR playbooks and runbooks | Procedural knowledge | Client-specific details | | SOC case histories | Agent trajectories | IP addresses, usernames and client identity | | Penetration test findings | Expert reasoning on real systems | Client confidentiality and open vulnerabilities | | Threat intelligence | Analysis text | TLP labels and source restrictions | | Training labs and challenges | RL environments | Exploit code and export controls | A lab with tasks that can be reset, attempted and scored automatically is an [RL environment](https://sellmydatatoai.net/glossary/rl-environment) in all but name. ## Can you license threat intelligence you received from others? You can license threat intelligence you received from others only if the source allows it. Under FIRST's [Traffic Light Protocol](https://www.first.org/tlp/) version 2.0, TLP:RED is for named recipients only, TLP:AMBER+STRICT stays within your organization, TLP:AMBER may go to your organization and its clients on a need-to-know basis, and TLP:GREEN stays within your community. Only TLP:CLEAR material may be passed on without limit, subject to copyright. Feeds bought from vendors carry their own license terms. License your own analysis and remove anything received under a sharing agreement. ## Do export controls apply to security data? Export controls can apply to security data that includes exploit tooling. The [UK Strategic Export Control List](https://assets.publishing.service.gov.uk/media/69415ba1f065108822537524/uk_export_control_list_2025.pdf), consolidated in December 2025, controls software specially designed or modified for the generation, command and control or delivery of intrusion software (entry 4D004) and technology for the development of intrusion software (4E001.c). The same list notes that 4E001.c does not control vulnerability disclosure or cyber incident response. Detection content and IR records are therefore lower risk than offensive tooling, but a training lab built around working exploits needs a specialist export control view before it goes to an overseas buyer. ## What other rights issues apply to cybersecurity data? Client reports are the main contractual issue. Penetration test and incident response reports describe a named client's systems and weaknesses, and engagement terms usually make them confidential. Removing the client's name does not always remove the risk, because system details can identify an organization. Logs and case notes carry IP addresses, hostnames and usernames. UK GDPR Article 4(1) lists an online identifier as a way a person can be identified, so treat log data as personal data until it is [de-identified](https://sellmydatatoai.net/glossary/de-identification). > **License your work, not your clients' secrets** Rules, playbooks, methods and labs you built are the cleanest assets. Client findings, received intelligence and working exploits each need a separate decision. ## How do you sell cybersecurity data through Sell My Data To AI? We start with a consultation. We do not take custody of your raw data, and full data moves only after you sign a license directly with the buyer. Every buyer is verified for intended use, security and retention commitments and onward transfer limits before seeing a sample, which matters when material describes attack techniques. [Book a consultation](https://sellmydatatoai.net/book-a-consultation?side=seller) to find out what your records are worth. ## Frequently asked questions ### How much does Mercor pay cybersecurity companies? Mercor's published listing puts cybersecurity company payouts at $20,000 to $750,000 per company, for firms with 20 to 2,000 staff, as recorded by ddiy.co in October 2026. Firms with in-house detection content, playbooks and de-identified case histories have the most material that can be licensed without client consent. ### Can we sell penetration test reports? You can rarely sell penetration test reports as they stand, because they describe a named client's systems and are confidential under the engagement terms. Your methodology, report templates and fully anonymized findings are a better basis. Check each contract and take independent legal advice before including client material. ### Does TLP:GREEN intelligence count as shareable? TLP:GREEN intelligence may be shared with peers and partner organizations in your community, but not through publicly accessible channels and not outside the community, according to FIRST's TLP 2.0 definitions. Licensing it to an AI company falls outside that. Only TLP:CLEAR material may be passed on without limit. ### Are capture-the-flag challenges worth anything to AI labs? Capture-the-flag challenges and training labs can become RL environments, which AI labs buy alongside agent trajectories. Value depends on whether tasks are replayable, scored automatically and documented. Check export controls on any exploit code, and see our guide to [agent trajectories and RL environments](https://sellmydatatoai.net/guides/agent-trajectories-and-rl-environments). ## Sources - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 - [Traffic Light Protocol (TLP)](https://www.first.org/tlp/), FIRST, accessed November 2026 - [UK Strategic Export Control List, December 2025](https://assets.publishing.service.gov.uk/media/69415ba1f065108822537524/uk_export_control_list_2025.pdf), GOV.UK (Export Control Joint Unit), accessed November 2026 - [UK GDPR Article 4: definitions](https://www.legislation.gov.uk/eur/2016/679/article/4), legislation.gov.uk, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Sell design and web agency data to AI companies URL: https://sellmydatatoai.net/sell-your-data/design-and-web-agencies Design and web agencies can license briefs, layered design files, revision rounds, client feedback and front-end code to AI companies, and Mercor publishes a range of $20,000 to $890,000 per agency. The rights work is heavy: client contracts often assign the finished work, stock images and fonts carry their own licenses, and freelancers own their work unless they assigned it. ## Key takeaways - AI buyers most want these records from design and web agencies: (1) briefs linked to concepts and final designs, (2) layered design files with version history and (3) client feedback rounds and change requests. - Published buyer range for design and web agencies: $20,000 to $890,000 per company. - Source: Mercor's published referral listing for design and web agencies, as reported by ddiy.co, checked November 2026. - Size buyers look for: 10 to 200 employees (Mercor's listing). - The first rights issue to clear: Client contracts that assign copyright in delivered work. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Briefs linked to concepts and final designs; Layered design files with version history; Client feedback rounds and change requests; Front-end codebases and component libraries; Brand guidelines and design systems; Accessibility and UX audits - Why buyers want them: Design and coding models learn from how a brief becomes a finished design through rounds of feedback. The revision chain, with the client's comments and the designer's response, is the part buyers cannot get from finished public websites. - Published buyer range: $20,000 to $890,000 per company (Mercor's published referral listing for design and web agencies, as reported by ddiy.co, checked November 2026) - Size buyers look for: 10 to 200 employees (Mercor's listing) - Rights issues: Client contracts that assign copyright in delivered work; Stock image licenses that bar machine learning use; Font licenses limited to the client project; Freelancer work without a written IP assignment; User research recordings and personal data in client sites ## What design agency data do AI companies buy? AI companies buy design agency data that shows how a brief turns into a finished design. The final website is public and has likely been crawled already. The brief, the three rejected concepts, the client's comments and the revised layout are not. - Briefs and discovery notes linked to the concepts and final designs that answered them - Layered design files from Figma, Sketch or Adobe tools, with version history - Client feedback rounds and change requests, with the designer's responses - Front-end codebases, component libraries and design tokens - Brand guidelines and design systems - Accessibility and UX audits with their recommendations Linked records sell best. A buyer training a model to design or to write front-end code wants the brief, each version and the reason for each change, in order. Layered files are also [image data](https://sellmydatatoai.net/buy-ai-training-data/image-datasets), and the code side overlaps with [source code datasets](https://sellmydatatoai.net/buy-ai-training-data/source-code-datasets). ## How much do AI companies pay for design agency data? AI companies publish a range of $20,000 to $890,000 for design and web agency data. That is Mercor's referral listing for agencies of 10 to 200 employees, as reported by ddiy.co in October 2026. The $890,000 ceiling is higher than Mercor's ceiling for software companies, at $500,000. The $20,000 floor is the lowest of the 11 Mercor listings ddiy.co reports, shared with cybersecurity companies and call centers, so the spread is wide. Where an agency lands depends largely on how much of its archive survives the rights checks below. Organization also affects price. Archives that keep each project's brief, versions and feedback together are easier to value than files scattered across personal drives and email threads. If your projects sit in Figma teams or a shared drive organized by client, record that structure in your manifest so a buyer can see how complete each project is. ## Do you own the work you delivered to clients? You may not own the work you delivered to clients. Many agency contracts assign copyright in the final deliverables to the client once fees are paid, and if yours do, you cannot license those deliverables without the client's agreement. Unused concepts, internal working files and rejected routes are often treated differently, but the contract decides. Under section 11 of the Copyright, Designs and Patents Act 1988 your agency is the first owner of work your employees create in the course of employment, subject to any agreement to the contrary. Read each client contract for what it assigns and when. Freelancers are the gap that catches agencies out. A freelance designer or developer is the first owner of what they create unless they assigned it to you in writing, so a missing assignment means a missing right. Our [vetting standard](https://sellmydatatoai.net/how-we-vet) checks IP assignments from staff and contractors. ## Can you include stock images and fonts? Stock images and fonts usually have to come out. Getty Images' content license agreement, for example, lists machine learning and AI use as a restricted use unless an additional license is bought, and bars use in connection with training, fine-tuning or other data ingestion. Other stock libraries and font foundries set their own terms, so check each one. The practical approach is to strip licensed assets from design files, or replace them with placeholders, and record the method in your [data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest). | Content | Usual rights position | Usual approach | | --- | --- | --- | | Final deliverables | Often assigned to the client | Include only with client agreement | | Unused concepts and working files | Depends on the contract | Check each contract | | Work by employees | Agency owns under CDPA section 11(2) | Include if not assigned to a client | | Work by freelancers | Freelancer owns unless assigned | Include only with a signed assignment | | Stock images | Licensed, often with AI restrictions | Remove or replace | | Fonts | Licensed under the font's terms | Remove or check the license | | User research recordings | Personal data of participants | Exclude | ## What about client confidentiality and personal data? Client confidentiality covers unreleased products, pitch work and anything shared under NDA, and that material should be excluded or cleared with the client. Pitches to prospects who did not appoint you are usually your own work, but they may contain the prospect's confidential briefing material, so remove anything the prospect supplied. User research recordings, interview notes and form submissions held for client sites contain personal data and rarely add enough value to justify the risk. You may decide to tell key clients before a sale rather than after. See [telling staff and clients about a data license](https://sellmydatatoai.net/guides/telling-staff-and-clients-about-a-data-sale) for how to approach it. ## How does selling agency data work with Sell My Data To AI? Selling agency data with Sell My Data To AI starts with a consultation about your file stores, project management tools and client contract types. We help you sort projects into cleared, needs client agreement and excluded, before any buyer is involved. Buyers you approve see a redacted sample under a mutual NDA, and you sign the license directly with your chosen buyer. [Book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller) to find out whether your archive is likely to interest buyers. ## Frequently asked questions ### Can a design agency sell work it has already handed to clients? Only if the agency still holds the rights. Many agency contracts assign copyright in final deliverables to the client on payment, and if yours do, you need the client's agreement. Unused concepts and working files may still be yours, depending on the contract wording, so check each contract before including a project. ### Do stock images in our design files stop a sale? Stock images do not stop a sale, but they usually have to be removed. Getty Images' license agreement lists machine learning and AI use as restricted unless an additional license is bought. Strip or replace licensed images and fonts in the files you deliver, and note the method in your data manifest. ### What if freelancers did some of our design work? Freelance work is owned by the freelancer unless they assigned it to you in writing, under section 11 of the Copyright, Designs and Patents Act 1988. Find the contracts that contain an assignment. Work without one should be left out, or you can ask the freelancer for a written assignment now. ### Is a small agency's archive worth selling? Mercor's referral listing, reported by ddiy.co, covers design and web agencies from 10 employees, with a published range starting at $20,000. A small agency with years of linked briefs, versions and feedback can still be of interest. A consultation tells you honestly whether buyers are likely to want your archive. ## Sources - [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), ddiy.co (figures taken from each buyer's own site), accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 11](https://www.legislation.gov.uk/ukpga/1988/48/section/11), legislation.gov.uk, accessed November 2026 - [Getty Images content license agreement](https://www.gettyimages.com/eula), Getty Images, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How engineering and architecture firms can sell data to AI companies URL: https://sellmydatatoai.net/sell-your-data/engineering-firms Engineering and architecture firms can license design calculations, specifications, design review comments, technical queries and drawings to AI companies. Architects usually keep copyright in their drawings and license it to clients, a stronger starting position than many industries have. Security-sensitive assets, government classified material and client confidentiality are the main reasons to hold records back. ## Key takeaways - AI buyers most want these records from engineering and architecture firms: (1) design calculations and the checks applied to them, (2) specifications and clause libraries and (3) design review comments, RFIs and technical query logs. - No AI buyer publishes a price range for engineering and architecture firms yet, so value depends on volume, years of history, structure and how cleanly the rights can be shown. - Size buyers look for: No engineering minimum is published. micro1 asks for 30 or more full-time staff, and Handshake AI for 20 or more staff and three years of trading. - The first rights issue to clear: Clients usually hold a license to use drawings, while copyright stays with the designer. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Design calculations and the checks applied to them; Specifications and clause libraries; Design review comments, RFIs and technical query logs; Drawings and BIM models where you hold the copyright; Method statements and design risk registers; Lessons-learned reports and project close-out reviews - Why buyers want them: Engineering records show expert reasoning checked against codes and standards, with a trail from query to answer. micro1's brief asks for project histories, process documentation and templates, and an engineering project file holds all three. - Size buyers look for: No engineering minimum is published. micro1 asks for 30 or more full-time staff, and Handshake AI for 20 or more staff and three years of trading. - Rights issues: Clients usually hold a license to use drawings, while copyright stays with the designer; Some appointments assign copyright to the client instead; Security-sensitive assets call for the security-minded approach in ISO 19650-5; Government classified material is handled under the Government Security Classifications Policy; Subconsultants' work and manufacturers' objects inside models belong to them Engineering and architecture firms start from a better rights position than most. Under a RIBA appointment the architect keeps copyright in the drawings and licenses it to the client, RIBA Journal explains, which reflects the position at law without a contract. The obstacles are confidentiality and security, not ownership. ## What engineering and architecture data do AI companies buy? AI companies buy engineering and architecture data that captures expert reasoning with a check attached. A calculation package shows a design decision and its verification. A technical query log shows a question, the reasoning and the resolution. micro1 asks for project histories, process documentation and templates, and says several years of records across multiple systems is generally a stronger fit. A practice with years of project files in a document management system matches that description. | Record | Who usually owns it | Main issue | | --- | --- | --- | | Calculations and design checks | The firm | Client and site identification | | Specifications and clause libraries | The firm, unless built on a licensed specification system | Third-party specification content | | RFIs and technical queries | Shared across the project team | Other parties' contributions | | Drawings and BIM models | The designer, licensed to the client | Security-sensitive assets | | Lessons-learned reports | The firm | Comments on named clients or contractors | ## Who owns the copyright in drawings and models? The designer usually owns the copyright in drawings and models. RIBA Journal states that under a RIBA contract the architect retains copyright and licenses it to the client, and that a client who has paid the fee holds a license to copy the drawings to build on that site. Section 11 of the Copyright, Designs and Patents Act 1988 gives copyright in employees' work to the employer. Check your appointments anyway. Some clients require an assignment of copyright, and novation on design and build projects can change who holds what. Models also contain subconsultants' work and manufacturers' objects that are not yours to license. ## Which engineering projects should you hold back? Engineering projects involving security-sensitive assets should be held back unless the client agrees in writing. [ISO 19650-5](https://www.iso.org/standard/74206.html), published in 2020, sets out a security-minded approach to managing sensitive information in BIM and across the life of an asset. Government work adds a second layer. [GOV.UK guidance for contractors](https://www.gov.uk/government/publications/government-security-classifications/guidance-16-contractors-and-contracting-authorities-html) says suppliers handle HMG classified information under the Government Security Classifications Policy and any extra requirements the contracting authority sets. Exclude anything marked OFFICIAL-SENSITIVE or classified higher. Site photos and progress reports show workers and members of the public. Remove faces and vehicle registrations, or leave the photos out. > **Strip the location** Location identifies an engineering project faster than a name. Remove site addresses, coordinates, grid references and client names from calculations and reports before [de-identification](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data) of the rest. ## How much is engineering data worth to AI buyers? No buyer publishes an engineering or architecture range. The nearest published figure is Mercor's construction company listing of $115,000 to $1 million for firms with 50 to 2,000 staff, recorded by ddiy.co in October 2026, but that applies to contractors, not design firms. micro1's general tiers start at $100,000 to $500,000 for a qualified partnership and rise to $3 million or more for highly unique operational data. If your practice also builds, see our page for [construction companies](https://sellmydatatoai.net/sell-your-data/construction-companies). ## How do you sell engineering data through Sell My Data To AI? We start with a consultation and a rights pre-check against your appointments. Every seller passes our eight-point vetting standard before a buyer sees anything, and every buyer is verified for intended use and onward transfer limits. You approve every buyer and every license term. Read [how to build a data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest), or [book a consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### Do architects own the copyright in their drawings? Architects usually own the copyright in their drawings. RIBA Journal explains that under a RIBA contract the architect retains copyright and licenses it to the client, which mirrors the position without a contract. Some appointments assign copyright to the client, so check each one before including drawings in a license. ### Can we license BIM models to AI companies? You can license BIM models where you hold the copyright and the asset is not security-sensitive. Models contain subconsultants' work and manufacturers' objects that need removing or clearing. For sensitive assets, ISO 19650-5 sets out a security-minded approach, and the client's written agreement should come first. ### Is there a published price for engineering data? No buyer publishes a price for engineering or architecture data. Mercor's construction company listing of $115,000 to $1 million, recorded by ddiy.co in October 2026, applies to contractors rather than design firms. We give you a realistic range for your own records in a [free data valuation](https://sellmydatatoai.net/data-valuation). ### What about defense and government projects? Defense and government projects should normally stay out of any AI data license. GOV.UK guidance says contractors handle HMG classified information under the Government Security Classifications Policy and any extra requirements the contracting authority sets. Ask the contracting authority before considering any reuse of project records. ## Sources - [Architects' copyright licences and infringements on drawings](https://www.ribaj.com/intelligence/legal-copyright-in-architects-drawings-licenses-and-infringement/), RIBA Journal, accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 11: first ownership of copyright](https://www.legislation.gov.uk/ukpga/1988/48/section/11), legislation.gov.uk, accessed November 2026 - [ISO 19650-5:2020 Security-minded approach to information management](https://www.iso.org/standard/74206.html), International Organization for Standardization, accessed November 2026 - [Government Security Classifications Policy, guidance 1.6: contractors and contracting authorities](https://www.gov.uk/government/publications/government-security-classifications/guidance-16-contractors-and-contracting-authorities-html), GOV.UK (Cabinet Office), accessed November 2026 - [Monetize your company's operational data](https://www.micro1.ai/data), micro1, accessed November 2026 - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Sell financial services data to AI companies URL: https://sellmydatatoai.net/sell-your-data/financial-services Financial services firms can license research notes, credit and investment memos, compliance reviews and procedure manuals to AI companies, but we have found no buyer that publishes a price range for the sector. Customer and market data carry the heaviest rules: GLBA notice and opt-out duties in the US, unlawful disclosure of inside information under UK MAR, PCI DSS and tipping off. ## Key takeaways - AI buyers most want these records from financial services firms: (1) investment research notes and memos, (2) credit and underwriting memos, de-identified and (3) compliance monitoring reports and file reviews. - No AI buyer publishes a price range for financial services firms yet, so value depends on volume, years of history, structure and how cleanly the rights can be shown. - Size buyers look for: No sector-wide minimum published. Mercor's listing for investment firms asks for 5 or more staff. - The first rights issue to clear: GLBA privacy notice and opt-out duties before sharing customer information with non-affiliates. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Investment research notes and memos; Credit and underwriting memos, de-identified; Compliance monitoring reports and file reviews; Policy and procedure manuals; Complaint case files, de-identified; Model documentation and validation reports - Why buyers want them: Finance models need examples of analysts and credit officers reasoning from evidence to a decision. Memos, reviews and the comments that challenged them are written by specialists and rarely published. - Size buyers look for: No sector-wide minimum published. Mercor's listing for investment firms asks for 5 or more staff - Rights issues: GLBA privacy notice and opt-out duties before sharing customer information with non-affiliates; Unlawful disclosure of inside information under UK MAR; Suspicious activity report records and tipping off under POCA section 333A; PCI DSS rules on card data in records and recordings; Market data vendor licenses restricting redistribution ## What financial services data do AI companies buy? AI companies buy financial services data that shows specialist reasoning from evidence to a decision. A credit memo that sets out the case, the risk committee's challenge and the final terms is the kind of document no public source contains. - Investment research notes and investment committee memos - Credit and underwriting memos, with customer details removed - Compliance monitoring reports and file review findings - Policy and procedure manuals - Complaint case files, de-identified, with the final decision - Model documentation and validation reports Firm-written manuals, procedures and model documentation are the simplest to clear. Anything built around a named customer or a live market position needs much more work. See [document and form data](https://sellmydatatoai.net/buy-ai-training-data/document-and-form-data) for how buyers use structured financial documents. Model documentation and validation reports are often overlooked. They explain how a model was built, tested and challenged, and they rarely contain customer data, though they may describe proprietary methods you want to keep out. Complaint files are another strong example once de-identified, because they record the issue, the investigation and the final response in order. ## How much do AI companies pay for financial services data? We have found no AI buyer that publishes a price range for financial services data. The buyers compared by ddiy.co in October 2026 give company-wide figures, and its report of Mercor's listings mentions investment firms with a minimum of 5 staff but no payout range. Company-wide figures give some context. micro1 publishes $100,000 to $2 million or more, Handshake AI reports $100,000 to $4 million paid to date and Scale AI gives an illustrative $10,000 to $1 million or more, all as reported by ddiy.co. None of these is a quote for a financial firm, and a [free data valuation](https://sellmydatatoai.net/data-valuation) is the way to get a realistic range. ## Can a US financial institution share customer data under GLBA? A US financial institution can share customer information with non-affiliated third parties only within the Gramm-Leach-Bliley Act's Privacy Rule. The FTC says financial institutions must tell customers about their information-sharing practices and explain their right to opt out if they do not want their information shared with certain third parties. The GLBA Safeguards Rule also requires an information security program with administrative, technical and physical safeguards for customer information. Combined with the privacy notices in force when the data was collected, which Frankfurt Kurnit Klein and Selz says companies must comply with, these rules make customer-level data a poor candidate for licensing. Exclude it or take specialist advice. ## Does inside information in your files stop a sale? Inside information in your files does not stop a sale, but every document containing it must come out. The FCA lists unlawful disclosure of inside information as one of the four civil offenses under UK MAR, alongside insider dealing and market manipulation. Deal files, research on listed issuers before publication and investment committee papers can all contain inside information at the time they were written. The safe approach is to exclude recent material entirely and include older research only after compliance has reviewed it. Record the review date and reviewer for each batch, so you can show the buyer and your regulator how each decision was made. | Record | Rule | Usual approach | | --- | --- | --- | | Customer account and transaction data | GLBA in the US, UK GDPR and customer confidentiality | Exclude | | Deal files and unpublished research | UK MAR unlawful disclosure | Exclude recent material, compliance review for older files | | Suspicious activity reports and related notes | POCA section 333A tipping off | Exclude | | Card data in records or call recordings | PCI DSS Requirement 3.3.1 | Exclude or redact | | Market data from vendor terminals | Vendor license terms | Remove pasted data, keep the analyst's commentary | | Policies, procedures and model documentation | Firm-owned | Include | ## What other financial services rules affect a data sale? Tipping off and card data rules also affect a financial services data sale. Section 333A of the Proceeds of Crime Act 2002 makes it an offense in the regulated sector to disclose a suspicious activity report in a way likely to prejudice an investigation, so all reporting records must be excluded. Firms that record calls should check card data. The PCI Security Standards Council says storing card validation codes in any digital audio recording after authorization breaches PCI DSS Requirement 3.3.1. Market data vendor licenses commonly restrict redistribution, so pasted prices and terminal screenshots should come out of research notes. Nixon Peabody's October 2026 alert also advises buyers to start with a limited sample in a controlled environment, which suits regulated sellers too. ## How does selling financial services data work with Sell My Data To AI? Selling financial services data with Sell My Data To AI starts with a consultation about your document stores, the years covered and your compliance team's view. We help you build an exclusion list and [data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest) your compliance officer can sign off before any buyer is involved. We verify every buyer's identity, intended use, security and retention commitments before they see a sample, see [how we vet](https://sellmydatatoai.net/how-we-vet). You sign the license directly with your chosen buyer. [Book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller) to start. ## Frequently asked questions ### Is there a published price for financial services data? We have found no buyer that publishes a price range for financial services data as of November 2026. ddiy.co's comparison of buyers gives only company-wide figures, such as micro1's $100,000 to $2 million or more. Mercor's listing for investment firms gives a minimum of 5 staff but no range. A valuation based on your own records is the reliable guide. ### Can a bank or lender license customer data for AI training? Customer data is the hardest category for a bank or lender to license. In the US, the FTC says GLBA requires financial institutions to tell customers about information sharing and offer an opt-out for certain third parties. UK firms face UK GDPR and customer confidentiality duties. Most firms exclude customer-level data and license firm-written material instead. ### Can investment research be licensed to AI companies? Older investment research can often be licensed after compliance review, once pasted vendor market data is removed. Recent research and deal files may contain inside information, and the FCA lists unlawful disclosure of inside information as a civil offense under UK MAR. Exclude anything that could still be inside information. ### Do compliance teams need to approve a data sale? Your compliance team should approve the exclusion list before any buyer sees a sample, because the rules on inside information, tipping off and customer data sit with them. We help prepare an exclusion list and data manifest they can review. You approve every buyer, sample and license term, and can withdraw before signing. ## Sources - [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), ddiy.co (figures taken from each buyer's own site), accessed November 2026 - [Gramm-Leach-Bliley Act](https://www.ftc.gov/business-guidance/privacy-security/gramm-leach-bliley-act), Federal Trade Commission, accessed November 2026 - [Market abuse](https://www.fca.org.uk/markets/market-abuse), Financial Conduct Authority, accessed November 2026 - [Proceeds of Crime Act 2002, section 333A: tipping off in the regulated sector](https://www.legislation.gov.uk/ukpga/2002/29/section/333A), legislation.gov.uk, accessed November 2026 - [Are audio/voice recordings permitted to contain sensitive authentication data? (FAQ 1210)](https://www.pcisecuritystandards.org/faqs/1210/), PCI Security Standards Council, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 - [Shopping for AI training data: five diligence questions before you buy](https://www.nixonpeabody.com/insights/alerts/2026/10/26/shopping-for-ai-training-data-five-diligence-questions-before-you-buy), Nixon Peabody, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How insurers and brokers can sell data to AI companies URL: https://sellmydatatoai.net/sell-your-data/insurance Insurers and brokers can license de-identified claims handling notes, underwriting referral rationale, policy wordings they drafted and operating procedures to AI companies. No buyer publishes an insurance-specific price, so value depends on volume, years covered and how cleanly health and fraud data can be removed. Settle the special category data question first, because the insurance condition in UK law does not cover licensing. ## Key takeaways - AI buyers most want these records from insurance companies and brokers: (1) claims handling notes and decision letters, de-identified, (2) underwriting referral notes explaining why risks were accepted, loaded or declined and (3) policy wordings, endorsements and clause libraries you drafted. - No AI buyer publishes a price range for insurance companies and brokers yet, so value depends on volume, years of history, structure and how cleanly the rights can be shown. - Size buyers look for: No insurance-specific minimum is published. micro1 asks for 30 or more full-time staff, and Handshake AI for 20 or more staff and three years of trading. - The first rights issue to clear: Health data in claims and underwriting files is special category data under UK GDPR Article 9. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Claims handling notes and decision letters, de-identified; Underwriting referral notes explaining why risks were accepted, loaded or declined; Policy wordings, endorsements and clause libraries you drafted; Underwriting guides, claims playbooks and procedure manuals; Complaint files and final response letters; Broker and insurer placement correspondence, with client names removed - Why buyers want them: Buyers want records of expert judgment applied to real cases. micro1's published brief asks for SOPs, process documentation, knowledge bases, templates and support tickets, and an insurance file holds all of them. - Size buyers look for: No insurance-specific minimum is published. micro1 asks for 30 or more full-time staff, and Handshake AI for 20 or more staff and three years of trading. - Rights issues: Health data in claims and underwriting files is special category data under UK GDPR Article 9; Fraud allegations and convictions are criminal offense data with their own rules; The Data Protection Act 2018 insurance condition covers insurance purposes, not licensing; Terms of business between brokers and insurers may limit who can reuse client files; US insurers fall under Gramm-Leach-Bliley privacy rules enforced by state regulators Insurance files record decisions and the reasons behind them. A claims handler's note explains why a claim was paid, reduced or declined, and an underwriter's referral note explains why a risk was accepted with a loading. That is the kind of material AI buyers say they want. micro1, which announced in October 2026 that it would spend $1 billion over 12 months on enterprise operational data, says it is looking for data that shows how work is actually performed within an organization. ## What insurance data do AI companies buy? AI companies buy insurance data that shows decisions being made on real cases. micro1's published brief lists SOPs, process documentation, knowledge bases, workflows, templates, CRM records, support tickets, project histories and internal communications, and names finance, legal and contracts, and customer support among the functions it wants. Structured policy data on its own is close to a customer list. The value sits in free text: the notes, letters and referral comments where a person explains a judgment. | Record | What it shows a model | Main issue to clear | | --- | --- | --- | | Claims notes and decision letters | How policy terms are applied to facts | Health and fraud data | | Underwriting referral notes | How risk is assessed and priced | Client identity and commercial terms | | Policy wordings and clause libraries | How cover is drafted | Wordings based on market or third-party forms | | Underwriting guides and claims playbooks | How work is organized | Usually low if written in-house | | Complaint files and final responses | How disputes are resolved | Complainant identity and health details | ## How much is insurance data worth to AI buyers? Insurance data has no published price. None of the buyers whose terms are public gives an insurance-specific range, so any figure quoted for an insurance archive is a negotiation, not a market rate. General ranges give a frame. micro1 publishes tiers of $100,000 to $500,000 for qualified partnerships, $500,000 to $3 million for large-scale or ongoing datasets, and $3 million or more for highly unique operational data. Mercor's published industry ranges run from $20,000 to $1.5 million, according to ddiy.co's review of buyer sites in October 2026. Volume, years covered and the share of free-text reasoning that survives de-identification decide where an archive lands. Our guide to [what your data is worth to AI companies](https://sellmydatatoai.net/guides/how-much-is-your-data-worth-to-ai) sets out the factors, and a [free data valuation](https://sellmydatatoai.net/data-valuation) applies them to your records. ## What rights issues do insurers and brokers need to clear? The main rights issue for insurers and brokers is sensitive personal data. Claims and underwriting files routinely hold health information, and the ICO lists health data as special category data under UK GDPR Article 9. ### Health and fraud data Fraud files raise a second category. The ICO says personal data about criminal allegations, proceedings or convictions is not special category data but carries similar rules and safeguards of its own. Counter-fraud notes, investigator reports and conviction checks fall into it. Daniel Goldberg of Frankfurt Kurnit Klein and Selz advises sellers to exclude high-risk data entirely rather than rely on de-identification, and names health information among the categories to leave out. For most insurance archives, that means removing medical reports and investigator files before [de-identification](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data) of the rest. ### Why the insurance condition does not help Schedule 1, paragraph 20 of the [Data Protection Act 2018](https://www.legislation.gov.uk/ukpga/2018/12/schedule/1/paragraph/20) allows processing of health and other sensitive data where it is necessary for an insurance purpose. The Act defines that as advising on, arranging, underwriting or administering a contract, handling a claim, or exercising rights and obligations connected with it. Licensing claims files to an AI company is none of those. Reuse for a new purpose is also tested under UK GDPR Article 8A, inserted by the Data (Use and Access) Act 2025 and fully in force from February 5, 2026. It weighs the link between the purposes, the context of collection, whether special category or criminal offense data is involved and the possible consequences for policyholders. ### Broker and insurer contracts Brokers hold client files under terms of business agreements with insurers and engagement terms with clients. Read both before you treat a placement file as yours to license. Our guide on [whether you can legally sell company data](https://sellmydatatoai.net/guides/can-you-legally-sell-company-data-for-ai) covers the wider ownership test. US insurers face another layer. The Gramm-Leach-Bliley Act's privacy provisions apply to insurers, and the US Government Accountability Office reported in 2002 that the Act requires state insurance authorities to enforce them through their own regulations. > **Rule of thumb** If a record would still identify a policyholder, claimant or witness once names are removed, or if it contains a medical report, leave it out. Procedures, wordings and anonymized reasoning carry far less risk and match what buyers' published briefs describe. ## How do you sell insurance data through Sell My Data To AI? You sell insurance data through Sell My Data To AI by starting with a consultation, not a data transfer. We do not take custody of your raw data. Buyers see redacted samples under a mutual NDA, and full data moves only after you sign a license directly with the buyer. Before any buyer sees a sample, your records pass our eight-point vetting standard, covering authority to sign, ownership, contract restrictions, privacy notices, third-party content, de-identification, sample quality and [manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest) accuracy. You approve every buyer, every sample and every license term. [Book a consultation](https://sellmydatatoai.net/book-a-consultation?side=seller) to find out whether your files are in demand. ## Frequently asked questions ### Can an insurer sell claims data to AI companies? An insurer can license claims data to AI companies only where it holds the right to do so and personal data is excluded or properly de-identified. Health and fraud information should normally come out entirely. The Data Protection Act 2018 insurance condition covers insurance purposes, not licensing, so take independent legal advice before relying on any other basis. ### Do brokers own their client files? Brokers hold client files under terms of business agreements with insurers and engagement terms with clients, and those contracts decide what reuse is allowed. Procedure manuals and templates your own staff wrote are a safer start. Under section 11 of the Copyright, Designs and Patents Act 1988, your company owns copyright in work employees create in the course of employment. ### Is there a published price for insurance data? No buyer whose terms are public publishes an insurance-specific price. micro1's general tiers start at $100,000 to $500,000 for qualified partnerships, and Mercor's industry ranges run from $20,000 to $1.5 million according to ddiy.co. An archive's value depends on volume, years covered and how much reasoning survives de-identification. ### Will policyholders find out their data was licensed? Plan as if policyholders will find out. The privacy notice in force when you collected their data governs what you may do with it, and in the US the FTC has warned that quietly adopting more permissive practices may be unfair or deceptive. Our guide to [telling staff and clients](https://sellmydatatoai.net/guides/telling-staff-and-clients-about-a-data-sale) covers how to handle it. ## Sources - [What is special category data?](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/lawful-basis/special-category-data/what-is-special-category-data/), Information Commissioner's Office, accessed November 2026 - [Data Protection Act 2018, Schedule 1, paragraph 20: insurance](https://www.legislation.gov.uk/ukpga/2018/12/schedule/1/paragraph/20), legislation.gov.uk, accessed November 2026 - [UK GDPR Article 8A: purpose limitation, further processing](https://www.legislation.gov.uk/eur/2016/679/article/8A), legislation.gov.uk, accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 11: first ownership of copyright](https://www.legislation.gov.uk/ukpga/1988/48/section/11), legislation.gov.uk, accessed November 2026 - [Monetize your company's operational data](https://www.micro1.ai/data), micro1, accessed November 2026 - [micro1 announces commitment to spend $1B on enterprise data acquisition over the next 12 months](https://www.micro1.ai/blog/micro1-commits-1b-to-enterprise-data-acquisition), micro1, accessed November 2026 - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz (Daniel Goldberg), accessed November 2026 - [Financial privacy: status of state actions on Gramm-Leach-Bliley Act's privacy provisions (GAO-02-361)](https://www.gao.gov/products/gao-02-361), US Government Accountability Office, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Sell law firm data to AI companies URL: https://sellmydatatoai.net/sell-your-data/law-firms Law firms can license precedents, markups, research memos and de-identified matter files to AI companies, and Mercor publishes a range of $50,000 to $1.2 million per firm. The limit is confidentiality: the SRA Codes require client affairs to stay confidential unless the client consents or the law permits disclosure, so firm-created know-how is the easiest material to clear. ## Key takeaways - AI buyers most want these records from law firms: (1) precedent and template banks with drafting notes, (2) contract markups with full redline history and (3) research memos and know-how notes. - Published buyer range for law firms: $50,000 to $1.2 million per company. - Source: Mercor's published referral listing for law firms, as reported by ddiy.co, checked November 2026. - Size buyers look for: 5 to 300 lawyers and staff (Mercor's listing). - The first rights issue to clear: SRA Codes paragraph 6.3: client affairs stay confidential, including former and deceased clients. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Precedent and template banks with drafting notes; Contract markups with full redline history; Research memos and know-how notes; Advice letters, pleadings and skeleton arguments, de-identified; Matter checklists and file opening procedures - Why buyers want them: Legal AI tools are trained and tested on how lawyers draft, revise and reason. Documents that show expert work in progress, with the changes and the reasons for them, teach a model more than finished documents. - Published buyer range: $50,000 to $1.2 million per company (Mercor's published referral listing for law firms, as reported by ddiy.co, checked November 2026) - Size buyers look for: 5 to 300 lawyers and staff (Mercor's listing) - Rights issues: SRA Codes paragraph 6.3: client affairs stay confidential, including former and deceased clients; CPR 31.22 limits use of documents disclosed by other parties to the proceedings; Suspicious activity report records, where disclosure can be tipping off under POCA section 333A; Legal professional privilege belongs to the client, not the firm; Counterparty documents received under NDA ## What law firm data do AI companies buy? AI companies buy law firm data that shows how lawyers actually draft, revise and reason. A signed contract on its own is worth less than the same contract with three rounds of tracked changes, the partner's comments and the file note explaining why a clause was rejected. Legal AI products are built to draft, review and answer questions the way a trained lawyer would. Training and testing them needs real examples of expert work, which only practicing lawyers produce. Mercor's listing covers firms of 5 to 300 lawyers and staff, so mid-sized and regional practices are in scope. - Precedents and template banks, with the drafting notes that explain each option - Contract markups that keep the full redline history between the parties - Research memos and know-how notes written by fee earners - Advice letters, pleadings and skeleton arguments, once de-identified - Matter checklists, file opening procedures and supervision notes Firm-created know-how is usually the cleanest part of the archive because the firm wrote it and owns it. Client files are more valuable to buyers and much harder to clear, as the sections below explain. See [expert domain data](https://sellmydatatoai.net/buy-ai-training-data/expert-domain-data) for how buyers use this kind of material. ## How much do AI companies pay for law firm data? AI companies publish a range of $50,000 to $1.2 million for law firm data. That is Mercor's referral listing for firms of 5 to 300 lawyers and staff, as reported by ddiy.co in October 2026. Treat a published range as a ceiling for strong archives, not a quote. AIData.Broker's analysis, reported by ddiy.co, puts the median of buyers' published offers at $280,000 per company but the median of public sales at $31,600. The same analysis finds an exclusive license fetches about 2.5 to 5 times a non-exclusive deal, which is why the [exclusivity decision](https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses) moves the price so much. What moves a law firm toward the top of a range is volume across many years, documents with revision history, a clear mix of practice areas and a rights position that has been checked. A [free data valuation](https://sellmydatatoai.net/data-valuation) gives you a written range before you commit to anything. ## Can a law firm sell client files without breaching confidentiality? A law firm can only license material that reveals a client's affairs if the client consents or the law permits it. Paragraph 6.3 of the SRA Codes of Conduct requires solicitors and firms to keep the affairs of current and former clients confidential unless disclosure is required or permitted by law or the client consents. The SRA's guidance on confidentiality of client information, last updated in June 2022, adds points that matter for a data license. The duty continues after the retainer ends and after the client dies, when it passes to their personal representatives. Consent must be clear about who receives the information, when and for what purpose. The SRA guidance does not address anonymized material. Whether a de-identified file still reveals a client's affairs is a judgment for your compliance officer for legal practice and your own professional conduct adviser. US firms face equivalent duties under their state's rules of professional conduct and should take local advice. ## What should a law firm leave out of a data sale? A law firm should leave out any material where a rule outside the firm's control prevents disclosure, whatever the client consent position. The table sets out the usual categories. | Material | Why it is a problem | Usual approach | | --- | --- | --- | | Documents disclosed by other parties in litigation | CPR 31.22 limits their use to the proceedings unless read in open court, permitted by the court or agreed by the disclosing party and owner | Exclude | | Suspicious activity report records and related file notes | POCA section 333A makes disclosure likely to prejudice an investigation an offense | Exclude | | Privileged advice for a named client | Privilege belongs to the client and the SRA duty of confidentiality applies | Exclude, or de-identify with client consent | | Counterparty documents received under NDA | Other companies' confidential information under contract | Exclude or seek consent | | HR, partner and payroll records | Personal data with little training value | Exclude | | Firm precedents and know-how | Firm-owned, low client risk once names are removed | Include | Frankfurt Kurnit Klein and Selz advises excluding high-risk data entirely rather than relying on de-identification, because once a model is trained the data cannot practically be pulled back out. For a law firm, that is a sensible default for crime, regulatory investigation and family files. ## How does de-identification work for legal documents? De-identification for legal documents means removing far more than client names. A hearing date, a court, an unusual asset and a town can identify a party as reliably as a name. - Case numbers, court names and hearing dates - Company registration numbers and registered addresses - Distinctive facts, such as a rare asset, a reported incident or a named property - Opposing counsel, expert witness and judge names - Signature blocks, email footers and document metadata Buyers will ask how you de-identified the files and whether the method was tested. Our [vetting standard](https://sellmydatatoai.net/how-we-vet) requires a documented method and a re-identification test on a sample. The guide to [de-identification for AI training data](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data) covers the methods in detail. ## How does selling law firm data work with Sell My Data To AI? Selling law firm data with Sell My Data To AI starts with a 30-minute consultation about what the firm holds and where it sits, usually a document management system, email and a practice management system. No records leave your firm at this stage. If buyers are likely to be interested, we help you build a [data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest) and an exclusion list, run our eight vetting checks and show a redacted, randomly drawn sample only to buyers you approve. You sign the license directly with the buyer after your own lawyer has reviewed it, and our success fee is agreed in writing before any introduction. [Book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller) to find out whether your archive is likely to interest buyers. ## Frequently asked questions ### Do AI companies want data from small law firms? Small law firms can be in scope. Mercor's referral listing, reported by ddiy.co, covers law firms from 5 to 300 lawyers and staff. Headcount matters less than the years of records held, the practice areas covered and how much of the archive is firm-created know-how that can be cleared without client consent. ### Does a client have to consent before their file is used to train AI? Under paragraph 6.3 of the SRA Codes of Conduct, client affairs stay confidential unless the client consents or the law requires or permits disclosure. The SRA guidance does not address anonymized material, so firms either seek consent or take advice on whether their de-identified files still reveal a client's affairs. Take independent advice before signing any license. ### Can a law firm license its precedent bank on its own? A precedent bank is often the easiest part of a law firm archive to license, because the firm created it and owns the copyright. Buyers value precedents more when drafting notes and usage guidance come with them. Check whether any precedents came from a subscription know-how service, because that material is licensed to you under the service's terms, so read them before including it. ### Will the buyer know which firm the data came from? The buyer you sign with will know your firm, because you contract with it directly. Other buyers will not: we never send your firm's name or any record to a buyer you have not approved. You can also ask for license terms that stop the buyer naming your firm publicly or in marketing. ### How long does a law firm data deal take? A law firm data deal usually takes longer than the buyer's own process, because exclusions and any client consents come first. Buyers' published timelines give a guide once a package is ready: Handshake AI says its process takes 7 to 14 business days, and Mercor says it pays 2 to 4 weeks after signing, as reported by ddiy.co. ## Sources - [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), ddiy.co (figures taken from each buyer's own site), accessed November 2026 - [How much is company data worth to AI companies](https://ddiy.co/how-much-is-company-data-worth-to-ai), ddiy.co (citing AIData.Broker), accessed November 2026 - [Confidentiality of client information](https://www.sra.org.uk/solicitors/guidance/confidentiality-client-information/), Solicitors Regulation Authority, accessed November 2026 - [Civil Procedure Rules 1998, rule 31.22: subsequent use of disclosed documents](https://www.legislation.gov.uk/uksi/1998/3132/rule/31.22), legislation.gov.uk, accessed November 2026 - [Proceeds of Crime Act 2002, section 333A: tipping off in the regulated sector](https://www.legislation.gov.uk/ukpga/2002/29/section/333A), legislation.gov.uk, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How logistics and freight companies can sell data to AI companies URL: https://sellmydatatoai.net/sell-your-data/logistics Logistics and freight companies can license dispatch decisions, exception handling records, customer service threads and operating procedures to AI companies. micro1 names operations and logistics among the functions it buys data from, though no buyer publishes a logistics price range. Driver location and tachograph data is personal data, so we recommend excluding it and licensing the workflow records instead. ## Key takeaways - AI buyers most want these records from logistics and freight companies: (1) dispatch, routing and load planning decisions with the reasons recorded, (2) exception records for delays, damage, shortages and refused deliveries and (3) customer service tickets and email threads with shippers and consignees. - No AI buyer publishes a price range for logistics and freight companies yet, so value depends on volume, years of history, structure and how cleanly the rights can be shown. - Size buyers look for: No logistics-specific minimum is published. micro1 asks for 30 or more full-time staff and primarily English documentation. - The first rights issue to clear: Vehicle tracking produces driver location data, which UK GDPR names as an identifier. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Dispatch, routing and load planning decisions with the reasons recorded; Exception records for delays, damage, shortages and refused deliveries; Customer service tickets and email threads with shippers and consignees; Freight quotes, tender responses and rate negotiation notes; Warehouse and depot procedures, safe systems of work and SOPs; Carrier and subcontractor onboarding and performance records - Why buyers want them: micro1's published brief names operations and logistics as a target function and asks for workflows, SOPs, support tickets and project histories. micro1 says it uses de-identified company data to build reinforcement learning environments. - Size buyers look for: No logistics-specific minimum is published. micro1 asks for 30 or more full-time staff and primarily English documentation. - Rights issues: Vehicle tracking produces driver location data, which UK GDPR names as an identifier; Tachograph records exist for drivers' hours enforcement, and reuse needs a compatible purpose; Dashcam and yard CCTV footage shows drivers, the public and number plates; Shipper contracts may make rates, volumes and consignee details confidential; Customs and trade documents reveal clients' suppliers, values and routes A logistics business makes a recorded decision every time a load is planned, rerouted or held. micro1, which said in October 2026 it would spend $1 billion over 12 months on enterprise operational data, lists operations and logistics as one of the functions it wants data from. ## What logistics data do AI companies buy? AI companies buy logistics data that shows how work gets done: who decided what, with what information, and what happened next. micro1's published brief asks for SOPs, process documentation, workflows, templates, support tickets, project histories and internal communications. Exception handling and customer service threads fit that brief closely. A delayed or damaged consignment produces a chain of messages, a decision and an outcome, which is the structure an [agent trajectory](https://sellmydatatoai.net/glossary/agent-trajectory) records. Depot procedures and safe systems of work add the rules the decisions were made under. ## Which logistics records should you hold back? Logistics sellers should hold back driver-level location and hours data unless it can be fully de-identified. [UK GDPR Article 4(1)](https://www.legislation.gov.uk/eur/2016/679/article/4) names location data as an identifier, so a tracked vehicle linked to a rota is personal data about the driver. | Record | Typical risk | Our usual recommendation | | --- | --- | --- | | SOPs, depot procedures, safe systems of work | Low if written in-house | License | | Customer service tickets and exception threads | Names, addresses and client terms | De-identify, then license | | Quotes, tenders and rate notes | Client and carrier confidentiality | Check contracts, remove rates if needed | | Telematics and GPS traces | Driver location data | Exclude or aggregate | | Tachograph downloads | Driver hours kept for enforcement | Exclude | | Dashcam and CCTV footage | Drivers, the public, number plates | Exclude unless fully cleared | ## What rights issues do logistics companies need to clear? ### Tachograph and telematics data Tachograph data is collected to meet drivers' hours rules. [GOV.UK guidance](https://www.gov.uk/guidance/drivers-hours-goods-vehicles/5-responsibilities-of-vehicle-operators) says operators must download vehicle units at least every 90 days and driver cards at least every 28 days, and be able to produce records to enforcement officers for 12 months. Licensing that data to an AI company is a new purpose. UK GDPR Article 8A, fully in force since February 5, 2026, weighs the link between the original and new purposes, the context of collection including the relationship between employer and driver, and the possible consequences for drivers. Data gathered to monitor workers under a legal obligation is hard to fit to those tests, so we recommend leaving it out. ### Customer and subcontractor contracts Shipper contracts may treat rates, volumes and consignee details as confidential. Daniel Goldberg of Frankfurt Kurnit Klein and Selz points out that email and chat contain other companies' confidential information, often covered by NDAs and commercial agreements, and holding a record does not give you the right to sell it. Customs entries and commercial invoices reveal who your clients buy from, what they pay and where goods move. Treat them as client confidential unless your terms say otherwise. ### Footage Dashcam and yard CCTV records drivers, members of the public and vehicle registrations. The ICO treats facial images as biometric data only where they are used to identify someone, but the footage is still personal data, and clearing it frame by frame is slow. > **Start with the paperwork, not the vehicles** Procedures, exception threads and customer service records carry the operational knowledge. Vehicle, driver and footage data carry most of the privacy risk. Separate the two before you build a [data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest). ## How much is logistics data worth to AI buyers? Logistics data has no published industry range. micro1's general tiers start at $100,000 to $500,000 for a qualified partnership and rise to $3 million or more for highly unique operational data. It prioritizes the US, then other Western markets, and says demand is strongest in the US, UK and Canada. micro1 also says several years of records across multiple systems is generally a stronger fit. A forwarder with a long run of ticket history in one platform and procedures in another matches that description better than a new operator with a single system. ## How do you sell logistics data through Sell My Data To AI? You start with a 30-minute consultation covering likely buyer interest, a realistic range and the rights issues to clear first. Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, agreed in writing before we introduce you to any buyer. Every buyer is verified for company identity, intended use, security and retention commitments and onward transfer limits before seeing a sample. [Book a consultation](https://sellmydatatoai.net/book-a-consultation?side=seller), or read [how to sell data to AI companies](https://sellmydatatoai.net/guides/how-to-sell-data-to-ai-companies) first. ## Frequently asked questions ### Can we sell our telematics data to AI companies? You can only license telematics data safely once it no longer identifies drivers. UK GDPR names location data as an identifier, and data gathered to monitor staff is hard to justify for a new commercial purpose under Article 8A. Aggregated or fully de-identified route data is a different question, but take independent legal advice first. ### Do our shipper contracts stop us licensing data? Shipper contracts may stop you licensing data that reveals rates, volumes or consignee details, so read them before preparing anything. Your own procedures and anonymized exception records are less likely to be caught. Our [licensing agreement checklist](https://sellmydatatoai.net/guides/ai-data-licensing-agreement-checklist) covers the clauses buyers will ask about. ### Does micro1 buy data from UK logistics firms? micro1 names operations and logistics as a target function and says demand is strongest in the US, UK and Canada, with the US prioritized. It asks for 30 or more full-time staff and primarily English documentation. Our profile of [micro1](https://sellmydatatoai.net/data-buyers/micro1) sets out its published terms. ### How long do we have to keep tachograph data? GOV.UK guidance says operators must be able to produce tachograph records to enforcement officers for 12 months. That duty exists for enforcement, not licensing, and it is one reason we recommend excluding tachograph downloads and driver card data from any AI data package. ## Sources - [Monetize your company's operational data](https://www.micro1.ai/data), micro1, accessed November 2026 - [micro1 announces commitment to spend $1B on enterprise data acquisition over the next 12 months](https://www.micro1.ai/blog/micro1-commits-1b-to-enterprise-data-acquisition), micro1, accessed November 2026 - [UK GDPR Article 4: definitions](https://www.legislation.gov.uk/eur/2016/679/article/4), legislation.gov.uk, accessed November 2026 - [UK GDPR Article 8A: purpose limitation, further processing](https://www.legislation.gov.uk/eur/2016/679/article/8A), legislation.gov.uk, accessed November 2026 - [Drivers' hours: goods vehicles, responsibilities of vehicle operators](https://www.gov.uk/guidance/drivers-hours-goods-vehicles/5-responsibilities-of-vehicle-operators), GOV.UK (DVSA), accessed November 2026 - [What is special category data?](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/lawful-basis/special-category-data/what-is-special-category-data/), Information Commissioner's Office, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz (Daniel Goldberg), accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Sell manufacturing data to AI companies URL: https://sellmydatatoai.net/sell-your-data/manufacturers Manufacturers can license quality records, maintenance logs, work instructions, sensor data and inspection images to AI companies, and Mercor publishes a range of $100,000 to $1.5 million per company, the highest ceiling in its reported listings. Export controls apply if any records contain controlled technology, and customer drawings made under NDA usually cannot be included. ## Key takeaways - AI buyers most want these records from manufacturers: (1) non-conformance reports, CAPA and 8D investigations, (2) maintenance work orders and fault logs and (3) work instructions and standard operating procedures. - Published buyer range for manufacturers: $100,000 to $1.5 million per company. - Source: Mercor's published referral listing for manufacturers, as reported by ddiy.co, checked November 2026. - Size buyers look for: 20 to 2,000 employees (Mercor's listing). - The first rights issue to clear: UK export controls: sending controlled technology overseas by email or cloud needs a license. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Non-conformance reports, CAPA and 8D investigations; Maintenance work orders and fault logs; Work instructions and standard operating procedures; Machine and sensor time series with event labels; Inspection images with pass or fail decisions; Engineering change notices - Why buyers want them: Industrial AI needs examples of faults, root causes and fixes from real production lines. Labeled quality and maintenance records are hard to synthesize and few companies outside manufacturing hold them. - Published buyer range: $100,000 to $1.5 million per company (Mercor's published referral listing for manufacturers, as reported by ddiy.co, checked November 2026) - Size buyers look for: 20 to 2,000 employees (Mercor's listing) - Rights issues: UK export controls: sending controlled technology overseas by email or cloud needs a license; US deemed export rules under EAR section 734.13(b); Customer-owned drawings and specifications held under NDA; Trade secrets in process parameters and recipes; Equipment vendor terms on machine telemetry ## What manufacturing data do AI companies buy? AI companies buy manufacturing data that links a fault to its cause and its fix. A non-conformance report, the 8D investigation that followed, the engineering change notice and the updated work instruction together show how a real problem was solved on a real line. - Non-conformance reports, corrective and preventive actions and 8D investigations - Maintenance work orders and fault logs from the maintenance management system - Work instructions and standard operating procedures, with revision history - Machine and sensor time series, especially where faults are labeled - Inspection images with the pass or fail decision recorded - Engineering change notices and the reasons given for them Labels are the difference between raw data and training data. A year of vibration readings is far more useful when the maintenance log shows the date a bearing failed. Inspection photos are [image data](https://sellmydatatoai.net/buy-ai-training-data/image-datasets) that buyers value when each image carries the inspector's decision. Structure matters more than file format. Exports from your maintenance and quality systems as CSV files or database tables, with consistent asset IDs and timestamps, let a buyer link a fault in one system to a sensor trace in another. Record those links, and which system each export came from, in your [data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest). ## How much do AI companies pay for manufacturing data? AI companies publish a range of $100,000 to $1.5 million for manufacturing data. That is Mercor's referral listing for manufacturers of 20 to 2,000 employees, as reported by ddiy.co in October 2026. The $1.5 million ceiling is the highest of the 11 Mercor listings ddiy.co reports, above law firms at $1.2 million and construction companies at $1 million. The 20-employee floor also means smaller specialist manufacturers are in scope. A [free data valuation](https://sellmydatatoai.net/data-valuation) gives you a written range for your own records. ## Do export controls apply to selling manufacturing data? Export controls apply to selling manufacturing data whenever the records contain controlled technology and the buyer, or anyone who accesses the files, is outside the UK. The Export Control Joint Unit's guidance on GOV.UK says emails containing controlled technology require an export license, and so does technology uploaded to the cloud and downloaded or accessed overseas. The same guidance says responsibility for compliance lies with the owner of the technology, not the service provider. For a data sale that means you, not the buyer or the broker. Military items are the obvious case, but dual-use control lists cover many civilian products, so check the lists rather than assuming you are out of scope. US rules add a second layer. The Bureau of Industry and Security treats a release of controlled technology to a foreign person inside the United States as a deemed export under section 734.13(b) of the Export Administration Regulations. If you or your buyer operate in the US, take export control advice before any technical records leave your systems. ## Can you include customer drawings and specifications? Customer drawings and specifications usually cannot be included. If you make parts to a customer's design, the drawings, tolerances and specifications are the customer's property and are normally supplied under an NDA or the purchase terms. Your own records about making those parts may still be usable once the customer's design details are removed. A non-conformance report can keep the defect type, cause and corrective action while dropping the part number, drawing reference and customer name. Treat this as a judgment for each customer contract, and see the [legal guide](https://sellmydatatoai.net/guides/can-you-legally-sell-company-data-for-ai) for the general position. ## What trade secrets should a manufacturer protect? A manufacturer should protect the process knowledge that gives it an edge, even if it owns that knowledge outright. Once a model is trained on records, Frankfurt Kurnit Klein and Selz notes, the data cannot practically be pulled back out. | Record | Competitive sensitivity | Usual approach | | --- | --- | --- | | Process parameters, recipes and settings | High | Exclude or license exclusively with tight use limits | | Quality investigations and corrective actions | Medium | Include with part and customer details removed | | Maintenance work orders and fault logs | Low to medium | Include | | Generic work instructions and safety procedures | Low | Include | | Customer drawings and specifications | Owned by the customer | Exclude | If you do include sensitive quality records, an exclusive license with one named buyer, a narrow field of use and a ban on onward transfer gives you more control than a non-exclusive deal sold to several buyers. See [exclusive vs non-exclusive data licenses](https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses). Check the terms of any connected equipment platform too. Some machine builders host telemetry on their own platforms, and the platform terms decide whether you can export and license that data. ## How does selling manufacturing data work with Sell My Data To AI? Selling manufacturing data with Sell My Data To AI starts with a consultation about your quality, maintenance and production systems and the years they cover. We flag export control and customer contract questions early, so they are cleared or excluded before a buyer sees anything. If any product sits close to a control list entry, allow time for a classification check. We then help you complete the manifest, run our eight [vetting checks](https://sellmydatatoai.net/how-we-vet) and show a redacted sample to buyers you approve. You sign the license directly with your chosen buyer. [Book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller) to start. ## Frequently asked questions ### Is sensor data from our machines worth selling? Sensor data is worth more when it is labeled with what happened, such as a fault, a stoppage or a quality reject recorded in your maintenance or quality system. Unlabeled readings are harder for buyers to use. Check the equipment platform's terms first, because some machine builders control how telemetry hosted on their systems can be exported. ### Who is responsible for export control compliance in a data sale? The owner of the technology is responsible. The Export Control Joint Unit's GOV.UK guidance says responsibility for compliance lies exclusively with the owner of the technology, not with the service provider. If your records contain controlled technology, you need to check whether a license is required before sending them to a buyer whose staff are overseas. ### Can we license data without revealing our processes to competitors? You can limit exposure by excluding process parameters and recipes, removing customer and part details, and agreeing license terms on onward transfer. Our buyer checks cover intended use and onward transfer limits before any sample is shared. You approve every buyer, so you can refuse any buyer connected to a competitor. ### What size of manufacturer do buyers look for? Mercor's referral listing, reported by ddiy.co, covers manufacturers of 20 to 2,000 employees, with a published range of $100,000 to $1.5 million. Other buyers set company-wide minimums instead, such as micro1's 30 or more full-time staff. Years of linked quality and maintenance records matter more than headcount alone. ## Sources - [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), ddiy.co (figures taken from each buyer's own site), accessed November 2026 - [Exporting military or dual-use technology: definitions and scope](https://www.gov.uk/government/publications/exporting-military-or-dual-use-technology-definitions/export-of-technology-remote-access-and-the-use-of-cloud-computing-services), Export Control Joint Unit, GOV.UK, accessed November 2026 - [What is a deemed export?](https://www.bis.gov/learn-support/deemed-exports/what-deemed-export), Bureau of Industry and Security, US Department of Commerce, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Sell marketing and SEO agency data to AI companies URL: https://sellmydatatoai.net/sell-your-data/marketing-agencies Marketing and SEO agencies can license strategy documents, edited content drafts, audits and annotated performance reports to AI companies, but we have found no buyer that publishes a price range for marketing agencies. Most campaign and customer data belongs to clients: under UK GDPR an agency acting as processor may use it only on the client's instructions. ## Key takeaways - AI buyers most want these records from marketing and seo agencies: (1) strategy documents and campaign briefs, (2) content drafts with editor changes and (3) SEO audits and technical recommendations. - No AI buyer publishes a price range for marketing and seo agencies yet, so value depends on volume, years of history, structure and how cleanly the rights can be shown. - Size buyers look for: No industry-specific minimum published. micro1 asks for 30 or more full-time staff across all industries. - The first rights issue to clear: Agency acting as processor for client customer data under UK GDPR Article 28. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Strategy documents and campaign briefs; Content drafts with editor changes; SEO audits and technical recommendations; Keyword research with intent mapping; Performance reports with analyst commentary; Ad copy variants with test results - Why buyers want them: Published marketing content is public and has largely been crawled already. Buyers want the reasoning behind it: how a strategist turned a brief and data into decisions, and how an editor changed a draft. - Size buyers look for: No industry-specific minimum published. micro1 asks for 30 or more full-time staff across all industries - Rights issues: Agency acting as processor for client customer data under UK GDPR Article 28; Client contracts on deliverables, confidentiality and performance figures; Exports from SEO, analytics and ad platforms under their own terms; Creator and influencer content licensed for campaigns only; Agency prospect lists collected for marketing, not AI training ## What marketing agency data do AI companies buy? AI companies buy marketing agency data that shows specialist reasoning, not finished output. A published blog post or ad is public. The strategy document that chose its angle, the editor's tracked changes and the report explaining why it outperformed are not. - Strategy documents and campaign briefs, with the research behind them - Content drafts with tracked editorial changes and comments - SEO audits and technical recommendations, with what was implemented - Keyword research with search intent mapping - Performance reports with the analyst's commentary - Ad copy variants alongside test results Paired records carry the most value. An original draft and the edited version, or two ad variants and the result of the test between them, are close to the [preference data](https://sellmydatatoai.net/glossary/preference-data) labs use to tune models. See [RLHF and preference data](https://sellmydatatoai.net/buy-ai-training-data/rlhf-preference-data) for how buyers use pairs like these. An SEO audit shows how this works. An audit that lists the issues found, the fixes recommended, which fixes were implemented and what changed afterward is a complete worked example. Keep the before and after evidence with the audit and redact the client's domain, so the chain survives de-identification. ## How much do AI companies pay for marketing agency data? We have found no AI buyer that publishes a price range for marketing agency data. ddiy.co's October 2026 comparison reports 11 of Mercor's 24 industry listings, and marketing agencies are not among them. The other buyers it compares publish company-wide figures only. The closest published comparison is Mercor's listing for design and web agencies at $20,000 to $890,000, but that is a different sector and not a guide to your price. Company-wide figures are wider still: AIData.Broker, reported by ddiy.co, puts the median of buyers' published offers at $280,000 per company and the median of public sales at $31,600. Without a published benchmark, a written valuation matters more. A [free data valuation](https://sellmydatatoai.net/data-valuation) gives you a view of likely buyers and a realistic range for your records. ## Whose data is in your client accounts? The data in your client accounts is usually your clients' data, not yours. Customer lists, CRM records, ad audiences and analytics properties are normally held for the client, and when the agency processes personal data for a client it is acting as the client's processor. Article 28(3)(a) of the UK GDPR says a processor processes personal data only on documented instructions from the controller. Article 28(10) says a processor that determines the purposes of processing is treated as a controller for it. Licensing a client's customer data to an AI company would fall outside your instructions, so exclude it. | Record | Usually controlled by | Usual approach | | --- | --- | --- | | Client CRM, email lists and ad audiences | The client | Exclude | | Client analytics and ad account data | The client, under platform terms | Exclude | | Strategy documents and audits written by your staff | The agency, subject to the client contract | Include with client names removed | | Content drafts and edits | Depends on the client contract | Check the contract | | Performance reports | Agency-written, with client-confidential figures | Include with figures and names removed, or with consent | | Creator and influencer content | The creator, licensed for the campaign | Exclude | ## Can you include exports from SEO and ad platforms? Exports from SEO, analytics and ad platforms are governed by each platform's terms, and many limit how exported data can be shared or resold. Check the terms of every tool whose exports appear in your audits and reports. A safe approach is to keep the analyst's commentary and recommendations and remove pasted tables of third-party platform data. The reasoning is what buyers want, and it is the part your staff wrote. Where an audit relied on a crawler or rank tracker, describe the tool and its settings in words instead of pasting its output. A buyer can still follow the reasoning, and you avoid redistributing another company's data. ## What about your agency's own prospect and contact lists? Your agency's own prospect and contact lists are personal data collected for marketing, and they rarely add value to an AI buyer. The ICO updated its purpose limitation guidance in 2026 after changes made by section 71 of the Data (Use and Access) Act 2025, and RPC's summary notes that reuse for a new purpose must be compatible with the original one. Contact lists on their own are among the data types buyers rarely want, so leaving them out costs little. See the guide to [UK GDPR, EU GDPR and CCPA when selling data](https://sellmydatatoai.net/guides/gdpr-and-selling-data-for-ai-training) for the wider position. ## How does selling agency data work with Sell My Data To AI? Selling agency data with Sell My Data To AI starts with a consultation about your document stores, reporting tools and client contract types. We help you separate agency-written work from client data, and record the split in a [data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest), before any buyer is involved. We tell you if we think buyers will not be interested. Buyers you approve see a redacted, randomly drawn sample under a mutual NDA. You sign the license directly with the buyer. [Book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller) to start. ## Frequently asked questions ### Is there a published price for marketing agency data? We have found no buyer that publishes a price range for marketing agency data as of November 2026. The buyers compared by ddiy.co list other sectors, and the nearest Mercor listing is for design and web agencies at $20,000 to $890,000. A valuation based on your own records is the only reliable guide for now. ### Can an agency sell client campaign data? Usually not. Client CRM records, audiences and analytics data are normally held for the client, and an agency processing personal data for a client is its processor. Article 28(3)(a) of the UK GDPR limits a processor to the controller's documented instructions. Exclude client customer data and focus on work your staff wrote. ### Are published blog posts and articles worth selling? Published content is rarely worth much on its own, because it is public and has often been crawled already. The drafts, editorial changes and briefs behind it are more useful to buyers. If the client owns the published content under your contract, you need their agreement to include drafts of it too. ### Do we need to tell clients before licensing agency data? You need client agreement for anything their contract makes confidential or assigns to them. Even where it does not, you may still choose to tell key clients first. Removing client names and figures protects confidentiality, but the contract wording decides what you can include, so take independent advice where it is unclear. ## Sources - [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), ddiy.co (figures taken from each buyer's own site), accessed November 2026 - [How much is company data worth to AI companies](https://ddiy.co/how-much-is-company-data-worth-to-ai), ddiy.co (citing AIData.Broker), accessed November 2026 - [UK GDPR, Article 28: processor](https://www.legislation.gov.uk/eur/2016/679/article/28), legislation.gov.uk, accessed November 2026 - [ICO publishes updated purpose limitation guidance](https://www.rpclegal.com/snapshots/data-protection/summer-2026/ico-publishes-updated-purpose-limitation-guidance/), RPC, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How media companies and video archives can license footage to AI companies URL: https://sellmydatatoai.net/sell-your-data/media-and-video-archives Media companies and video archive owners can license footage to AI companies where they hold every rights layer in it. Versos reports OpenAI paying $1 to $4 per minute of video and Adobe about $3 per minute. Performers' consents, music, contributor releases and footage licensed in from third parties each need clearing before an archive can be offered. ## Key takeaways - AI buyers most want these records from media companies and video archives: (1) raw footage and rushes you shot or commissioned, (2) finished programs where you hold all rights layers and (3) shot logs, transcripts and descriptive metadata. - No AI buyer publishes a price range for media companies and video archives yet, so value depends on volume, years of history, structure and how cleanly the rights can be shown. - Size buyers look for: Priced by the minute or hour of footage, not by headcount. - The first rights issue to clear: Performers' rights under CDPA section 182 and the consents in their contracts. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Raw footage and rushes you shot or commissioned; Finished programs where you hold all rights layers; Shot logs, transcripts and descriptive metadata; Edit decision lists and versions showing how footage was cut; Stock and b-roll libraries with model and property releases; Archive catalogs with descriptions and rights records - Why buyers want them: Video model developers license footage at scale, and clips paired with transcripts or descriptions give a model both the video and text about it. Versos's list of places to sell video for AI training includes Google, Meta, NVIDIA, Runway, Moonvalley and Synthesia. - Size buyers look for: Priced by the minute or hour of footage, not by headcount. - Rights issues: Performers' rights under CDPA section 182 and the consents in their contracts; Music in footage is a separate copyright layer; Commissioned programs may leave some rights with the broadcaster; Agency footage cut into programs stays under the agency's license; Contributors and members of the public need releases or removal Video is one of the few data types with a reported per-minute price. Versos, a video licensing network, reported in February 2026 that OpenAI licenses video directly at $1 to $4 per minute and Adobe at about $3 per minute. An archive of 10,000 hours is 600,000 minutes, so the rights position of each hour matters. ## How much do AI companies pay for video footage? AI companies pay $1 to $4 per minute for video in OpenAI's case and about $3 per minute in Adobe's, as reported by Versos in February 2026. These are reported figures, not published rate cards, and Versos gives no prices for the other buyers it lists. Licensing agents work differently. [Troveo](https://sellmydatatoai.net/data-buyers/troveo) acts for rights holders on a revenue share with no fees, uses time-limited licenses and says full-company deals start at six figures, according to ddiy.co's review of buyer sites in October 2026. | Route | How you are paid | License term | | --- | --- | --- | | Direct license to a model developer | Per minute or per hour, as reported by Versos | Negotiated deal by deal | | Licensing agent such as Troveo | Revenue share, no fees | Time-limited | | Broker such as Sell My Data To AI | Paid by the buyer, our success fee only on completion | Agreed with each buyer | ## What footage do AI companies want? AI companies want footage where the rights are clear and each clip can be described. A clip with a transcript, shot log or description gives a model the video and the text about it, which is why archive metadata is worth preparing. Our page on [video training data](https://sellmydatatoai.net/buy-ai-training-data/video-training-data) sets out what buyers check. Raw rushes can be easier to license than finished programs. They are often free of the music, graphics and archive inserts added in the edit, which cuts the number of rights layers to clear. ## What rights do you need to license video for AI training? You need every rights layer in the footage, not just ownership of the master. Getphyllo's guide to training data licenses lists per-layer clearance as a requirement, because video contains music, images, performances and trademarks that may each belong to someone different. ### Performers Under [section 182 of the Copyright, Designs and Patents Act 1988](https://www.legislation.gov.uk/ukpga/1988/48/section/182), recording a live performance without the performer's consent infringes their rights. The consent in a performer's contract may be limited to broadcast or distribution, so check whether it reaches new uses before licensing footage for training. ### Music, commissioned programs and agency footage Music is normally licensed for a specific production and use. Commissioned programs may leave some rights with the broadcaster or production company, depending on the terms of trade. Agency footage cut into your programs stays under the agency's license, so it has to be removed or cleared. ### People in the footage Contributors, interviewees and the public are identifiable in video. The ICO says images used to uniquely identify a person are biometric data, a special category, and footage of identifiable people is personal data in any case. Versos advises sellers to hold clear, transferable rights and to include model releases and third-party clearances where needed. > **Rushes first** If your finished programs carry heavy music and archive clearances, start the [data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest) with rushes and b-roll you shot and own outright. ## How do you license an archive through Sell My Data To AI? We start with a consultation and a rights pre-check on a sample of the archive. Buyers see redacted samples under a mutual NDA, and full footage moves only after you sign a license directly with the buyer. Our seller vetting checks ownership, chain of title and third-party content before any buyer sees a frame. Read our guide to [licensing video and audio archives](https://sellmydatatoai.net/guides/sell-video-and-audio-for-ai-training), or [book a consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### How much is a minute of video worth to AI companies? A minute of video is worth $1 to $4 to OpenAI and about $3 to Adobe, according to figures Versos reported in February 2026. Those are reported rates for direct licensing, not published price lists. Footage with clear rights, descriptions and transcripts is easier to license than footage with open clearance questions. ### Can we license programs we made for a broadcaster? You can license commissioned programs only for the rights your terms of trade left with you. Some rights may sit with the broadcaster, and music, archive inserts and performers' consents were usually cleared for the original use. Rushes you shot and own outright are often simpler. Take independent advice on each commission. ### Do performers need to consent to AI training use? Performers' rights under section 182 of the Copyright, Designs and Patents Act 1988 protect against recording without consent, and what later uses are allowed depends on the performer's contract. If a contract does not cover licensing for AI training, treat that footage as uncleared until you have legal advice or a new consent. ### Is it better to use a licensing agent or license directly? A licensing agent such as Troveo takes a revenue share and uses time-limited licenses, while direct licenses are negotiated deal by deal. A broker helps you compare routes and buyers before you commit. Our guide to [marketplaces, brokers and direct deals](https://sellmydatatoai.net/guides/ai-data-marketplaces-compared) sets out the trade-offs. ## Sources - [Top 10 places to sell your video content for AI training data](https://www.versos.ai/blog/sell-video-content-ai-training-data), Versos, accessed November 2026 - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 - [Best public data providers for AI training and grounding](https://www.getphyllo.com/post/best-public-data-providers-ai-training-grounding), Getphyllo, accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 182: consent required for recording of live performance](https://www.legislation.gov.uk/ukpga/1988/48/section/182), legislation.gov.uk, accessed November 2026 - [What is special category data?](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/lawful-basis/special-category-data/what-is-special-category-data/), Information Commissioner's Office, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Sell medical practice data to AI companies URL: https://sellmydatatoai.net/sell-your-data/medical-practices Medical practices and clinics can license de-identified clinical notes, letters, protocols and dictation to AI companies, and Mercor publishes a range of $80,000 to $820,000 per practice. Patient confidentiality governs everything: GMC guidance says anonymized information must be used in preference to identifiable information wherever possible, and removing names and addresses alone is unlikely to be enough. ## Key takeaways - AI buyers most want these records from medical practices and clinics: (1) de-identified clinical notes and consultation records, (2) referral and clinic letters and (3) triage records with outcomes. - Published buyer range for medical practices and clinics: $80,000 to $820,000 per company. - Source: Mercor's published referral listing for medical practices, as reported by ddiy.co, checked November 2026. - Size buyers look for: 10 or more clinicians or 2,500 or more patients, de-identified (Mercor's listing). - The first rights issue to clear: Common law duty of confidentiality and GMC confidentiality guidance. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: De-identified clinical notes and consultation records; Referral and clinic letters; Triage records with outcomes; Coded diagnoses and prescribing, de-identified; Clinical protocols and care pathways; Dictation audio with transcripts - Why buyers want them: Clinical AI needs real examples of how clinicians record, reason and refer. De-identified records with outcomes are scarce and cannot be replaced by synthetic text without losing what makes them useful. - Published buyer range: $80,000 to $820,000 per company (Mercor's published referral listing for medical practices, as reported by ddiy.co, checked November 2026) - Size buyers look for: 10 or more clinicians or 2,500 or more patients, de-identified (Mercor's listing) - Rights issues: Common law duty of confidentiality and GMC confidentiality guidance; National data opt-out for publicly funded care in England; Health data is special category data under UK GDPR Article 9; HIPAA de-identification standard for US providers; Re-identification risk in rare conditions and free text ## What medical practice data do AI companies buy? AI companies buy medical practice data that shows how clinicians record, reason and decide, once it has been de-identified. Mercor's listing for medical practices specifies de-identified records. - Clinical notes and consultation records, de-identified - Referral letters and clinic letters - Triage records with the outcome of each case - Coded diagnoses, investigations and prescribing - Clinical protocols, care pathways and practice templates - Dictation audio with its transcript Protocols, pathways and templates written by the practice contain no patient information, so they are the simplest part of the archive to clear. Clinical records are worth more and carry far more risk. Coded records are easier for buyers to use when the coding system is stated, so record whether you use SNOMED CT or another scheme, and which version, in your manifest. See [expert domain data](https://sellmydatatoai.net/buy-ai-training-data/expert-domain-data) for how buyers use clinical material. ## How much do AI companies pay for medical practice data? AI companies publish a range of $80,000 to $820,000 for medical practice data. That is Mercor's referral listing for practices with 10 or more clinicians or 2,500 or more patients, with records de-identified, as reported by ddiy.co in October 2026. The size test is unusual because it is based on patients as well as staff, so a small practice with a large list can qualify. Not every broker will handle this material. SellMyData, for example, excludes person-linked health records from what it brokers. Practices that belong to a group or federation should first confirm who is the data controller for the records, because only the controller can authorize a license. A [free data valuation](https://sellmydatatoai.net/data-valuation) tells you whether your records are realistic to clear. ## What does GMC guidance say about using patient records? GMC guidance says anonymized information should be used in preference to identifiable information wherever possible, and that explicit consent is normally needed to use identifiable information for purposes other than direct care or local clinical audit. That comes from the GMC's confidentiality guidance on using and disclosing patient information for secondary purposes. The same guidance warns that simply removing a patient's name, age, address or other personal identifiers is unlikely to be enough to anonymize information. Section 251 of the NHS Act 2006 can set aside the duty of confidentiality for defined purposes, but only where anonymized information cannot be used and consent is not practicable. That route is not available where anonymized information would serve the purpose. For a practice, that leaves two realistic routes: properly anonymized records, or records from patients who gave explicit consent. Involve your Caldicott guardian or data protection officer before any work starts. Dictation audio needs extra care. The recording carries the clinician's voice as well as spoken patient details, so transcripts are usually easier to de-identify than the audio itself. Health data is special category data under Article 9 of the UK GDPR, whatever format it is in. ## Does the national data opt-out apply? The national data opt-out applies to confidential patient information from publicly funded care in England when it is used for purposes beyond individual care. NHS England says it covers all organizations providing or coordinating publicly funded health or adult social care in England, including private and independent providers. NHS England also says the opt-out does not apply to information that is anonymized in line with the ICO's code of practice on anonymization. So the opt-out is a further reason to anonymize thoroughly, not a separate route to selling identifiable records. ## How does HIPAA de-identification work for US practices? HIPAA de-identification works through one of two methods set out in section 164.514 of the HIPAA Privacy Rule, according to the HHS Office for Civil Rights guidance. The Safe Harbor list includes names, geographic subdivisions smaller than a state, most date elements, contact details, record and account numbers, biometric identifiers and full-face photos. | Method | What it requires | Practical note | | --- | --- | --- | | Safe Harbor | Remove 18 categories of identifiers of the patient and their relatives, employers or household members, with no actual knowledge that the rest could identify anyone | Free text often still holds identifiers after field removal | | Expert Determination | A qualified expert finds the risk of re-identification is very small and documents the methods and results | Better suited to rich clinical text that buyers value | Frankfurt Kurnit Klein and Selz advises excluding high-risk data entirely rather than relying on de-identification. For a practice, that means leaving out rare conditions, sexual health, mental health and records of children unless an expert has assessed them. See [de-identification for AI training data](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data). ## How does selling medical practice data work with Sell My Data To AI? Selling medical practice data with Sell My Data To AI starts with a consultation about your clinical system, list size, years of records and who acts as data controller. We tell you plainly if we think the rights position cannot be cleared. If it can, we help you document the anonymization method, run a re-identification test on a sample as our [vetting standard](https://sellmydatatoai.net/how-we-vet) requires, and share samples only with buyers you approve. You sign directly with the buyer after independent legal advice. [Book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller) to start. ## Frequently asked questions ### Can a GP practice sell patient data to AI companies? A GP practice cannot sell identifiable patient data without explicit consent, under GMC confidentiality guidance and the common law duty of confidentiality. Properly anonymized data is treated differently, and the national data opt-out does not apply to it. Removing names and addresses alone is unlikely to be enough, so take independent advice before starting. ### Is removing names enough to anonymize clinical notes? No. GMC guidance says simply removing a patient's name, age, address or other personal identifiers is unlikely to be enough. Free text can mention relatives, places, dates and rare conditions that identify a patient. A documented method and a re-identification test on a sample are needed before any buyer sees the data. ### What size of practice do AI buyers look for? Mercor's referral listing, reported by ddiy.co, asks for 10 or more clinicians or 2,500 or more patients, with records de-identified, and publishes a range of $80,000 to $820,000. Because the test includes patient numbers, a practice with few clinicians but a large patient list can still qualify. ### Do US clinics have to use Safe Harbor de-identification? No. The HHS guidance describes two methods under section 164.514 of the HIPAA Privacy Rule: Safe Harbor, which removes 18 categories of identifiers, and Expert Determination, where a qualified expert documents that re-identification risk is very small. Expert Determination is often better suited to the rich clinical text buyers value. ## Sources - [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), ddiy.co (figures taken from each buyer's own site), accessed November 2026 - [Confidentiality: using and disclosing patient information for secondary purposes](https://www.gmc-uk.org/professional-standards/the-professional-standards/confidentiality/using-and-disclosing-patient-information-for-secondary-purposes), General Medical Council, accessed November 2026 - [Understanding the national data opt-out](https://digital.nhs.uk/services/national-data-opt-out/understanding-the-national-data-opt-out), NHS England, accessed November 2026 - [Guidance regarding methods for de-identification of protected health information](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html), US Department of Health and Human Services, accessed November 2026 - [UK GDPR, Article 9: processing of special categories of personal data](https://www.legislation.gov.uk/eur/2016/679/article/9), legislation.gov.uk, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How publishers and education companies can sell data to AI companies URL: https://sellmydatatoai.net/sell-your-data/publishers-and-education Publishers and education companies can license back catalogs they hold the rights to, editorial records and assessment items to AI companies. Mercor's published listing puts publisher and education company payouts at $35,000 to $370,000 for firms with 10 to 500 staff. Check author contracts first, because an older contract may not give you AI licensing rights at all. ## Key takeaways - AI buyers most want these records from publishers and education companies: (1) back catalog text where you hold the licensing rights, (2) edited manuscripts, reviewer reports and copy-edit queries and (3) assessment items with mark schemes, model answers and examiner commentary. - Published buyer range for publishers and education companies: $35,000 to $370,000 per company. - Source: Mercor's published listing for publishers and education companies, recorded by ddiy.co, checked October 2026. - Size buyers look for: 10 to 500 staff (Mercor). - The first rights issue to clear: Author contracts may not grant you AI training or sublicensing rights. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Back catalog text where you hold the licensing rights; Edited manuscripts, reviewer reports and copy-edit queries; Assessment items with mark schemes, model answers and examiner commentary; Course materials, lesson plans and worked examples; Learner support tickets and tutor feedback, de-identified; Style guides and editorial procedures - Why buyers want them: Buyers value text with clean rights and records of expert judgment, such as an editor's changes or a marker's commentary. Assessment items with expected answers and scoring rules are already in the shape of an evaluation dataset. - Published buyer range: $35,000 to $370,000 per company (Mercor's published listing for publishers and education companies, recorded by ddiy.co, checked October 2026) - Size buyers look for: 10 to 500 staff (Mercor) - Rights issues: Author contracts may not grant you AI training or sublicensing rights; Freelance editors, illustrators and item writers own their work unless assigned in writing; Permissions for third-party images and extracts may be limited to one format or edition; Student records from US schools cannot be redisclosed without consent under FERPA; Learner work and personal data, often about children, should be excluded Publishing and education are the sectors where the AI copyright question is most public. On July 20, 2026 a US federal court in California granted final approval to the $1.5 billion Bartz v. Anthropic settlement over books used without permission, which the Authors Guild puts at about $3,000 per work. Licensed, rights-cleared text is the alternative. ## How much do AI companies pay for publishing and education data? AI companies pay publishers and education companies $35,000 to $370,000 per company, according to Mercor's published listing for the sector, which covers firms with 10 to 500 staff. ddiy.co recorded the range from Mercor's referral page in October 2026. That range sits below Mercor's listings for law firms ($50,000 to $1.2 million) and manufacturers ($100,000 to $1.5 million). Large publishers with major catalogs negotiate separately, and the reported agreements in our [AI data deals tracker](https://sellmydatatoai.net/ai-data-deals) show that end of the market. ## What publishing and education data do AI companies buy? AI companies buy publishing and education data that is rights-cleared and shows expert work. A back catalog matters only where you hold the rights to license it for training. | Record | What a buyer uses it for | Rights check | | --- | --- | --- | | Back catalog text | Training on edited, high-quality writing | Author contract grants AI licensing rights | | Edited manuscripts and reviewer reports | How experts improve text | Author and reviewer agreement | | Assessment items and mark schemes | Evaluation datasets with expected answers | Item ownership and exam security | | Course materials and worked examples | Domain teaching data | Third-party extracts and images | | Learner support and tutor feedback | Feedback and tutoring behavior | Learner personal data, often children's | Assessment items fit a format buyers name directly. An [evaluation dataset](https://sellmydatatoai.net/buy-ai-training-data/evaluation-datasets) is a set of tasks with expected answers and scoring rules, which is what an item bank with mark schemes already is. ## Do you own the AI rights in your back catalog? You own the AI rights in your back catalog only if your contracts give them to you. [Section 11 of the Copyright, Designs and Patents Act 1988](https://www.legislation.gov.uk/ukpga/1988/48/section/11) makes the author the first owner of copyright, so a publisher's rights come from the author contract, and a contract granting print and ebook rights may say nothing about AI training. Freelance editors, illustrators and item writers own their work unless they assigned it in writing, while employees' work belongs to the company. Audit contracts title by title before you build a [data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest), and leave out titles where the position is unclear. The UK government's report on copyright and AI, published on March 18, 2026 under the Data (Use and Access) Act 2025, says a broad copyright exception with an opt-out is no longer its preferred way forward and proposes not to intervene in the AI licensing market for now. For UK rights holders, licensing remains a matter of contract. ## Can education companies license student and learner data? Education companies should exclude student and learner data from any license. Under FERPA, US schools disclose personally identifiable information from education records only on condition that the recipient will not pass it to any other party without consent, as [34 CFR 99.33](https://www.ecfr.gov/current/title-34/subtitle-A/part-99/subpart-D/section-99.33) sets out. The US Department of Education treats records as de-identified only once all personally identifiable information is removed, including information that alone or in combination is linkable to a specific student. UK providers holding children's data should take the same approach and leave it out. > **License retired items, not live ones** Models can reproduce parts of their training data, as Frankfurt Kurnit Klein and Selz warns. License retired assessment items, and keep live papers and reusable item banks out of any package. ## How do you sell publishing data through Sell My Data To AI? We start with a consultation and a rights pre-check. Our eight-point vetting standard covers ownership and chain of title, contract restrictions and third-party content, the checks that matter most for a back catalog. Buyers see redacted samples under a mutual NDA, and you approve every buyer and every license term. Read the guide to [exclusive and non-exclusive licenses](https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses), or [book a consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### How much does Mercor pay publishers for data? Mercor's published listing puts publisher and education company payouts at $35,000 to $370,000 per company, for firms with 10 to 500 staff. ddiy.co recorded the range from Mercor's referral page in October 2026. Where you land depends on rights clarity, volume and how much expert editorial material you hold. ### Can we license books whose contracts do not mention AI? If an author contract does not grant you the right to license text for AI training, assume you do not have it. Copyright starts with the author, and a publisher holds only the rights the contract transfers. You can ask authors for a written variation, often with a share of the fee, and should take independent legal advice. ### What did the Anthropic settlement decide? The Bartz v. Anthropic settlement resolved authors' copyright claims over books used to train AI models. A US federal court granted final approval on July 20, 2026 for $1.5 billion, which the Authors Guild puts at about $3,000 per work. It is a settlement, so it does not set a legal rule on training. ### Can an edtech company sell student data? An edtech company should not sell student personal data. In the US, FERPA's redisclosure rule bars recipients of education records from passing them on without consent, and UK learner data is often children's data. Teaching materials, retired assessment items and fully de-identified feedback are the licensable assets. ## Sources - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 - [Court grants final approval of $1.5 billion Anthropic copyright settlement](https://authorsguild.org/news/court-grants-final-approval-anthropic-copyright-settlement/), The Authors Guild, accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 11: first ownership of copyright](https://www.legislation.gov.uk/ukpga/1988/48/section/11), legislation.gov.uk, accessed November 2026 - [Report on copyright and artificial intelligence](https://www.gov.uk/government/publications/report-and-impact-assessment-on-copyright-and-artificial-intelligence/report-on-copyright-and-artificial-intelligence), GOV.UK (DSIT and IPO), accessed November 2026 - [34 CFR 99.33: limitations on the redisclosure of information](https://www.ecfr.gov/current/title-34/subtitle-A/part-99/subpart-D/section-99.33), eCFR (US Government), accessed November 2026 - [What constitutes de-identified records and information?](https://studentprivacy.ed.gov/faq/what-constitutes-de-identified-records-and-information), US Department of Education, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz (Daniel Goldberg), accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How real estate firms can sell data to AI companies URL: https://sellmydatatoai.net/sell-your-data/real-estate Estate and letting agents can license in-house property descriptions, valuation reasoning, lettings maintenance tickets and procedures to AI companies, but not the personal data they collected for anti-money laundering checks. Regulation 41 of the Money Laundering Regulations 2017 limits that data to crime prevention. Public data such as sold prices has little value, because buyers can collect it themselves. ## Key takeaways - AI buyers most want these records from real estate firms: (1) property descriptions and particulars written by your own staff, (2) valuation and appraisal notes explaining comparables and price advice and (3) lettings maintenance tickets, contractor instructions and resolutions. - No AI buyer publishes a price range for real estate firms yet, so value depends on volume, years of history, structure and how cleanly the rights can be shown. - Size buyers look for: No real estate minimum is published. micro1 asks for 30 or more full-time staff, and Handshake AI for 20 or more staff and three years of trading. - The first rights issue to clear: Personal data collected for AML checks may only be used for crime prevention under MLR 2017 regulation 41. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Property descriptions and particulars written by your own staff; Valuation and appraisal notes explaining comparables and price advice; Lettings maintenance tickets, contractor instructions and resolutions; Vendor, applicant, tenant and landlord email threads, de-identified; Sales progression notes and chasers through to completion; Procedure manuals, compliance checklists and templates - Why buyers want them: Buyers want records of how agents reason and handle cases, not listings they can already see. micro1's brief asks for workflows, templates, CRM records and support tickets, which describes most of a sales or lettings file. - Size buyers look for: No real estate minimum is published. micro1 asks for 30 or more full-time staff, and Handshake AI for 20 or more staff and three years of trading. - Rights issues: Personal data collected for AML checks may only be used for crime prevention under MLR 2017 regulation 41; Freelance photographers own copyright in property photos unless they assigned it; Vendors, buyers, tenants and landlords are identifiable throughout correspondence; Franchise and portal agreements may restrict reuse of listing content; Public data such as sold prices has little licensing value Real estate firms hold two kinds of data: public facts about property and private records of how agents work. AI buyers can already collect the first. The second, from valuation reasoning to maintenance ticket histories, is what a license can sell. ## What real estate data do AI companies buy? AI companies buy real estate data that records professional judgment and casework. micro1, which committed $1 billion in October 2026 to buying enterprise operational data, asks for SOPs, process documentation, workflows, templates, CRM records, support tickets and project histories. Free text fits that brief best. A valuer's note on why one comparable was discounted, or a property manager's thread from a tenant reporting a leak through to the contractor's sign-off, shows a decision and its outcome. Data that anyone can download is a weak asset. HM Land Registry publishes price paid data openly, so sold prices and listing histories add little to a license. | Record | Licensing value | Main issue | | --- | --- | --- | | In-house property descriptions | Writing and local property knowledge | Photos and floor plans may belong to a photographer | | Valuation and appraisal notes | Price reasoning from comparables | Vendor identity and addresses | | Lettings maintenance tickets | Fault diagnosis and contractor handling | Tenant names and access details | | Sales progression notes | Case management through to completion | Buyer, seller and solicitor details | | AML and ID checks | None for licensing | Restricted by MLR 2017 regulation 41 | | Sold prices and listing history | Low, widely available | Portal terms | ## Can estate agents sell anti-money laundering records? Estate agents cannot license the personal data they collected for anti-money laundering checks. [Regulation 41 of the Money Laundering Regulations 2017](https://www.legislation.gov.uk/uksi/2017/692/regulation/41) says personal data obtained for the Regulations may only be processed to prevent money laundering, terrorist financing or proliferation financing, unless another enactment permits the use or the data subject consents. HMRC is the supervisory authority for estate agency businesses, and for letting agency businesses handling rents of 10,000 euros or more a month, according to GOV.UK. Treat ID documents, source of funds evidence and customer risk assessments as outside any data license. ## Who owns property photos and floor plans? Property photos and floor plans belong to whoever took or drew them unless the copyright was assigned. Section 11 of the Copyright, Designs and Patents Act 1988 makes the author the first owner, with an exception for work made by employees in the course of employment. If a freelance photographer supplies your images under a license for marketing, check whether it covers any other use. If it does not mention AI training, assume it does not cover it, and strip the images or get written permission before they go into a package. ## What other rights issues apply to real estate data? Correspondence identifies vendors, buyers, tenants, landlords and solicitors. The privacy notice in force when you collected that data governs what you may do with it, and UK GDPR Article 8A asks whether licensing is compatible with the original purpose. Frankfurt Kurnit Klein and Selz notes that removing identifying details from millions of emails is much harder than removing names from structured fields. Our guide to [de-identification for AI training data](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data) explains the methods buyers expect. Franchise and portal agreements can also limit how listing content is reused, so read them alongside your own terms of business. > **Where to start** Procedure manuals, templates and in-house property descriptions are the cleanest records an agency holds. Build the package around them, then add de-identified casework once the privacy questions are answered. ## How do you sell real estate data through Sell My Data To AI? You begin with a consultation. Sellers pay nothing to be assessed, and our success fee is agreed in writing before any introduction. Our eight-point vetting standard checks authority to sign, ownership, contract restrictions and privacy notices, so AML records and third-party photos are caught before a buyer sees a sample. See [how we vet](https://sellmydatatoai.net/how-we-vet) for the full standard, read [whether you can legally sell company data](https://sellmydatatoai.net/guides/can-you-legally-sell-company-data-for-ai), or [book a consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### Can letting agents sell tenant data to AI companies? Letting agents should not sell identifiable tenant data to AI companies. Identity checks collected for AML purposes are restricted by regulation 41 of the Money Laundering Regulations 2017, and other tenant data is limited by the privacy notice you gave at the time. De-identified maintenance tickets and procedures are a safer basis for a license. ### Is sold price data worth anything to AI buyers? Sold price data is worth little on its own, because HM Land Registry publishes price paid data openly. Your valuation notes explaining how you reached a price are different. They record professional reasoning that is not public, and that reasoning is what a buyer would license. ### Do we need consent from vendors and landlords? Consent is one route, but it is rarely practical across thousands of past clients. The usual alternative is to remove personal data and license de-identified records. Whether that is enough depends on your privacy notice and the de-identification method, so take independent legal advice before licensing client correspondence. ### Which buyers take real estate data? No buyer publishes a real estate brief or price range. General operational data buyers such as [micro1](https://sellmydatatoai.net/data-buyers/micro1) ask for workflows, templates and CRM records from established firms with 30 or more staff. We tell you in one consultation whether your records are in demand and which buyers fit them. ## Sources - [The Money Laundering, Terrorist Financing and Transfer of Funds Regulations 2017, regulation 41](https://www.legislation.gov.uk/uksi/2017/692/regulation/41), legislation.gov.uk, accessed November 2026 - [Who needs to register for money laundering supervision](https://www.gov.uk/guidance/money-laundering-regulations-who-needs-to-register), GOV.UK (HMRC), accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 11: first ownership of copyright](https://www.legislation.gov.uk/ukpga/1988/48/section/11), legislation.gov.uk, accessed November 2026 - [UK GDPR Article 8A: purpose limitation, further processing](https://www.legislation.gov.uk/eur/2016/679/article/8A), legislation.gov.uk, accessed November 2026 - [Monetize your company's operational data](https://www.micro1.ai/data), micro1, accessed November 2026 - [micro1 announces commitment to spend $1B on enterprise data acquisition over the next 12 months](https://www.micro1.ai/blog/micro1-commits-1b-to-enterprise-data-acquisition), micro1, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz (Daniel Goldberg), accessed November 2026 - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How recruitment agencies can sell data to AI companies URL: https://sellmydatatoai.net/sell-your-data/recruitment-agencies Recruitment agencies can license job descriptions, anonymized shortlisting decisions, interview feedback and placement outcomes to AI companies, but not candidates' CVs and details without consent. Regulation 28 of the Conduct of Employment Agencies and Employment Businesses Regulations 2003 bars UK agencies from disclosing work-seekers' information except for work-finding services. Buyers building hiring tools also face high-risk rules under the EU AI Act. ## Key takeaways - AI buyers most want these records from recruitment agencies: (1) job descriptions and role briefs, with client names removed, (2) shortlisting decisions and the reasons recorded and (3) interview feedback and assessment notes, de-identified. - No AI buyer publishes a price range for recruitment agencies yet, so value depends on volume, years of history, structure and how cleanly the rights can be shown. - Size buyers look for: No recruitment minimum is published. micro1 asks for 30 or more full-time staff, and Handshake AI for 20 or more staff and three years of trading. - The first rights issue to clear: Regulation 28 of the Conduct Regulations 2003 limits disclosure of work-seeker information. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Job descriptions and role briefs, with client names removed; Shortlisting decisions and the reasons recorded; Interview feedback and assessment notes, de-identified; Recruiter and candidate messaging, de-identified; Placement outcomes and retention data in aggregate; Salary benchmarking and market reports you authored - Why buyers want them: Recruitment records capture how people are matched to roles, from brief to shortlist to placement. Buyers using them for hiring tools fall under the EU AI Act's high-risk category, so they need clear provenance. - Size buyers look for: No recruitment minimum is published. micro1 asks for 30 or more full-time staff, and Handshake AI for 20 or more staff and three years of trading. - Rights issues: Regulation 28 of the Conduct Regulations 2003 limits disclosure of work-seeker information; Diversity monitoring data is special category data; Criminal record and background checks are criminal offense data; Client job briefs and fee terms are confidential; The EU AI Act treats recruitment AI as high-risk, raising buyers' provenance demands The asset most agencies think of first, the candidate database, is the hardest to license. A CV database is mostly personal data, and UK law adds a sector-specific rule on top of UK GDPR. The records that show how recruiters work, from briefs to shortlisting reasons to outcomes, carry far less risk. ## What recruitment data do AI companies buy? AI companies buy recruitment data that shows matching and judgment. micro1's brief asks for workflows, templates, CRM records and internal communications, and an agency's ATS and CRM hold all of these. | Record | What it shows | Main issue | | --- | --- | --- | | Job descriptions and briefs | How roles are defined | Client confidentiality | | Shortlisting reasons | How candidates are matched | Candidate identity | | Interview feedback | How people are assessed | Candidate identity and fairness | | Recruiter messaging | Candidate and client handling | Personal data throughout | | CVs and candidate profiles | Work histories | Regulation 28 and UK GDPR | | Diversity and background checks | None for licensing | Special category and criminal offense data | ## Can recruitment agencies sell candidate CVs? UK recruitment agencies cannot sell candidate CVs without each candidate's consent. [Regulation 28 of the Conduct of Employment Agencies and Employment Businesses Regulations 2003](https://www.legislation.gov.uk/uksi/2003/3319/regulation/28) says an agency or employment business must not disclose information about a work-seeker without their prior consent, except to provide work-finding services, for legal proceedings or to the work-seeker's professional body. That rule sits alongside UK GDPR. A license to an AI company is not a work-finding service. Whether [de-identified](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data) records still count as information about a work-seeker is a question for your lawyer, so take advice before licensing anything derived from candidate files. ## Which candidate data must always be excluded? Diversity monitoring and background check data must always be excluded. The ICO lists racial or ethnic origin, religious beliefs, health, sex life and sexual orientation as special category data, and treats personal data about criminal allegations, proceedings or convictions as criminal offense data with its own rules. Daniel Goldberg of Frankfurt Kurnit Klein and Selz recommends excluding high-risk data entirely rather than relying on de-identification. For an agency, that means equal opportunities forms, right-to-work documents, criminal record checks and any health or adjustment information. ## Why do AI buyers care about recruitment data provenance? AI buyers care about recruitment data provenance because hiring tools are regulated. Annex III, point 4 of the [EU AI Act](https://eur-lex.europa.eu/eli/reg/2024/1689/oj) lists AI systems intended for the recruitment or selection of people as high-risk, and Article 10 requires the training data of high-risk systems to meet data governance criteria, including examination for possible biases. A buyer in that position will ask where every record came from, how candidates were told and whether the data reflects biased past decisions. Our [data provenance guide](https://sellmydatatoai.net/guides/data-provenance-explained) explains what they look for. > **Client briefs are confidential too** Job briefs, fee terms and hiring managers' comments are your clients' confidential information. Remove client names and anything that could identify a role before including briefs. ## How do you sell recruitment data through Sell My Data To AI? We start with a consultation. Our eight-point vetting standard checks contract restrictions, privacy notices and the de-identification method, with a re-identification test run on a sample. Buyers see redacted samples under a mutual NDA, and you approve every buyer and every license term. Read [whether selling company data to AI is safe](https://sellmydatatoai.net/guides/is-selling-company-data-to-ai-safe), or [book a consultation](https://sellmydatatoai.net/book-a-consultation?side=seller). ## Frequently asked questions ### Is it legal to sell a candidate database to AI companies? In the UK, selling identifiable candidate data to an AI company is very hard to reconcile with regulation 28 of the Conduct Regulations 2003, which bars disclosing work-seekers' information without consent except for work-finding services. UK GDPR also applies. Take independent legal advice, and build any license around de-identified process records. ### Do any buyers publish a price for recruitment data? No buyer publishes a recruitment-specific price. micro1's general tiers start at $100,000 to $500,000 for a qualified partnership, for companies with 30 or more full-time staff. We tell you in one consultation whether your records are in demand and what a realistic range looks like. ### Can we license job descriptions? You can usually license job descriptions you wrote once client names and identifying details are removed. Check your terms of business, because some clients treat role briefs and fee terms as confidential. Published job adverts are public, so they carry less value than internal briefs and shortlisting notes. ### Does the EU AI Act affect a UK agency selling data? The EU AI Act mainly affects the buyer, not the UK agency, but it shapes what buyers ask for. AI systems for recruitment or selection are high-risk under Annex III, so buyers building them need documented provenance and data governance. Expect detailed questions about how candidates were informed. ## Sources - [The Conduct of Employment Agencies and Employment Businesses Regulations 2003, regulation 28](https://www.legislation.gov.uk/uksi/2003/3319/regulation/28), legislation.gov.uk, accessed November 2026 - [What is special category data?](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/lawful-basis/special-category-data/what-is-special-category-data/), Information Commissioner's Office, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz (Daniel Goldberg), accessed November 2026 - [Regulation (EU) 2024/1689 (Artificial Intelligence Act)](https://eur-lex.europa.eu/eli/reg/2024/1689/oj), EUR-Lex, accessed November 2026 - [Monetize your company's operational data](https://www.micro1.ai/data), micro1, accessed November 2026 - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # How retail and ecommerce businesses can sell data to AI companies URL: https://sellmydatatoai.net/sell-your-data/retail-and-ecommerce Retailers and ecommerce businesses can license customer support conversations, product copy they wrote, returns reasons and merchandising decisions to AI companies. Customer purchase histories are personal data, and in California handing them over for value is a sale under the CCPA unless de-identified to its standard. Supplier images and descriptions usually belong to the brand, not to you. ## Key takeaways - AI buyers most want these records from retail and ecommerce businesses: (1) customer support tickets, chats and email threads, de-identified, (2) product descriptions, buying guides and size guides written in-house and (3) returns reasons and resolution notes. - No AI buyer publishes a price range for retail and ecommerce businesses yet, so value depends on volume, years of history, structure and how cleanly the rights can be shown. - Size buyers look for: No retail minimum is published. micro1 asks for 30 or more full-time staff, and Handshake AI for 20 or more staff and three years of trading. - The first rights issue to clear: Customer purchase histories are personal data governed by the privacy notice at collection. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Customer support tickets, chats and email threads, de-identified; Product descriptions, buying guides and size guides written in-house; Returns reasons and resolution notes; Product question and answer threads; Merchandising, buying and range planning notes; Site search queries and zero-result logs, de-identified - Why buyers want them: micro1's published brief names customer support and sales and CRM among the functions it wants data from. A support thread records a request, the steps taken and the outcome, which is the structure of an agent trajectory. - Size buyers look for: No retail minimum is published. micro1 asks for 30 or more full-time staff, and Handshake AI for 20 or more staff and three years of trading. - Rights issues: Customer purchase histories are personal data governed by the privacy notice at collection; Under the CCPA, transferring personal information for value is a sale unless de-identified; Supplier images and descriptions belong to whoever created them, often the brand; Customer reviews belong to their authors unless your terms grant a wide enough license; Payment card data should be excluded entirely Retailers hold two very different assets. Transaction records are mostly personal data about customers. Support conversations, product copy and buying decisions record how the business works, and they are much closer to what AI buyers publish briefs for. ## What retail and ecommerce data do AI companies buy? AI companies buy retail and ecommerce data that captures work being done. micro1's published brief names customer support and sales and CRM as target functions and asks for support tickets, CRM records, templates and knowledge bases. A support ticket that runs from a complaint through order lookup and refund rules to a resolution is close to an [agent trajectory](https://sellmydatatoai.net/buy-ai-training-data/agent-trajectory-data). Product copy and buying guides your team wrote are [expert domain data](https://sellmydatatoai.net/buy-ai-training-data/expert-domain-data) for the categories you sell. | Record | Value to a buyer | Main issue | | --- | --- | --- | | Support tickets and chats | Request, steps and resolution | Customer names, addresses and order details | | In-house product copy and guides | Category expertise | Mixed with supplier text | | Returns reasons | Why products fail customers | Low once de-identified | | Merchandising and buying notes | Commercial reasoning | Supplier terms and price confidentiality | | Purchase histories and loyalty data | Limited without context | Personal data and CCPA sale rules | | Card payment data | None | Exclude | ## Can you sell customer purchase data for AI training? Customer purchase data can only be licensed within the limits of the privacy notice in force when you collected it. Daniel Goldberg of Frankfurt Kurnit Klein and Selz notes that companies must honor the privacy statements that applied at collection, and that the US Federal Trade Commission has warned quietly adopting more permissive practices may be unfair or deceptive. In California, the CCPA treats handing over personal information for value as a sale, unless the data meets the Act's de-identification standard: technical measures, a public commitment not to re-identify and contractual limits on recipients. In the UK, [Article 8A of UK GDPR](https://www.legislation.gov.uk/eur/2016/679/article/8A) asks whether the new purpose is compatible with the original one, weighing the customer relationship and the possible consequences for customers. ## Who owns product images, descriptions and reviews? Product images and descriptions belong to whoever created them. Under section 11 of the Copyright, Designs and Patents Act 1988 the author is the first owner, and an employer owns work its staff create in the course of employment. Text and photos supplied by brands are theirs, provided to you for selling their products. Customer reviews follow the same rule. The reviewer owns the copyright unless your terms of use assign it or grant you a license wide enough to cover AI training. Check the terms that applied when each review was posted, and remove supplier content before you build a [data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest). > **Exclude card data** Payment card numbers add nothing to a training dataset and carry heavy security obligations. Remove them, along with any field that could be used to take a payment, before records leave your systems. ## How much is retail data worth to AI buyers? No buyer publishes a retail or ecommerce range. micro1's general tiers start at $100,000 to $500,000 for a qualified partnership. Handshake AI says it has paid $100,000 to $4 million to date and looks for 20 or more staff and three or more years of trading, according to ddiy.co's review of its site in October 2026. The volume of support history and the share of content you wrote yourself drive the range. Our guide to [what your data is worth](https://sellmydatatoai.net/guides/how-much-is-your-data-worth-to-ai) explains the factors, and [exclusive and non-exclusive licenses](https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses) price very differently. ## How do you sell retail data through Sell My Data To AI? We start with a consultation and never take custody of your raw data. Buyers see redacted samples under a mutual NDA, and full data moves only after you sign a license directly with the buyer. You approve every buyer, every sample and every license term, and you can withdraw at any point before signing. [Book a consultation](https://sellmydatatoai.net/book-a-consultation?side=seller) to find out which of your records are in demand. ## Frequently asked questions ### Can an online shop sell customer data to AI companies? An online shop can license de-identified support records and content it wrote, but selling identifiable customer data is limited by the privacy notice it gave customers and by laws such as UK GDPR and the CCPA. We recommend excluding raw customer data and licensing the records that show how the business works. ### Can we license product descriptions supplied by brands? You cannot license product descriptions supplied by brands unless the brand agrees. Copyright stays with the author, and brand content is provided for selling the brand's products. Descriptions, buying guides and size guides written by your own employees belong to your company under the Copyright, Designs and Patents Act 1988. ### Are customer reviews ours to sell? Customer reviews belong to the people who wrote them unless your terms of use assign copyright or grant you a license. Check whether that license covers sublicensing and AI training, and which version of your terms applied when each review was posted. Reviews also carry names and personal details that need removing. ### What does the CCPA mean for selling retail data? Under the CCPA, giving personal information to another business for value counts as a sale unless the data meets the Act's de-identification standard. That standard requires technical measures, a public commitment not to re-identify and contractual limits on recipients. Our [UK GDPR, EU GDPR and CCPA guide](https://sellmydatatoai.net/guides/gdpr-and-selling-data-for-ai-training) explains the steps. ## Sources - [Monetize your company's operational data](https://www.micro1.ai/data), micro1, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz (Daniel Goldberg), accessed November 2026 - [UK GDPR Article 8A: purpose limitation, further processing](https://www.legislation.gov.uk/eur/2016/679/article/8A), legislation.gov.uk, accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 11: first ownership of copyright](https://www.legislation.gov.uk/ukpga/1988/48/section/11), legislation.gov.uk, accessed November 2026 - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Sell software company data and source code to AI companies URL: https://sellmydatatoai.net/sell-your-data/software-companies Software companies can license source code with full git history, pull requests, issue trackers, incident postmortems and support tickets to AI companies, and Mercor publishes a range of $50,000 to $500,000 per company. Before a sale you need to strip third-party code and secrets, confirm contractors assigned their IP, and remove customer data. ## Key takeaways - AI buyers most want these records from software companies: (1) repositories with full commit history, (2) pull requests with review comments and (3) issue trackers linked to commits. - Published buyer range for software companies: $50,000 to $500,000 per company. - Source: Mercor's published referral listing for software companies, as reported by ddiy.co, checked November 2026. - Size buyers look for: 15 to 500 employees (Mercor's listing). - The first rights issue to clear: Open source and third-party code inside your repositories. This page covers 4 more. - Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, and the fee is agreed in writing before we introduce you to any buyer. ## At a glance - Records buyers ask for: Repositories with full commit history; Pull requests with review comments; Issue trackers linked to commits; Incident postmortems and runbooks; Support tickets and internal engineering chat; Test suites and CI logs - Why buyers want them: Coding models and agents learn from how real teams change code over time, why a review rejected a change and how an incident was fixed. Public code is already widely available to labs, so private codebases with history are what buyers ask for. - Published buyer range: $50,000 to $500,000 per company (Mercor's published referral listing for software companies, as reported by ddiy.co, checked November 2026) - Size buyers look for: 15 to 500 employees (Mercor's listing) - Rights issues: Open source and third-party code inside your repositories; Contractor-written code without a written IP assignment; Customer data in logs, tickets and test fixtures; Code written for clients under services contracts; Secrets and credentials in commit history ## What software company data do AI companies buy? AI companies buy software company data that shows how code changes and why. A repository with its full commit history, linked pull requests and review comments teaches a coding model far more than a snapshot of the current code. - Repositories with full commit history and branch structure - Pull requests with review comments and requested changes - Issue trackers where tickets link to the commits that closed them - Incident postmortems, runbooks and on-call notes - Support tickets that escalated to engineering - Test suites, CI logs and architecture decision records DataVendor, a marketplace for code and agent tasks, says labs pay more for human-written code, complex products and systems languages such as Rust, Go, C and C++. Linked engineering records can also become [agent trajectory data](https://sellmydatatoai.net/buy-ai-training-data/agent-trajectory-data) or evaluation tasks, which some buyers price separately. Retired products count too. A product you closed years ago can still be useful if its repository and tickets survived, and MIT Technology Review's October 2026 reporting describes labs buying data from startups that have shut down. ## How much do AI companies pay for software company data? AI companies publish a range of $50,000 to $500,000 for software company data. That is Mercor's referral listing for software companies of 15 to 500 employees, as reported by ddiy.co in October 2026. Real sales can land well below that. MIT Technology Review reported in October 2026 that a shut-down crypto startup received about $25,000 for its codebase, with about $15,000 more pending for its workspace data. A small codebase in a common web stack sits at the low end, and a large, long-lived product in a systems language with linked tickets sits higher. Read [how to sell source code to AI companies](https://sellmydatatoai.net/guides/sell-source-code-to-ai-companies) for pricing factors. ## Does selling your code give away your product? Selling your code to an AI company does not hand the buyer your product, but it does put your code into a model you do not control. The buyer is licensing the code to train models, not to run your service, and a license can say so. The risk to weigh is that a model trained on your code may later help someone build something similar. If you are still trading, you may prefer a license with tight use limits, or sell only once a product is retired. Companies that have closed face fewer such trade-offs, see [selling data from a closed company](https://sellmydatatoai.net/guides/selling-data-from-a-closed-company). ## Which code can you legally license? You can license code your company owns and that no other license restricts. Under section 11 of the Copyright, Designs and Patents Act 1988, your company owns code written by employees in the course of their employment, but a contractor is the first owner of code they wrote unless they assigned it to you in writing. | Code or content | Rights position | Usual approach | | --- | --- | --- | | Code written by employees | Owned by the company under CDPA section 11(2) | Include | | Code written by contractors | Owned by the contractor unless assigned | Include only with a signed assignment | | Open source dependencies and vendored libraries | Licensed to you under each project's license | Remove or list separately in the manifest | | Code built for a client under a services contract | Often owned by or licensed to the client | Check the contract, usually exclude | | Commercial SDKs and licensed components | Licensed for your product only | Remove | | Secrets, keys and credentials | Security risk | Remove from the full history and rotate | Buyers run their own license scans, so an accurate list of third-party code in your [data manifest](https://sellmydatatoai.net/guides/how-to-build-a-data-manifest) builds trust quickly. Removing a secret from the latest commit is not enough because it remains in the history. Tools such as git filter-repo can rewrite history to remove secrets and excluded files, but run the clean-up before you count commits for the manifest, so the numbers match what is delivered. ## What customer data sits in engineering records? Engineering records often hold customer data in logs, database fixtures, support tickets and incident timelines. If you run a SaaS product, you are usually a processor for your customers' data, and Article 28(3)(a) of the UK GDPR says a processor processes personal data only on documented instructions from the controller. Licensing that data to an AI company would not be on your customers' instructions. Article 28(10) adds that a processor who determines the purposes of processing is treated as a controller for it. Strip customer data from tickets, logs and fixtures before any sample is drawn, and see [support ticket data](https://sellmydatatoai.net/buy-ai-training-data/support-ticket-data) for how buyers expect it to be cleaned. ## How does selling software company data work with Sell My Data To AI? Selling software company data with Sell My Data To AI starts with a consultation about your repositories, languages, years of history and linked systems such as Jira, GitHub or Linear. No code leaves your systems at this stage. We then help you prepare a manifest of repositories and exclusions, run our [eight vetting checks](https://sellmydatatoai.net/how-we-vet) and share a redacted sample with buyers you approve. Buyers' tools never get direct access to your live systems: a fixed export moves only after you sign the license. [Book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=seller) to start. ## Frequently asked questions ### Do AI companies buy code from small software companies? They do. Mercor's referral listing, reported by ddiy.co, covers software companies of 15 to 500 employees, and MIT Technology Review reported a shut-down startup selling its codebase for about $25,000. Smaller codebases usually sell for less, but long commit histories, linked tickets and systems languages raise interest at any size. ### Do we need to remove open source code before selling? You need to identify open source code and either remove it or list it separately, because it is licensed to you rather than owned by you. Buyers scan for it, and an accurate list in your manifest avoids disputes later. Code your own employees wrote stays in, as your company owns it under section 11(2) of the CDPA. ### Can we sell code while the product is still live? You can, with care. The license should limit the buyer to training and evaluation, bar redistribution of the code and set rules on onward transfer. You could also license only retired products or older versions. We check every buyer's intended use and onward transfer limits before they see a sample. ### Are pull request reviews worth more than the code itself? Pull request reviews often add a large share of the value, because they show why a change was accepted or rejected. DataVendor notes labs pay more for human-written code from complex products. A repository with linked reviews and tickets is a richer package than code alone, and buyers price it that way. ## Sources - [How to sell your company's data to AI companies](https://ddiy.co/sell-company-data-to-ai/), ddiy.co (figures taken from each buyer's own site), accessed November 2026 - [AI training data marketplaces](https://datavendor.ai/blog/ai-training-data-marketplaces), DataVendor, accessed November 2026 - [AI labs are buying data from dead startups to train their models](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models), MIT Technology Review, accessed November 2026 - [Copyright, Designs and Patents Act 1988, section 11](https://www.legislation.gov.uk/ukpga/1988/48/section/11), legislation.gov.uk, accessed November 2026 - [UK GDPR, Article 28: processor](https://www.legislation.gov.uk/eur/2016/679/article/28), legislation.gov.uk, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Buy workplace communications data for AI training URL: https://sellmydatatoai.net/buy-ai-training-data/workplace-communications-data Workplace communications data is licensed email, chat and meeting text from real companies, used to train and test AI agents on how work gets done. Public prices are few: MIT Technology Review reported about $15,000 for one closed startup's workspace data, while Google won Spirit Airlines' archive of 100 million emails and 500 million Teams messages with a $10 million bid. ## Key takeaways - AI teams use workplace communications data for (1) training agents on multi-step business workflows, (2) fine-tuning assistants for email and chat drafting and (3) reinforcement learning environments built from real work. - Usually priced per company archive, as a flat license fee. - Common delivery formats: MBOX or EML email exports, Slack or Teams JSON exports and JSONL threads with message metadata. - Check before you license workplace communications data: (1) US: employee messages protected under NLRA Section 7, (2) US: CCPA coverage of California employees' personal information and (3) US: privacy notice in force when the messages were sent. - Every seller we shortlist passes 8 vetting checks before you see a sample. ## Buyer's spec - Used for: Training agents on multi-step business workflows; Fine-tuning assistants for email and chat drafting; Reinforcement learning environments built from real work; Evaluating judgment in ambiguous workplace situations - Formats: MBOX or EML email exports; Slack or Teams JSON exports; JSONL threads with message metadata; Parquet tables for large archives - Usually priced: Per company archive, as a flat license fee - Check before you sign: US: employee messages protected under NLRA Section 7; US: CCPA coverage of California employees' personal information; US: privacy notice in force when the messages were sent; Third-party confidential information under NDAs; HR records excluded; De-identification method and re-identification test; Thread completeness and date range in the manifest; UK and EU: lawful basis under UK GDPR or EU GDPR, or anonymization ## What is workplace communications data? Workplace communications data is the written record of how a company runs: email, Slack and Teams channels, meeting notes and the documents linked from them. AI buyers value it because it shows requests, decisions, handoffs and disagreements in context. MIT Technology Review reported in October 2026 that Surge AI looks for archives covering situations from "disagreements to tricky judgment calls", and that it uses purchased data to build reinforcement learning gyms where AI agents practice real tasks. ## What do AI teams use workplace communications for? AI teams use workplace communications mainly to train and evaluate agents that work inside business tools. A thread that runs from request to approval to delivery gives a model a worked example of a multi-step task with a known outcome. - Agent training: turning real request and reply chains into [agent trajectories](https://sellmydatatoai.net/glossary/agent-trajectory) and tasks - Assistant fine-tuning: drafting, summarizing and triaging messages in a company's own register - Evaluation: testing whether a model escalates, declines or asks for clarification at the right moment - Environment building: seeding simulated inboxes and channels for an [RL environment](https://sellmydatatoai.net/glossary/rl-environment) ## How much does workplace communications data cost? Workplace communications data has no list price, and most deals are private. The public figures that exist show a wide spread, driven by archive size and by how much context survives de-identification. | Reported deal | What was licensed | Reported value | Source | | --- | --- | --- | --- | | Closed crypto startup, 2026 | Workspace data in Google, Notion, Figma and Slack | About $15,000, pending at the time of reporting | MIT Technology Review, October 2026 | | Spirit Airlines, August 2026 | 100 million emails, 500 million Teams messages and 176,000 employee records | $10 million winning bid by Google | MIT Technology Review, October 2026 | | Closed startups generally | Code and workspace data | Typically $10,000 to $100,000 per company | Fortune, September 2026, citing SimpleClosure | micro1 lists internal communications among the operational records it licenses, with published company payouts starting at $100,000 to $500,000 for a qualified enterprise partnership. Those are payouts to data owners rather than prices to buyers, but they set what sellers expect. Our guide to [AI training data pricing models](https://sellmydatatoai.net/guides/ai-training-data-pricing-models) compares the common deal structures. ## What rights issues come with workplace communications? Workplace communications carry more third-party rights than most data types. Daniel Goldberg of Frankfurt Kurnit Klein and Selz notes that email and chat often hold other companies' confidential information covered by NDAs and commercial agreements, which can limit what a seller is free to license. Employee messages raise a second issue in the US. Goldberg points out that messages about pay, scheduling and working conditions may be protected concerted activity under Section 7 of the National Labor Relations Act, whether or not staff are unionized. State privacy law applies to staff too. The California Privacy Protection Agency says the CCPA covers California residents who are employees or job applicants, and that the exemptions for employment-related personal information expired on December 31, 2022. If you hold UK or EU data, the messages are personal data, so the seller needs a lawful basis under UK GDPR or EU GDPR, or must anonymize them. - Privacy notices: the seller must follow the notice in force when the messages were written - HR records: suppliers interviewed by MIT Technology Review leave them out - Client content: privileged or confidential client material removed, not just masked - Withdrawal: Goldberg warns that data used to train a model cannot practically be pulled back out ## What should you check before buying workplace communications data? Before buying workplace communications data, check that the seller's vetting covers the rights above and that the sample reflects the whole archive. Every seller we shortlist passes our [eight-point vetting standard](https://sellmydatatoai.net/how-we-vet), including contract restrictions, the privacy notice and a re-identification test on a sample. Ask for a [representative sample](https://sellmydatatoai.net/glossary/representative-sample) drawn at random across years and teams, and compare it with the [data manifest](https://sellmydatatoai.net/glossary/data-manifest). Our [due diligence checklist](https://sellmydatatoai.net/guides/ai-training-data-due-diligence-checklist) lists the questions to put to any seller. To source communications archives that have already passed these checks, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer) and tell us which workflows you are training for. ## Frequently asked questions ### Can I buy Slack data for AI training? Slack data can be licensed for AI training when the company that owns the workspace holds the rights to share it and has removed personal and third-party confidential content. The license comes from that company, not from Slack. Check the seller's privacy notice, client contracts and de-identification method before you sign. ### Is workplace email data de-identified before sale? Workplace email data should be de-identified before any buyer receives it, with names, addresses, signatures and other direct identifiers removed and indirect identifiers reviewed. The CCPA treats data as deidentified only if the business takes reasonable measures, publicly commits not to re-identify it and binds recipients by contract. Under UK GDPR, pseudonymized data is still personal data. Ask whether a re-identification test was run. ### Why do AI labs want internal company emails? AI labs want internal company emails because they record real multi-step work with outcomes attached: requests, approvals, disputes and fixes. MIT Technology Review reported in October 2026 that Surge AI uses such archives to build reinforcement learning gyms where AI agents practice real tasks. ### Are HR records included in communications datasets? HR records are normally left out of communications datasets. Suppliers interviewed by MIT Technology Review in October 2026 said they exclude HR records, and our seller vetting checks that personal data the seller cannot lawfully disclose is removed before any sample is shared. ## Sources - [AI labs are buying data from dead startups to train their models](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models), MIT Technology Review, accessed November 2026 - [Little-known AI startup Micro1 tries to trump Google's bid for bankrupt Spirit Airlines' data](https://fortune.com/2026/09/14/micro1-tries-to-beat-googles-bid-to-buy-bankrupt-spirit-airlines-data-for-12-5-million-highlighting-value-of-data-to-ai-companies/), Fortune, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 - [Monetize your company's operational data](https://www.micro1.ai/data), micro1, accessed November 2026 - [Frequently asked questions](https://cppa.ca.gov/faq.html), California Privacy Protection Agency, accessed November 2026 - [California Civil Code section 1798.140 (CCPA definitions)](https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?sectionNum=1798.140.&lawCode=CIV), California Legislative Information, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Buy customer support ticket data for AI training URL: https://sellmydatatoai.net/buy-ai-training-data/support-ticket-data Customer support ticket data is the record of real customer problems and how agents resolved them, licensed from helpdesk systems to train and evaluate support assistants and AI agents. micro1 lists support tickets among the operational records it licenses, with company payouts from $100,000. The value sits in resolved threads with outcomes, and the main risk is customer personal data inside the text. ## Key takeaways - AI teams use customer support ticket data for (1) fine-tuning customer support assistants, (2) training ticket triage and routing models and (3) agent training on resolution workflows. - Usually priced per ticket archive, as a flat license fee. - Common delivery formats: Helpdesk exports in JSON or CSV, JSONL conversation threads and linked knowledge base articles in Markdown or HTML. - Check before you license customer support ticket data: (1) US: CCPA and other state privacy laws for customer personal information, (2) US: privacy notice in force when customers wrote in, which the FTC expects companies to honor and (3) customer personal data removed from ticket text and attachments. - Every seller we shortlist passes 8 vetting checks before you see a sample. ## Buyer's spec - Used for: Fine-tuning customer support assistants; Training ticket triage and routing models; Agent training on resolution workflows; Evaluation sets built from resolved tickets with known outcomes - Formats: Helpdesk exports in JSON or CSV; JSONL conversation threads; Linked knowledge base articles in Markdown or HTML; Parquet for large ticket histories - Usually priced: Per ticket archive, as a flat license fee - Check before you sign: US: CCPA and other state privacy laws for customer personal information; US: privacy notice in force when customers wrote in, which the FTC expects companies to honor; Customer personal data removed from ticket text and attachments; Payment card data and credentials excluded; Customer terms permit the use; Resolution status and outcome fields present; Knowledge base articles are the seller's own content; Ticket counts and date range match the manifest; UK and EU: lawful basis under UK GDPR or EU GDPR, or anonymization ## What is customer support ticket data? Customer support ticket data is the history of customer requests held in a helpdesk: the opening message, every agent and customer reply, internal notes, tags, priority, resolution status and satisfaction ratings. The most useful archives keep the link between each ticket and the knowledge base article or fix that closed it. micro1's data partnerships page names support tickets alongside SOPs, project histories, CRM records and internal communications as operational data it licenses from companies with 30 or more full-time employees. ## What do AI teams use support ticket data for? AI teams use support ticket data to train assistants that answer customers and agents that take actions to resolve their problems. A resolved ticket is a worked example with a starting state, a sequence of steps and a measurable result. - [Supervised fine-tuning](https://sellmydatatoai.net/glossary/supervised-fine-tuning) on agent replies that customers rated highly - Triage and routing models trained on tags, priority and the team that solved each ticket - Agent training where a model must look up an order, apply a policy and close the case - Evaluation sets of tickets with known correct resolutions, kept as a [hold-out set](https://sellmydatatoai.net/glossary/hold-out-set) ## How much does support ticket data cost? Support ticket data has no published buyer price list. The clearest public signals are the payouts that training-data companies advertise to the companies whose records they license. | Buyer | Published company payout | Relevance to support data | Source | | --- | --- | --- | --- | | micro1 | $100,000 to $500,000 for a qualified partnership, $500,000 to $3 million for large-scale or ongoing datasets | Lists support tickets and CRM records | micro1 data partnerships page, November 2026 | | Mercor | $20,000 to $250,000 for call centers with 50 to 5,000 agents | Licenses recorded customer conversations | Mercor listing via ddiy.co, October 2026 | These are payouts to data owners. A buyer's total cost also includes cleaning, redaction review and the work of [evaluating dataset samples](https://sellmydatatoai.net/guides/how-to-evaluate-dataset-samples) before signing. ## What are the privacy risks in support ticket data? Support ticket data is full of customer personal data, often typed into free text where automated redaction misses it. Names, email addresses, order numbers, postal addresses and screenshots all need handling, and anything that remains must be disclosed in line with the privacy laws that apply, such as the CCPA for California customers. Card details are a specific hazard. The PCI Security Standards Council states that sensitive authentication data, which includes card verification codes, must not be stored after authorization even if encrypted, so any ticket containing it should be removed rather than masked. The seller must also respect the privacy notice that applied when customers wrote in. FTC staff warned in February 2024 that adopting more permissive data practices, such as using data for AI training, through a surreptitious, retroactive amendment to a privacy policy may be unfair or deceptive. Under the CCPA, at California Civil Code section 1798.140(ad), making a consumer's personal information available to a third party for monetary or other valuable consideration is a sale, unless the data meets the statute's deidentified standard. If you hold UK or EU tickets, UK GDPR or EU GDPR requires a lawful basis for the disclosure or proper anonymization. ## How should support ticket data be delivered? Support ticket data is easiest to use when delivered as one JSON or JSONL record per ticket, with messages in order, timestamps, author roles such as customer, agent or bot, tags, priority, status and satisfaction score. Internal notes belong in a separate field so they can be included or excluded. Ask for linked knowledge base articles with stable IDs, and for a field showing whether each reply came from a person or an automated system. Replies from an old chatbot teach a new model to imitate that chatbot. ## What should you check before buying support ticket data? Before buying support ticket data, confirm that outcome fields survived the export and that redaction was tested on free text and attachments, not only structured fields. Ask for a random sample with the redaction method documented, then check it against the [data manifest](https://sellmydatatoai.net/glossary/data-manifest). Our [due diligence checklist](https://sellmydatatoai.net/guides/ai-training-data-due-diligence-checklist) covers the contract terms, and [de-identification for AI training data](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data) explains the methods to expect. To source vetted ticket archives, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer). ## Frequently asked questions ### Can a company sell its support tickets to AI companies? A company can license its support tickets to AI companies if its customer terms and privacy notice allow the disclosure, or if customer personal data is removed to an anonymization standard. micro1 names support tickets among the operational records it licenses. Client contracts can also restrict reuse, so they need checking first. ### Do support tickets count as personal data? Support tickets usually contain personal information, because customers include names, email addresses, order numbers and account details in their messages. Under California's CCPA they stay personal information unless deidentified to the standard in Civil Code section 1798.140(m). Under UK GDPR and EU GDPR they remain personal data until no one can reasonably identify the customer, and pseudonymized tickets still count. ### What makes a support ticket dataset valuable? A support ticket dataset is valuable when each ticket keeps its full thread, its resolution status and the article or action that fixed the problem. That turns tickets into worked examples with known outcomes, which buyers can use for fine-tuning, agent training and evaluation sets. ### What format should support ticket data be delivered in? Support ticket data is best delivered as JSON or JSONL threads exported from the helpdesk, with one record per ticket and fields for messages, timestamps, tags, status and satisfaction rating. Large archives can be supplied as Parquet. Linked knowledge base articles should come as Markdown or HTML with stable IDs. ## Sources - [Monetize your company's operational data](https://www.micro1.ai/data), micro1, accessed November 2026 - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 - [Why is storage of sensitive authentication data after authorization not permitted?](https://www.pcisecuritystandards.org/faq/articles/Frequently_Asked_Question/For-PCI-DSS-why-is-storage-of-sensitive-authentication-data-SAD-after-authorization-not-permitted-even-when-there-are-no-primary-account-numbers-PANs-in-an-environment), PCI Security Standards Council, accessed November 2026 - [AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive](https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/02/ai-other-companies-quietly-changing-your-terms-service-could-be-unfair-or-deceptive), Federal Trade Commission, accessed November 2026 - [California Civil Code section 1798.140 (CCPA definitions)](https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?sectionNum=1798.140.&lawCode=CIV), California Legislative Information, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Buy source code datasets for AI training URL: https://sellmydatatoai.net/buy-ai-training-data/source-code-datasets Source code datasets are licensed private codebases, ideally with their commit history, pull requests, issues and runbooks, used to train and evaluate coding models and agents. DataVendor says labs pay more for code written mostly by people and for systems languages such as Rust, Go, C and C++. MIT Technology Review reported one closed startup's codebase selling for about $25,000. ## Key takeaways - AI teams use source code datasets for (1) pre-training and fine-tuning coding models, (2) training software engineering agents on real repository tasks and (3) building evaluation tasks from issues and the commits that fixed them. - Usually priced per repository or codebase, as a flat license fee. - Common delivery formats: Git repository bundles with full history, ZIP or tar archives of a fixed snapshot and JSONL of pull requests, reviews and issues. - Check before you license source code datasets: (1) US copyright: employee code owned as work made for hire, contractor code assigned in a signed writing, (2) open-source components and their licenses identified and (3) secrets, keys and credentials removed from code and history. - Every seller we shortlist passes 8 vetting checks before you see a sample. ## Buyer's spec - Used for: Pre-training and fine-tuning coding models; Training software engineering agents on real repository tasks; Building evaluation tasks from issues and the commits that fixed them; Code review and debugging models trained on pull request history - Formats: Git repository bundles with full history; ZIP or tar archives of a fixed snapshot; JSONL of pull requests, reviews and issues; Runbooks and incident notes in Markdown - Usually priced: Per repository or codebase, as a flat license fee - Check before you sign: US copyright: employee code owned as work made for hire, contractor code assigned in a signed writing; Open-source components and their licenses identified; Secrets, keys and credentials removed from code and history; Client-owned code excluded or cleared; Share of AI-generated code disclosed; Commit history intact and matching the manifest ## What is a source code dataset? A source code dataset is a licensed copy of private software: the code itself plus, in the stronger packages, the history that explains it. DataVendor describes production code as private examples of how developers build, test and maintain working software, and notes that public repositories rarely preserve the full operational record. Context adds value. DataVendor lists pull requests, issue tickets, runbooks and incident notes as code-adjacent material that explains why the code changed. ## What do labs pay most for in source code? Labs pay most for source code written mainly by people, in systems languages, from large engineering teams. DataVendor states that labs pay more for human-written code and for Rust, Go, C and C++, and that high-quality codebases from large engineering teams sell to the most lab buyers on its platform. | Factor | What buyers prefer | | --- | --- | | Authorship | Code written mostly by people | | Language | Systems languages such as Rust, Go, C and C++ | | Team | Codebases from large engineering teams | | Context | Pull requests, issue tickets, runbooks and incident notes | | History | The full operational record, which public repositories rarely keep | ## How much does a source code dataset cost? Source code datasets range from tens of thousands of dollars to the high six figures, depending on size, history and language. The public figures are few and come mostly from the supply side. - About $25,000 for a closed crypto startup's codebase, reported by MIT Technology Review in October 2026 - $50,000 to $500,000 for software companies with 15 to 500 employees, from Mercor's published listing for private code and its history - $20,000 to $890,000 for design and web agencies' design systems, shipped websites and the code behind them, also from Mercor's listing MIT Technology Review also reported that buyers' valuations rest on research theses estimating how many useful tasks a body of code might yield. When you price a codebase, ask the seller how many issues have linked fixes and tests, since that is what a task estimate counts. ## What rights problems affect source code datasets? Source code datasets fail diligence most often over ownership and open-source licenses. Under 17 USC 201(b), code an employee writes within the scope of employment is a work made for hire owned by the employer. Code from contractors usually needs a transfer signed by the contractor under 17 USC 204(a), because the US Copyright Office's Circular 30 limits commissioned works made for hire to nine listed categories with a signed written agreement. The seller must also identify open-source components whose licenses impose conditions. If you are buying from a UK seller, section 11 of the Copyright, Designs and Patents Act 1988 likewise leaves contractors with copyright unless they assign it in writing. - Copyleft components such as GPL code, which carry their own license conditions - Client-owned code written under services contracts - Vendored third-party libraries and SDKs under commercial terms - Secrets in commit history, which survive deletion from the current code Every seller we shortlist passes our [vetting standard](https://sellmydatatoai.net/how-we-vet), which includes chain of title and a check that third-party content and open-source code are identified and either cleared or removed. ## How should a source code dataset be delivered? A source code dataset should be delivered as a fixed export: a git bundle or archive of each repository with its full history, plus a separate export of pull requests, reviews and issues from the code hosting platform. Frankfurt Kurnit Klein and Selz advises against giving buyer tools direct access to company systems, and that applies to code hosts as much as to email. Ask for a manifest per repository listing languages, lines of code, commit count, date range and number of contributors, so you can check the delivery against what you paid for. ## What should you check before buying source code? Before buying source code, run a license scan on a sample, confirm the commit history matches the manifest and ask what share of the code was generated by AI tools. On DataVendor's account of buyer preferences, a history full of tool-generated commits is worth less. Our guide on [how to sell source code to AI companies](https://sellmydatatoai.net/guides/sell-source-code-to-ai-companies) shows what sellers prepare, and [agent trajectories and RL environments](https://sellmydatatoai.net/guides/agent-trajectories-and-rl-environments) covers turning repositories into tasks. Software firms can read about [selling software company data](https://sellmydatatoai.net/sell-your-data/software-companies). To source vetted codebases, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer). ## Frequently asked questions ### Can AI companies buy private source code? AI companies can buy licenses to private source code from the companies that own it. DataVendor runs a marketplace for code repositories, and Mercor publishes payouts of $50,000 to $500,000 for software companies with 15 to 500 employees. Under US copyright law the seller must own employee code as work made for hire and hold signed assignments for contractor code. ### Does AI-generated code reduce a codebase's value? AI-generated code does reduce a codebase's value to most training buyers. DataVendor states that labs pay more for code written mostly by people, so ask the seller to estimate the share of tool-generated code and the date from which AI coding tools were used across the team. ### What about open-source code inside a private repository? Open-source code inside a private repository must be identified before sale, because its license still applies. Copyleft licenses such as the GPL impose conditions on distribution. Our seller vetting checks that third-party content and open-source code are identified and either cleared or removed from the dataset. ### Is commit history worth including in a code dataset? Commit history is worth including in a code dataset because it shows how and why the code changed. DataVendor notes that public repositories rarely preserve the full operational record, and that pull requests, issues, runbooks and incident notes add the context labs value. ## Sources - [AI training data marketplaces](https://datavendor.ai/blog/ai-training-data-marketplaces), DataVendor, accessed November 2026 - [AI labs are buying data from dead startups to train their models](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models), MIT Technology Review, accessed November 2026 - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 - [17 U.S. Code section 201: ownership of copyright](https://www.law.cornell.edu/uscode/text/17/201), Legal Information Institute, Cornell Law School, accessed November 2026 - [17 U.S. Code section 204: execution of transfers of copyright ownership](https://www.law.cornell.edu/uscode/text/17/204), Legal Information Institute, Cornell Law School, accessed November 2026 - [Works Made for Hire (Circular 30)](https://www.copyright.gov/circs/circ30.pdf), US Copyright Office, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Buy agent trajectory data and RL environments URL: https://sellmydatatoai.net/buy-ai-training-data/agent-trajectory-data Agent trajectory data records each step an AI agent or a human expert takes to complete a task: the observations, clicks, tool calls and the result. Labs use it to train and evaluate agents, often alongside a replayable environment. DataVendor reports that post-training work now creates stronger demand for evals, environments and real workflow traces. ## Key takeaways - AI teams use agent trajectory data for (1) supervised fine-tuning of computer-use and tool-using agents, (2) reinforcement learning with replayable environments and (3) reward model and verifier training on successful and failed runs. - Usually priced per task or trajectory, or per environment with its taskset. - Common delivery formats: JSONL step logs with observation, action and result, screenshots or screen recordings paired with action logs and containerized environments such as Docker images. - Check before you license agent trajectory data: (1) each trajectory has a verified outcome, (2) environment replays reproduce the recorded result and (3) US: personal information in screenshots and observations removed under the CCPA and other state privacy laws. - Every seller we shortlist passes 8 vetting checks before you see a sample. ## Buyer's spec - Used for: Supervised fine-tuning of computer-use and tool-using agents; Reinforcement learning with replayable environments; Reward model and verifier training on successful and failed runs; Evaluation of agents on held-out tasks - Formats: JSONL step logs with observation, action and result; Screenshots or screen recordings paired with action logs; Containerized environments such as Docker images; Task definitions with scoring functions - Usually priced: Per task or trajectory, or per environment with its taskset - Check before you sign: Each trajectory has a verified outcome; Environment replays reproduce the recorded result; US: personal information in screenshots and observations removed under the CCPA and other state privacy laws; Rights to the software inside the environment; Split between human and model-generated runs disclosed; No overlap with public benchmarks ## What is agent trajectory data? Agent trajectory data is a step-by-step log of an agent completing a workflow, linking each action to what the agent saw and what happened next. DataVendor defines trajectories as records of the steps an AI agent takes while completing a workflow, and notes that they connect model actions to measurable outcomes. A trajectory can come from a model run or from a person doing the task while every action is logged. OpenAI's 2022 InstructGPT paper is an early, well-documented example of training on human demonstrations, although its demonstrations were written answers rather than tool use. The glossary defines an [agent trajectory](https://sellmydatatoai.net/glossary/agent-trajectory) in brief. ## What is the difference between a trajectory and an RL environment? A trajectory is a recording, and an RL environment is the sandbox in which the task can be run again. DataVendor describes replayable environments as the software or sandbox needed to run the task again, and on its marketplace environments cannot be listed on their own, only as tasksets with their environments. | Asset | What it contains | Main use | | --- | --- | --- | | Trajectory | Observations, actions and outcome for one run | Supervised fine-tuning and reward modeling | | Taskset | Task prompts with scoring rules or expected results | Evaluation and RL rewards | | RL environment | Runnable software or sandbox for the tasks | Reinforcement learning at scale | Our guide to [agent trajectories and RL environments](https://sellmydatatoai.net/guides/agent-trajectories-and-rl-environments) covers each asset in more depth, and the glossary defines an [RL environment](https://sellmydatatoai.net/glossary/rl-environment). ## Why is demand for agent trajectory data rising? Demand for agent trajectory data is rising because post-training has become a main route to better agents. DataVendor states that post-training work now creates stronger demand for evals, environments and real workflow traces. Real company archives feed this too. MIT Technology Review reported in October 2026 that Surge AI buys company data to build reinforcement learning gyms where clients' agents practice tasks such as debugging complex IT architecture or drafting appellate briefs. The same article noted that no company has published tests comparing real-world and alternative training data, so the size of the gain is still unclear. ## Where does agent trajectory data come from? Agent trajectory data can come from logs of AI products where agents already act for users, from expert demonstrations recorded for the purpose, or from company records rebuilt into tasks. MIT Technology Review reported that buyers value a codebase by estimating how many useful tasks it might yield, and the same reasoning applies to ticket queues and approval workflows. Product logs need care. Prompt and response logs, agent run traces and user ratings all describe agent behavior, but they belong to the product's users as much as to the company, so they need a privacy notice that covers the use or proper de-identification before licensing. California residents can also opt out of the sale or sharing of their personal information under the CCPA, according to the California Attorney General. ## How is agent trajectory data priced? Agent trajectory data is priced per task, per trajectory or as a bundle of environment and taskset, depending on the seller. There is no public price index, and DataVendor warns that a listed price or a large catalog does not establish what an asset will earn. Ask whether a quote covers failed runs as well as successful ones. Failed runs with a recorded reason can be useful for training verifiers, so check whether the seller kept them. Our guide to [AI training data pricing models](https://sellmydatatoai.net/guides/ai-training-data-pricing-models) compares per-unit and flat-fee structures. ## What should you check before buying agent trajectories? Before buying agent trajectories, replay a random sample in the environment and confirm the recorded outcome. A trajectory that does not reproduce cannot be trusted as a training signal. - Outcome verification method, automatic or human - Share of trajectories produced by models rather than people - Personal data in screenshots, form fields and observations - Licenses for any commercial software inside the environment - Overlap with public benchmarks that would contaminate evaluation To source vetted trajectory data and environments, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer). Our [due diligence checklist](https://sellmydatatoai.net/guides/ai-training-data-due-diligence-checklist) covers the contract side, and [evaluation datasets](https://sellmydatatoai.net/buy-ai-training-data/evaluation-datasets) explains how to keep test tasks clean. ## Frequently asked questions ### What does an agent trajectory contain? An agent trajectory contains the sequence of observations, actions and results from one attempt at a task. For a computer-use agent that means screenshots or page states, each click, keystroke or tool call, and the final outcome. Good trajectories also record whether the outcome was verified and how. ### Can company workflows be turned into agent training data? Company workflows can be turned into agent training data when the records show each step and its result, for example ticket histories, code changes with tests or approval chains in email. MIT Technology Review reported that Surge AI uses purchased company data to build reinforcement learning gyms for agents. ### Are human or model-generated trajectories better? Human and model-generated trajectories serve different purposes. Human demonstrations show expert behavior a model cannot yet produce, while model runs with verified outcomes supply volume for reinforcement learning. Ask the seller to disclose the split, since buyers pay for different things in each. ### Can an RL environment be bought on its own? An RL environment is usually bought together with tasks to run in it. DataVendor states that environments cannot be listed on its marketplace on their own, only as tasksets with their environments, because an environment without tasks and scoring rules gives an agent nothing to learn from. ## Sources - [AI training data marketplaces](https://datavendor.ai/blog/ai-training-data-marketplaces), DataVendor, accessed November 2026 - [AI labs are buying data from dead startups to train their models](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models), MIT Technology Review, accessed November 2026 - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155), arXiv (Ouyang et al., 2022), accessed November 2026 - [California Consumer Privacy Act (CCPA)](https://oag.ca.gov/privacy/ccpa), State of California Department of Justice, Office of the Attorney General, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Buy private evaluation datasets for AI models URL: https://sellmydatatoai.net/buy-ai-training-data/evaluation-datasets Evaluation datasets are sets of tasks with expected answers and scoring rules, kept out of training so they measure what a model can really do. Their value depends on staying private: a 2024 study by Hugh Zhang and colleagues found accuracy drops of up to 8% when leading models were tested on fresh grade school maths problems instead of a public benchmark. ## Key takeaways - AI teams use evaluation datasets for (1) measuring model capability on private tasks, (2) regression testing between model versions and (3) grading agents with automatic or rubric scoring. - Usually priced per task or per set, with exclusive access priced higher. - Common delivery formats: JSONL of prompts, reference answers and metadata, rubrics and scoring scripts and containerized tasks with unit tests. - Check before you license evaluation datasets: (1) tasks never published or shared with other buyers, (2) reference answers checked by a second expert and (3) scoring rules written down and reproducible. - Every seller we shortlist passes 8 vetting checks before you see a sample. ## Buyer's spec - Used for: Measuring model capability on private tasks; Regression testing between model versions; Grading agents with automatic or rubric scoring; Procurement tests comparing vendors' models - Formats: JSONL of prompts, reference answers and metadata; Rubrics and scoring scripts; Containerized tasks with unit tests; Inter-annotator agreement records - Usually priced: Per task or per set, with exclusive access priced higher - Check before you sign: Tasks never published or shared with other buyers; Reference answers checked by a second expert; Scoring rules written down and reproducible; Exclusivity and non-disclosure terms in the license; US copyright in each task held by the seller, as work made for hire or by signed assignment; Difficulty spread across the set; Author and provenance of each task ## What is an evaluation dataset? An evaluation dataset is a fixed set of tasks, each with an expected answer and a rule for scoring a model's response. It is used to measure a model, not to train it, so it must be kept apart from training data as a [hold-out set](https://sellmydatatoai.net/glossary/hold-out-set). DataVendor lists evaluation datasets among the assets labs buy and states that post-training work now creates stronger demand for evals. The glossary entry on the [evaluation dataset](https://sellmydatatoai.net/glossary/evaluation-dataset) gives a short definition. ## Why does an evaluation dataset need to stay private? An evaluation dataset needs to stay private because once its tasks appear in training data, scores stop measuring ability and start measuring memory. Researchers call this contamination. The effect can be measured. In a 2024 study, Hugh Zhang and colleagues built GSM1k to mirror the public GSM8k maths benchmark and observed accuracy drops of up to 8% for leading models, with several model families showing evidence of systematic overfitting across almost all model sizes. ## What makes a good evaluation dataset? A good evaluation dataset is hard, unambiguous and checked by experts. OpenAI's SWE-bench Verified, published in August 2024, shows the effort involved: OpenAI worked with 93 software developers experienced in Python to screen tasks and kept 500 samples verified as non-problematic. Rubric-scored tasks need the rubric itself to be tested. Two experts scoring the same responses should reach similar marks, and the seller should report how often they agreed. | Property | What to look for | | --- | --- | | Ambiguity | Each task has one defensible answer or a written rubric | | Verification | Reference answers checked by a second expert | | Difficulty | A spread of tasks, not only easy or impossible ones | | Coverage | Tasks that match the deployment you care about | | Freshness | No overlap with public benchmarks or web text | ## How are evaluation datasets priced? Evaluation datasets are priced per task or per set, and exclusivity weighs heavily because a shared evaluation set loses value each time it is shared. ddiy.co's 2026 pricing guide puts an exclusive deal at about 2.5 to 5 times a non-exclusive one across company data. Our guide to [exclusive vs non-exclusive data licenses](https://sellmydatatoai.net/guides/exclusive-vs-non-exclusive-data-licenses) sets out how the terms differ, and [AI training data pricing models](https://sellmydatatoai.net/guides/ai-training-data-pricing-models) compares per-task and flat-fee structures. ## Who builds evaluation datasets? Evaluation datasets are built by AI labs in-house, by training-data companies that recruit domain experts, and by companies whose own records contain tasks with known answers. A resolved support ticket, a closed legal matter or a fixed software bug each pairs a problem with a verified outcome. Company records need conversion before they work as evaluations. Each task must be rewritten to stand alone, the expected answer separated out and a scoring rule written. Ask the seller who did that work and whether the original records were also licensed to anyone for training. If outside experts wrote the tasks, check for signed copyright assignments, because the US Copyright Office's Circular 30 treats commissioned work as made for hire only in nine listed categories, and only with a signed written agreement. ## How should evaluation data be delivered? Evaluation data should be delivered with tasks, reference answers and scoring code in separate files, each task carrying a stable ID, author, difficulty label and creation date. For agent tasks, the environment and tests should come in a container that reproduces the scoring on your own systems. ## What should you check before buying an evaluation dataset? Before buying an evaluation dataset, ask who wrote each task, whether any task has been shown to another buyer and how the reference answers were checked. Search a sample of tasks against public benchmarks and the open web for overlap. Then fix the terms: the exclusivity period, a ban on the seller reusing tasks and what happens if a leak is found. Read how to [evaluate a dataset sample](https://sellmydatatoai.net/guides/how-to-evaluate-dataset-samples), or [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer) to source vetted evaluation sets. ## Frequently asked questions ### What is the difference between training data and evaluation data? Training data is what a model learns from, while evaluation data is what it is tested on. Evaluation data must be kept out of training, because a model that has seen the test tasks scores well by memory. That is why evaluation sets are licensed with strict exclusivity and non-disclosure terms. ### How do I know an evaluation set has not leaked? Knowing an evaluation set has not leaked starts with the seller's records of who has seen it. Then search a sample of tasks against public benchmarks and web text. The 2024 GSM1k study by Hugh Zhang and colleagues showed accuracy drops of up to 8% on fresh problems, a sign of contamination. ### Who writes good evaluation tasks? Good evaluation tasks are written by domain experts and checked by a second expert. OpenAI's SWE-bench Verified used 93 software developers experienced in Python to screen tasks for well-specified issues and appropriate tests, keeping 500 verified samples. Ask sellers for author credentials and review records. ### Should I buy an evaluation set exclusively? An evaluation set bought exclusively keeps its value longer, because every additional licensee raises the chance of leakage into training data. Exclusivity costs more: ddiy.co's 2026 pricing guide puts exclusive deals at about 2.5 to 5 times non-exclusive ones. Agree the exclusivity period and remedies in writing. ## Sources - [A careful examination of large language model performance on grade school arithmetic](https://arxiv.org/abs/2405.00332), arXiv (Zhang et al., 2024), accessed November 2026 - [Introducing SWE-bench Verified](https://openai.com/index/introducing-swe-bench-verified/), OpenAI, accessed November 2026 - [AI training data marketplaces](https://datavendor.ai/blog/ai-training-data-marketplaces), DataVendor, accessed November 2026 - [How much is company data worth to AI? (2026 prices)](https://ddiy.co/how-much-is-company-data-worth-to-ai/), ddiy.co, accessed November 2026 - [Works Made for Hire (Circular 30)](https://www.copyright.gov/circs/circ30.pdf), US Copyright Office, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Buy RLHF and preference data for model alignment URL: https://sellmydatatoai.net/buy-ai-training-data/rlhf-preference-data RLHF and preference data are human judgments on model outputs, usually pairs of responses with one marked as better, used to train reward models and align models with human preferences. OpenAI's 2022 InstructGPT paper showed the effect: outputs from a 1.3 billion parameter model trained this way were preferred to those of the 175 billion parameter GPT-3. ## Key takeaways - AI teams use rlhf and preference data for (1) reward model training, (2) reinforcement learning from human feedback and (3) direct preference optimization and similar methods. - Usually priced per comparison or per rated prompt. - Common delivery formats: JSONL with prompt, chosen and rejected responses, ranked lists of several responses per prompt and rubric scores with rater IDs and rationales. - Check before you license rlhf and preference data: (1) rater qualifications and guidelines supplied, (2) agreement between raters measured and (3) model that produced the responses and its terms. - Every seller we shortlist passes 8 vetting checks before you see a sample. ## Buyer's spec - Used for: Reward model training; Reinforcement learning from human feedback; Direct preference optimization and similar methods; Safety tuning with labeled red-team data - Formats: JSONL with prompt, chosen and rejected responses; Ranked lists of several responses per prompt; Rubric scores with rater IDs and rationales; Rater guidelines and calibration records - Usually priced: Per comparison or per rated prompt - Check before you sign: Rater qualifications and guidelines supplied; Agreement between raters measured; Model that produced the responses and its terms; Rater consent and pay records; Offensive content flagged and labeled; No duplication with public preference datasets ## What is RLHF and preference data? Preference data is a record of human choices between model outputs: for a given prompt, which response is better, or how a set of responses ranks. RLHF, reinforcement learning from human feedback, uses those choices to train a reward model and then tunes the main model against it. The glossary covers [RLHF](https://sellmydatatoai.net/glossary/rlhf) and [preference data](https://sellmydatatoai.net/glossary/preference-data) in brief. The InstructGPT paper by Long Ouyang and colleagues at OpenAI, published in March 2022, describes the standard recipe: collect labeler demonstrations for [supervised fine-tuning](https://sellmydatatoai.net/glossary/supervised-fine-tuning), then collect rankings of model outputs to fine-tune further with RLHF. ## What does a preference dataset look like? A preference dataset is usually a file of prompts with paired responses. Anthropic's public HH-RLHF dataset on Hugging Face is a clear example: each line of its JSONL files holds a pair of texts, one chosen and one rejected, and it is released under the MIT license. The same dataset card shows why documentation matters. Anthropic states that the preference data is meant to train preference or reward models and is not meant for supervised training of dialogue agents, because training dialogue agents on it is likely to lead to harmful models. | Format | What raters do | Typical use | | --- | --- | --- | | Pairwise | Pick the better of two responses | Reward models | | Ranked | Order several responses to one prompt | Reward models with more comparisons per prompt | | Rubric scored | Score responses against written criteria | Evaluation and targeted tuning | | Edited | Rewrite a response into a better one | Supervised fine-tuning | ## What makes preference data good quality? Good preference data comes from qualified raters following written guidelines, with agreement between raters measured. For expert domains such as law or medicine, the rater's credentials matter as much as the volume of judgments. Ask for the rater guidelines, a calibration record and agreement scores. Low agreement is not always a fault, since some prompts are genuinely contested, but the seller should know which ones and flag them. ## How is preference data priced? Preference data is priced per comparison or per rated prompt. No buyer publishes a rate card, so treat any quote as a starting point and compare it with the cost of commissioning the same work from raters you recruit yourself. Our guide to [AI training data pricing models](https://sellmydatatoai.net/guides/ai-training-data-pricing-models) sets out the common structures, including per-unit pricing and fixed project fees. ## Who supplies preference data? Preference data is supplied by training-data companies that run rater pools, by AI product companies that collect user ratings, and by firms whose experts review model output as part of their work. Profiles of training-data companies such as [Scale AI](https://sellmydatatoai.net/data-buyers/scale-ai) and [Mercor](https://sellmydatatoai.net/data-buyers/mercor) are on our data buyers pages. User ratings from AI products, such as thumbs up and thumbs down on responses, are preference data too. They are less controlled than rater judgments and usually tied to user accounts, so they need anonymization and a privacy notice that covers the use before licensing. ## How should preference data be delivered? Preference data should be delivered as JSONL with one comparison per line, including the prompt, each response, the rater's choice, a rater ID, a timestamp and any written rationale. Rater guidelines should come as a versioned document, since a change to the guidelines changes what the labels mean. ## What should you check before buying preference data? Before buying preference data, check which model generated the responses being compared and that its terms allow its outputs to be used this way. Confirm that raters consented and were paid, and that offensive content from red-teaming is labeled as such. Check a random sample for duplication with public sets such as HH-RLHF. [Book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer) to source vetted preference data, and use our [due diligence checklist](https://sellmydatatoai.net/guides/ai-training-data-due-diligence-checklist) for the contract. ## Frequently asked questions ### What is the difference between RLHF data and fine-tuning data? RLHF data records which of several responses a person preferred, while fine-tuning data shows the single response a model should give. The InstructGPT paper used both: labeler demonstrations for supervised fine-tuning, then rankings of model outputs to train a reward model for reinforcement learning. ### Can preference data be used for supervised training? Preference data is not always suitable for supervised training. Anthropic's HH-RLHF dataset card states that its data is meant for preference or reward models and not for supervised training of dialogue agents, because that is likely to produce harmful models. Check the intended use stated by any seller. ### Who should rate responses in a preference dataset? Responses in a preference dataset should be rated by people qualified in the subject, following written guidelines. General raters suit everyday helpfulness tasks, while legal, medical or coding prompts need practitioners. Ask the seller for rater credentials, guidelines and agreement scores before you buy. ### Are public preference datasets free to use commercially? Public preference datasets carry their own licenses, so commercial use depends on each one. Anthropic's HH-RLHF is released under the MIT license, for example, but other sets restrict commercial use. Licensed private preference data gives you judgments that competitors have not trained on. ## Sources - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155), arXiv (Ouyang et al., 2022), accessed November 2026 - [Anthropic/hh-rlhf dataset card](https://huggingface.co/datasets/Anthropic/hh-rlhf), Hugging Face, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Buy expert domain data from professional firms URL: https://sellmydatatoai.net/buy-ai-training-data/expert-domain-data Expert domain data is the working record of professionals such as lawyers, accountants, engineers and clinicians: case files, workpapers, drawings and notes that show expert judgment applied to real problems. Mercor's published listings put company payouts at $50,000 to $1.2 million for law firms and $25,000 to $865,000 for accounting and tax firms. ## Key takeaways - AI teams use expert domain data for (1) fine-tuning models for professional tasks, (2) building expert evaluation sets and rubrics and (3) agent training on professional workflows. - Usually priced per company archive, as a flat license fee. - Common delivery formats: Document sets in PDF, DOCX and native formats, structured exports from practice management systems and JSONL of tasks with expert reference answers. - Check before you license expert domain data: (1) US: HIPAA de-identification for medical records, (2) US: IRC Section 7216 limits on tax return information and (3) client confidentiality, attorney-client privilege and professional duties. - Every seller we shortlist passes 8 vetting checks before you see a sample. ## Buyer's spec - Used for: Fine-tuning models for professional tasks; Building expert evaluation sets and rubrics; Agent training on professional workflows; RL environments for drafting and review tasks - Formats: Document sets in PDF, DOCX and native formats; Structured exports from practice management systems; JSONL of tasks with expert reference answers; CAD and BIM files for engineering and construction - Usually priced: Per company archive, as a flat license fee - Check before you sign: US: HIPAA de-identification for medical records; US: IRC Section 7216 limits on tax return information; Client confidentiality, attorney-client privilege and professional duties; Privileged and regulated material excluded; Credentials of authors and reviewers; Files complete from instruction to outcome; Client contracts that restrict reuse; Clients and counterparties de-identified; UK and EU: client records are personal data under UK GDPR or EU GDPR unless anonymized ## What is expert domain data? Expert domain data is the record of skilled professional work, kept by the firms that did it. It includes closed matter files at law firms, tax returns and workpapers at accountancy practices, project records and drawings at construction companies, and de-identified clinical notes at medical practices. Labs buy it to teach models complete professional tasks, from instruction to finished output. MIT Technology Review reported in October 2026 that Surge AI builds reinforcement learning gyms where agents practice tasks such as drafting appellate briefs. ## How much do AI companies pay for expert domain data? AI companies pay from tens of thousands of dollars to over $1 million per firm for expert domain data, depending on the profession and the size of the archive. Mercor's published listings, compiled by ddiy.co in October 2026, give these company payouts. | Profession | Records licensed | Company size | Published payout | | --- | --- | --- | --- | | Law firm | Complete files from closed matters and completed deals | 5 to 300 attorneys and staff | $50,000 to $1.2 million | | Accounting and tax firm | Tax returns, workpapers and bookkeeping records | 5 to 200 employees | $25,000 to $865,000 | | Construction company | Project records, drawings and site photos | 50 to 2,000 employees | $115,000 to $1 million | | Manufacturer | Production, maintenance and inspection records | 20 to 2,000 employees | $100,000 to $1.5 million | | Medical practice | De-identified patient records with physician notes | 10 or more clinicians or 2,500 or more patients | $80,000 to $820,000 | | Consulting firm | Client deliverables, models and drafts | 10 to 500 employees | $150,000 to $300,000 | These are payouts to the firms supplying data, not buyer prices, and each range spans small and large firms. The Next Web reported that micro1 has committed $1bn over 12 months to buying and licensing operational company data, a sign of the competition for good archives. Firms on the supply side can read our pages for [law firms](https://sellmydatatoai.net/sell-your-data/law-firms), [accounting and tax firms](https://sellmydatatoai.net/sell-your-data/accounting-firms) and [medical practices](https://sellmydatatoai.net/sell-your-data/medical-practices). ## What rights issues affect expert domain data? Expert domain data carries professional duties as well as privacy law. Client files sit inside a relationship governed by confidentiality, and often by contract terms that restrict reuse. Our seller vetting covers contract restrictions, the privacy notice in force when the records were made and the de-identification method, so shortlisted firms have already cleared those checks. You still need independent legal advice on your own intended use. - Medical records: HIPAA allows de-identification by expert determination or by removing 18 listed identifiers under 45 CFR 164.514, and HHS says de-identified health information is no longer protected health information - Tax files: US preparers face IRC Section 7216, which prohibits knowingly or recklessly disclosing or using tax return information - Legal files: attorney-client privilege and client confidentiality - Engineering drawings: client ownership of designs under some services contracts If you are buying UK or EU records, client files are also personal data under UK GDPR or EU GDPR unless anonymized, and health records are special category data. ## How should expert domain data be delivered? Expert domain data should be delivered as complete files grouped by matter, engagement or project, with documents in native formats where possible and a manifest listing each file's type, date and role in the workflow. Drafts add value: Mercor's consulting listing, for example, covers client deliverables and the drafts behind them. Ask for a short note of each file's outcome, such as the matter result or the audit opinion, so the record can be turned into tasks with known answers. ## What should you check before buying expert domain data? Before buying expert domain data, check that each file runs from instruction to outcome and that the authors held the relevant qualifications. Our guide to [data provenance](https://sellmydatatoai.net/guides/data-provenance-explained) lists the records that prove where each file came from. Ask how clients and counterparties were de-identified, who reviewed the result and whether any files were excluded for privilege or regulation. Every seller we shortlist passes our [eight-point vetting standard](https://sellmydatatoai.net/how-we-vet). [Book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer) to source vetted professional archives. ## Frequently asked questions ### Why do AI labs want professional firm records? AI labs want professional firm records because they show expert judgment applied to real problems from instruction to finished output. MIT Technology Review reported that Surge AI uses such data to build gyms where agents practice tasks like drafting appellate briefs. ### Can law firm files be licensed for AI training? Law firm files can be licensed for AI training only after privilege, client confidentiality and contract restrictions have been cleared, and client identities removed. Mercor publishes payouts of $50,000 to $1.2 million for complete files from closed matters at firms of 5 to 300 attorneys and staff. ### Can tax returns be used as AI training data? Tax returns can only be used as AI training data within the rules on tax return information. In the US, IRC Section 7216 prohibits preparers from knowingly or recklessly disclosing or using tax return information, with fines of up to $1,000 per violation. Take specialist advice before licensing tax files. ### How is medical data de-identified for AI training? Medical data is de-identified for AI training under HIPAA in the US by one of two methods in 45 CFR 164.514: expert determination that re-identification risk is very small, or the safe harbor method of removing 18 listed identifiers. HHS says de-identified health information is no longer protected health information. In the UK, health data is special category data under UK GDPR. ## Sources - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 - [AI labs are buying data from dead startups to train their models](https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models), MIT Technology Review, accessed November 2026 - [micro1 commits $1bn to buy company data for training AI agents](https://thenextweb.com/news/micro1-1bn-enterprise-data-ai-training-citi-hercules), The Next Web, accessed November 2026 - [Section 7216 frequently asked questions](https://www.irs.gov/node/3756), Internal Revenue Service, accessed November 2026 - [Guidance regarding methods for de-identification of protected health information](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html), US Department of Health and Human Services, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Buy speech and audio data for AI training URL: https://sellmydatatoai.net/buy-ai-training-data/speech-and-audio-data Speech and audio data is recorded human speech and sound, licensed with transcripts and speaker metadata to train speech recognition, voice and audio models. The rights questions are sharper than for text: a voiceprint is a biometric identifier under Illinois' Biometric Information Privacy Act, and California's CCPA treats biometric information processed to identify a consumer as sensitive personal information. ## Key takeaways - AI teams use speech and audio data for (1) automatic speech recognition training, (2) text-to-speech and voice model training and (3) speaker diarization across accents and conditions. - Usually priced per hour of audio, or per archive for company recordings. - Common delivery formats: WAV or FLAC lossless audio, time-aligned transcripts in JSON, SRT or TextGrid and speaker metadata such as language, accent and age band. - Check before you license speech and audio data: (1) US: BIPA written release for voiceprints, and its ban on profiting from them, (2) US: recording made lawfully, including in all-party consent states such as California and (3) consent from every recorded speaker. - Every seller we shortlist passes 8 vetting checks before you see a sample. ## Buyer's spec - Used for: Automatic speech recognition training; Text-to-speech and voice model training; Speaker diarization across accents and conditions; Audio understanding and sound event models - Formats: WAV or FLAC lossless audio; Time-aligned transcripts in JSON, SRT or TextGrid; Speaker metadata such as language, accent and age band; Consent records linked to speaker IDs - Usually priced: Per hour of audio, or per archive for company recordings - Check before you sign: US: BIPA written release for voiceprints, and its ban on profiting from them; US: recording made lawfully, including in all-party consent states such as California; Consent from every recorded speaker; Rights to any music or third-party audio; Voice cloning and biometric uses covered or excluded; Transcript accuracy measured on a sample; Speaker, language and accent balance; Personal data spoken aloud redacted; UK and EU: biometric data processed to identify a speaker is special category data ## What is speech and audio training data? Speech and audio training data is recorded sound with the metadata a model needs to learn from it: transcripts aligned to time, speaker labels, language and recording conditions. Sources include broadcast and podcast archives, commissioned recordings, meeting recordings and [call recordings](https://sellmydatatoai.net/buy-ai-training-data/call-recording-data). ## What do AI teams use speech data for? AI teams use speech data to train models that recognize speech, generate voices and understand audio. Each use needs different material. | Use | What the data needs | | --- | --- | | Speech recognition | Accurate time-aligned transcripts across accents and noise conditions | | Text-to-speech | Clean studio-quality recordings of consenting speakers | | Speaker diarization | Multi-speaker audio labeled with who spoke when | | Low-resource languages | Speech in languages poorly covered by public data | Language coverage is a real gap. Meta's No Language Left Behind project, which covers 200 languages including 150 low-resource ones, states that reliable translation data is not available for many languages. See [multilingual text data](https://sellmydatatoai.net/buy-ai-training-data/multilingual-text-data) for the text side. ## How is speech and audio data priced? Speech and audio data is priced by the hour for commissioned or catalog recordings, and by the archive for company recordings. Mercor's published listing puts call centers at $20,000 to $250,000 for recorded customer conversations from operations with 50 to 5,000 agents, as compiled by ddiy.co in October 2026. Ask whether a quote covers audio only or audio with verified transcripts, and whether consent records come with it. Our guide to [AI training data pricing models](https://sellmydatatoai.net/guides/ai-training-data-pricing-models) compares per-hour and flat-fee structures. ## What rights and consent issues affect speech data? Speech data raises consent and biometric questions that text does not. Illinois' Biometric Information Privacy Act lists a voiceprint as a biometric identifier. Section 15(b) requires a private entity to inform the person in writing and receive a written release before collecting one, and section 15(c) bars an entity holding biometric identifiers from selling, leasing, trading or otherwise profiting from them. California's CCPA defines biometric information, when used or intended to be used to establish identity, to include voice recordings from which a voiceprint can be extracted, and treats biometric information processed to uniquely identify a consumer as sensitive personal information. If you plan to train voice cloning or speaker identification models, agree the license terms for those uses separately. Recording law matters too. Federal law at 18 USC 2511(2)(d) permits a recording where one party to the communication consents, but California Penal Code section 632 makes it an offense to record a confidential communication without the consent of all parties. Recordings must have been lawfully made before they can be lawfully licensed. ### If you hold UK or EU recordings UK GDPR Article 4 defines biometric data as personal data from specific technical processing that allows or confirms the unique identification of a person. Under Article 9, biometric data processed to uniquely identify someone is special category data, which needs an additional condition for processing. ## Where does licensed speech data come from? Licensed speech data comes from broadcasters and podcast publishers, from companies that record meetings or calls, and from commissioned collections where speakers are recruited and paid to record. Commissioned data gives the cleanest consent position, because the recording was made for training from the start. ## How should speech data be delivered? Speech data should be delivered as lossless WAV or FLAC files, each with a time-aligned transcript, a speaker ID and metadata on language, accent, recording device and environment. Consent records should link to speaker IDs rather than names, so the dataset itself carries no direct identifiers. Ask the seller to state the sample rate and channel layout in the [data manifest](https://sellmydatatoai.net/glossary/data-manifest), along with total hours by language and the number of speakers. Long stretches of silence or music inflate the hours without adding training value. ## What should you check before buying speech data? Before buying speech data, ask for consent records linked to speaker IDs and listen to a random sample against its transcripts. Check for music, broadcasts or other third-party audio in the background, which needs separate clearance. Our [due diligence checklist](https://sellmydatatoai.net/guides/ai-training-data-due-diligence-checklist) covers the contract, and our guide to [licensing video and audio archives](https://sellmydatatoai.net/guides/sell-video-and-audio-for-ai-training) explains what sellers prepare. To source vetted speech and audio, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer). ## Frequently asked questions ### Is voice data biometric data? Voice data becomes biometric data when it is processed to identify or verify a person. Illinois' Biometric Information Privacy Act lists a voiceprint as a biometric identifier, and California's CCPA counts voice recordings from which a voiceprint can be extracted as biometric information when used to establish identity. UK GDPR defines biometric data as personal data from specific technical processing that allows unique identification. Plain recordings still count as personal information. ### Can podcast or broadcast audio be licensed for AI training? Podcast and broadcast audio can be licensed for AI training by whoever holds the rights to the recordings, and that often includes contributors and music owners as well as the publisher. Check contributor agreements and music licenses for any mention of AI training before relying on them. ### What transcripts should come with speech data? Transcripts that come with speech data should be time-aligned to the audio, labeled by speaker and checked by a person on at least a sample. Ask the seller for a measured accuracy figure and the method used. Machine transcripts without review are useful only as a starting point. ### Do speakers need to consent to AI training? Whether speakers need to consent to AI training depends on where they were recorded and the intended use. In the US, Illinois' BIPA requires a written release before a voiceprint is collected, and all-party consent laws such as California Penal Code section 632 govern the original recording. Under UK GDPR, biometric data processed to identify someone is special category data. Take independent legal advice. ## Sources - [Biometric Information Privacy Act, 740 ILCS 14/10 definitions](https://lig.ilga.gov/legislation/ilcs/documents/074000140K10.htm), Illinois General Assembly, accessed November 2026 - [Biometric Information Privacy Act, 740 ILCS 14](https://www.ilga.gov/legislation/ilcs/ilcs3.asp?ActID=3004&ChapterID=57), Illinois General Assembly, accessed November 2026 - [California Civil Code section 1798.140 (CCPA definitions)](https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?sectionNum=1798.140.&lawCode=CIV), California Legislative Information, accessed November 2026 - [California Penal Code section 632](https://law.justia.com/codes/california/code-pen/part-1/title-15/chapter-1-5/section-632), Justia, accessed November 2026 - [18 U.S. Code section 2511: interception and disclosure of wire, oral, or electronic communications prohibited](https://www.law.cornell.edu/uscode/text/18/2511), Legal Information Institute, Cornell Law School, accessed November 2026 - [UK GDPR Article 4: definitions](https://www.legislation.gov.uk/eur/2016/679/article/4), legislation.gov.uk, accessed November 2026 - [No Language Left Behind](https://ai.meta.com/research/no-language-left-behind/), Meta AI, accessed November 2026 - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Buy video training data for AI models URL: https://sellmydatatoai.net/buy-ai-training-data/video-training-data Video training data is licensed footage, with metadata and rights records, used to train video generation, computer vision and robotics models. Public prices are quoted per minute: Bloomberg reported in April 2024 that Adobe paid an average of about $2.62 per minute of submitted video, and Versos reports OpenAI paying $1 to $4 per minute. ## Key takeaways - AI teams use video training data for (1) text-to-video and video generation models, (2) computer vision and action recognition and (3) robotics and embodied AI. - Usually priced per minute of footage. - Common delivery formats: Original-resolution masters in MP4, MOV or ProRes, shot-level metadata and captions in JSON and rights and release records per asset. - Check before you license video training data: (1) US copyright in every layer: footage, music, performances and logos, (2) US: BIPA and similar state laws if face geometry is scanned from footage and (3) model and location releases for identifiable people. - Every seller we shortlist passes 8 vetting checks before you see a sample. ## Buyer's spec - Used for: Text-to-video and video generation models; Computer vision and action recognition; Robotics and embodied AI; Video understanding and captioning - Formats: Original-resolution masters in MP4, MOV or ProRes; Shot-level metadata and captions in JSON; Rights and release records per asset; Frame-level annotations where labeled - Usually priced: Per minute of footage - Check before you sign: US copyright in every layer: footage, music, performances and logos; US: BIPA and similar state laws if face geometry is scanned from footage; Model and location releases for identifiable people; Original resolution, not compressed web copies; Duplicates and near-duplicates removed; Personal and sensitive content handled; Withdrawal terms for models already trained ## What is video training data? Video training data is footage licensed for model training, together with the metadata and rights records that let a buyer use it safely. Versos, which runs a video archive licensing network, reports buyers seeking unpublished footage, film and TV libraries, corporate video and niche or large-scale libraries. Footage from robotics and first-person cameras is a separate category, used to teach models how people and objects move. Its rights questions center on the people and places in shot rather than on music or performances. ## How much does video training data cost? Video training data is priced per minute, and the public figures run from about $1 to about $7.25 per minute. | Buyer | Reported rate | Source | | --- | --- | --- | | Adobe | Average about $2.62 per minute, up to about $7.25 | Bloomberg, April 2024 | | OpenAI | $1 to $4 per minute | Versos blog | | Adobe | About $3 per minute | Versos blog | Bloomberg reported that Adobe asked contributors for footage of people in everyday actions such as walking or expressing emotions, simple anatomy shots of hands, feet or eyes, and people interacting with objects such as smartphones or fitness equipment. Briefs like these show that buyers often specify content in advance. ## What rights does video training data need? Video training data needs clearance for every layer inside the frame, not only the footage itself. Getphyllo's guidance on training data licenses calls this per-layer clearance for complex content, since video can contain music, images, performances and trademarks owned by others. - Copyright in the footage, confirmed by [chain of title](https://sellmydatatoai.net/glossary/chain-of-title) - Music and sound recordings in the soundtrack - Performers' and contributors' consent - Model releases for identifiable people - Trademarks, artworks and other copyright works in shot Versos makes the same point for sellers: confirm clear, transferable rights, obtain model releases and third-party clearances where needed, and use anonymization or consent protocols for personal or sensitive material. Under US copyright law, an exclusive license counts as a transfer of copyright ownership under 17 USC 101, and 17 USC 204(a) requires a transfer to be in writing and signed by the rights owner. Ask for the signed grant behind any exclusive rights the seller claims. Illinois' Biometric Information Privacy Act lists a scan of face geometry as a biometric identifier, which matters if faces in the footage will be analyzed. ## What makes video data more useful to AI teams? Video data is more useful when it comes as original masters with shot-level descriptions that let you filter by content. Versos notes that without a semantic, searchable structure even valuable assets can stay hidden. Ask for resolution, frame rate and codec in the [data manifest](https://sellmydatatoai.net/glossary/data-manifest), and check whether clips were cut from longer programs, since that affects both rights and duplication. ## Who sells video for AI training? Video for AI training is sold by broadcasters, film and TV libraries, stock footage agencies, creators and companies with corporate video archives. Versos lists Google, Meta, NVIDIA, Runway, Moonvalley, Synthesia and Getty Images among the places where creators can license video. ## How should video training data be delivered? Video training data should be delivered as original masters, with a manifest listing each clip's duration, resolution, frame rate, codec and rights record. Per-clip descriptions help you filter footage and make duplicate detection easier. Ask how the minutes were counted. A per-minute price on a catalog padded with near-duplicates, slates or black frames costs more per useful minute than the headline rate suggests. ## What should you check before buying video data? Before buying video data, sample at random across the catalog, check resolution and duplicates, and audit the rights records for a subset of clips. Agree withdrawal terms that say what happens to models already trained if a rights holder objects, which Getphyllo lists as one of six things a training data license should cover. Media owners can read about [selling media and video archives](https://sellmydatatoai.net/sell-your-data/media-and-video-archives), and our guide to [licensing video and audio archives](https://sellmydatatoai.net/guides/sell-video-and-audio-for-ai-training) explains the seller's side. See the [Versos profile](https://sellmydatatoai.net/data-buyers/versos) for one licensing network. To source vetted footage, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer). ## Frequently asked questions ### How much do AI companies pay per minute of video? AI companies pay roughly $1 to $7.25 per minute of video on public figures. Bloomberg reported in April 2024 that Adobe paid an average of about $2.62 per minute, rising to about $7.25, and Versos reports OpenAI paying $1 to $4 per minute. Ask what a quoted rate covers before comparing. ### What kinds of video do AI buyers want? AI buyers want video that fills specific gaps. Versos reports demand for unpublished footage, film and TV libraries and corporate video, and Bloomberg reported Adobe asking for people walking, expressing emotions, interacting with objects and close shots of hands, feet and eyes. Original masters are preferred. ### Does licensed footage need music clearance for AI training? Licensed footage does need separate music clearance for AI training if the soundtrack includes music the footage owner does not control. Getphyllo's guidance calls for per-layer clearance because video can contain music, images, performances and trademarks owned by others. Muted or stem-separated delivery is one option. ### Can I remove video from a model after training? Video cannot practically be removed from a model after training, a point Daniel Goldberg of Frankfurt Kurnit Klein and Selz makes for all training data. That is why withdrawal terms matter: they set out what happens to raw copies, future training runs and existing models if a rights holder objects. ## Sources - [Adobe is buying videos for $3 per minute to build AI model](https://www.bnnbloomberg.ca/adobe-is-buying-videos-for-3-per-minute-to-build-ai-model-1.2057606.amp.html), Bloomberg via BNN Bloomberg, accessed November 2026 - [Top 10 places to sell your video content for AI training data](https://www.versos.ai/blog/sell-video-content-ai-training-data), Versos, accessed November 2026 - [Best public data providers for AI training and grounding](https://www.getphyllo.com/post/best-public-data-providers-ai-training-grounding), Getphyllo, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 - [17 U.S. Code section 101: definitions](https://www.law.cornell.edu/uscode/text/17/101), Legal Information Institute, Cornell Law School, accessed November 2026 - [17 U.S. Code section 204: execution of transfers of copyright ownership](https://www.law.cornell.edu/uscode/text/17/204), Legal Information Institute, Cornell Law School, accessed November 2026 - [Biometric Information Privacy Act, 740 ILCS 14](https://www.ilga.gov/legislation/ilcs/ilcs3.asp?ActID=3004&ChapterID=57), Illinois General Assembly, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Buy licensed image datasets for AI training URL: https://sellmydatatoai.net/buy-ai-training-data/image-datasets Image datasets are licensed collections of photographs, illustrations or scans with captions and rights records, used to train image generation and computer vision models. Licensed archives are now a business: Shutterstock's chief executive told Bloomberg that its AI licensing operation produced $104 million in revenue in 2023. Scraped sets carry real risk, as LAION-5B's withdrawal in December 2023 showed. ## Key takeaways - AI teams use image datasets for (1) text-to-image generation, (2) image classification and object detection and (3) vision-language models and captioning. - Usually priced per image, or per collection for large archives. - Common delivery formats: Original-resolution JPEG, PNG or TIFF, captions and tags in JSON or Parquet and bounding boxes or masks in COCO format. - Check before you license image datasets: (1) US copyright: owner of every image, with signed transfers where rights were acquired, (2) US: BIPA if face geometry will be scanned, since photographs alone are excluded and (3) model and property releases. - Every seller we shortlist passes 8 vetting checks before you see a sample. ## Buyer's spec - Used for: Text-to-image generation; Image classification and object detection; Vision-language models and captioning; Document and specialist imaging models - Formats: Original-resolution JPEG, PNG or TIFF; Captions and tags in JSON or Parquet; Bounding boxes or masks in COCO format; Per-image rights and release records - Usually priced: Per image, or per collection for large archives - Check before you sign: US copyright: owner of every image, with signed transfers where rights were acquired; US: BIPA if face geometry will be scanned, since photographs alone are excluded; Model and property releases; Illegal and harmful content screened; Machine-generated captions disclosed; Duplicates and near-duplicates removed; Location metadata stripped where it identifies people ## What is an image dataset for AI training? An image dataset is a collection of images paired with text or labels, plus the records that prove the seller can license each one. Captions drive text-to-image models, and labels such as bounding boxes drive detection models. Specialist images, such as medical scans, satellite imagery or industrial inspection photos, are sold with expert labels and carry the rules of their own sector, including health data law for medical images. ## Who sells licensed image datasets? Licensed image datasets come from stock libraries, photo agencies, publishers and companies with large in-house image archives. Shutterstock is the best documented: its chief executive Paul Hennessy told Bloomberg in June 2024 that its AI licensing business produced $104 million in revenue the previous year. See our [Shutterstock data licensing profile](https://sellmydatatoai.net/data-buyers/shutterstock-data-licensing). Company archives count too. Mercor's published listing includes site photos, alongside project records and drawings, in construction company deals worth $115,000 to $1 million for firms of 50 to 2,000 employees, as compiled by ddiy.co in October 2026. ## Why does provenance matter so much for images? Provenance matters for images because scraped collections have contained illegal material that no buyer can lawfully hold. In December 2023 the Stanford Internet Observatory reported more than 3,200 images of suspected child sexual abuse in LAION-5B, and LAION took its datasets down to make them safe before republishing. A licensed archive with a known owner, a documented upload history and moderation records reduces that risk sharply. Our guide to [data provenance](https://sellmydatatoai.net/guides/data-provenance-explained) lists the records to ask for. ## What does an image rights record contain? An image rights record shows who made each image and why the seller can license it. Ask for these fields for every image, and check them in full for a random subset. - Creator and date of creation - License or assignment giving the seller the right to license it - Model release for identifiable people - Property release for private property or artworks where needed - Moderation history showing the image was reviewed ## How are image datasets priced? Image datasets are priced per image for smaller sets and per collection for large archives. Published per-image training rates are rare, so most prices are negotiated from the type of license. | License type | Reported range | | --- | --- | | Research | $5,000 to $15,000 | | Commercial | $25,000 to $500,000 and above | | Professional, annual | $25,000 to $50,000 a year | | Enterprise, annual | $50,000 to $500,000 a year | | Exclusive or custom | $1 million and above | Getphyllo notes that a research license rarely permits production training, so check the license type before comparing prices. ## How should an image dataset be delivered? An image dataset should be delivered as original-resolution files with a manifest listing each image's ID, dimensions, source, license record and caption. Labels for detection or segmentation should use a standard format such as COCO, with the labeling guidelines attached. Ask for a duplicate report or perceptual hash values, since a stock archive can hold the same photograph at several crops and sizes. Our guide on [how to evaluate a dataset sample](https://sellmydatatoai.net/guides/how-to-evaluate-dataset-samples) covers the tests to run. ## What should you check before buying an image dataset? Before buying an image dataset, audit rights for a random subset, screen the full set with hash-matching tools for known illegal content and check captions against the images. Strip location data from EXIF metadata where it could identify people's homes, and confirm model releases for identifiable faces. Illinois' Biometric Information Privacy Act lists a scan of face geometry as a biometric identifier but excludes photographs themselves, so the risk rises when face templates are extracted. California's CCPA counts face imagery from which a faceprint can be extracted as biometric information when it is used or intended to be used to establish identity. If you hold UK or EU images, UK GDPR Article 4 gives facial images as an example of biometric data when they are processed to identify a person. Our [due diligence checklist](https://sellmydatatoai.net/guides/ai-training-data-due-diligence-checklist) covers the contract. To source vetted image archives, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer). ## Frequently asked questions ### Is it safe to train on scraped image datasets? Training on scraped image datasets carries legal and content risk, because no one has checked the rights or the content of every image. The Stanford Internet Observatory found more than 3,200 images of suspected child sexual abuse in LAION-5B in December 2023, after which LAION took its datasets down. ### How much do licensed image datasets cost? Licensed image datasets cost anything from a few thousand dollars for a research license to $1 million or more for exclusive deals. Getphyllo's reported market ranges, which are not quotes, put commercial licenses at $25,000 to $500,000 and above. Research licenses rarely permit production training. ### Do images of people need model releases for AI training? Images of identifiable people generally need model releases or another clear right of use before AI training. In the US, Illinois' BIPA covers scans of face geometry though not photographs themselves, and California's CCPA treats face imagery used to establish identity as biometric information. Under UK GDPR, facial images processed to identify a person are biometric data. Ask the seller for release records and check a random sample. ### Are captions included with licensed image datasets? Captions for licensed image datasets come from people, from original metadata or from a model, and their quality varies. Ask the seller which source each caption came from, since machine captions can repeat the errors of the model that wrote them. ## Sources - [Shutterstock's AI-licensing business generated $104 million](https://news.bloomberglaw.com/ip-law/shutterstocks-ai-licensing-business-generated-104-million), Bloomberg Law, accessed November 2026 - [AI image dataset is pulled after child sex abuse pictures discovered](https://petapixel.com/2023/12/20/ai-image-dataset-is-pulled-after-child-sex-abuse-pictures-discovered), PetaPixel, accessed November 2026 - [Best public data providers for AI training and grounding](https://www.getphyllo.com/post/best-public-data-providers-ai-training-grounding), Getphyllo, accessed November 2026 - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 - [Biometric Information Privacy Act, 740 ILCS 14](https://www.ilga.gov/legislation/ilcs/ilcs3.asp?ActID=3004&ChapterID=57), Illinois General Assembly, accessed November 2026 - [California Civil Code section 1798.140 (CCPA definitions)](https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?sectionNum=1798.140.&lawCode=CIV), California Legislative Information, accessed November 2026 - [UK GDPR Article 4: definitions](https://www.legislation.gov.uk/eur/2016/679/article/4), legislation.gov.uk, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Buy multilingual text data for AI training URL: https://sellmydatatoai.net/buy-ai-training-data/multilingual-text-data Multilingual text data is licensed writing in languages other than English, or parallel text across languages, used to train and evaluate models outside English. The gap is large: English makes up about 42.6% of pages in Common Crawl's CC-MAIN-2026-43 crawl, and Meta's No Language Left Behind project states that reliable translation data is not available for many languages. ## Key takeaways - AI teams use multilingual text data for (1) pre-training and fine-tuning in under-served languages, (2) machine translation from parallel text and (3) multilingual evaluation sets. - Usually priced per million words or tokens, or per archive. - Common delivery formats: Plain text or JSONL with language codes, aligned parallel corpora in TMX or TSV and translation memories exported from CAT tools. - Check before you license multilingual text data: (1) written by native speakers, not machine translated, (2) language and dialect labels verified and (3) US copyright: a translation is a derivative work, so rights in the source text still apply. - Every seller we shortlist passes 8 vetting checks before you see a sample. ## Buyer's spec - Used for: Pre-training and fine-tuning in under-served languages; Machine translation from parallel text; Multilingual evaluation sets; Localization and cultural alignment - Formats: Plain text or JSONL with language codes; Aligned parallel corpora in TMX or TSV; Translation memories exported from CAT tools; Metadata on dialect, region and register - Usually priced: Per million words or tokens, or per archive - Check before you sign: Written by native speakers, not machine translated; Language and dialect labels verified; US copyright: a translation is a derivative work, so rights in the source text still apply; Translator and client rights in translation memories; US: personal information in correspondence removed under the CCPA and other state privacy laws; UK and EU: UK GDPR or EU GDPR for personal data in correspondence; Domain and register spread documented ## What is multilingual text data? Multilingual text data is human-written text in languages other than English, or text aligned across two or more languages so each passage can be matched to its translation. Sources include publishers, translation agencies, broadcasters, public bodies and companies that operate in several languages. ## Why do AI teams buy multilingual text? AI teams buy multilingual text because public web data is dominated by a few languages. In Common Crawl's CC-MAIN-2026-43 crawl, English accounts for about 42.6% of pages, followed by Russian at 6.3%, German at 5.8%, Japanese at 4.9% and French at 4.8%. | Language | Share of pages | | --- | --- | | English | 42.6% | | Russian | 6.3% | | German | 5.8% | | Japanese | 4.9% | | French | 4.8% | Many languages have far less. Meta's No Language Left Behind project, which translates between 200 languages including 150 low-resource ones, states that reliable translation data is not available for many languages. ## What makes multilingual data useful? Multilingual data is most useful when it is written by native speakers, in a language or dialect that public data covers poorly, across domains such as law, health and customer service. Ask how much was machine translated, since a model trained on machine translation learns that system's output rather than native usage. Parallel text, where each sentence is matched to its translation, is the raw material for translation models. Translation memories held by agencies can supply it, but each one sits under the rights of the client who owned the source and the translator who wrote the target. US copyright law spells this out. 17 USC 101 lists a translation as a derivative work, and under 17 USC 103(b) copyright in a derivative work extends only to the material its author contributed, leaving the source owner's rights unchanged. ## How is multilingual text priced? Multilingual text is priced per million words or tokens for catalog sales, or as a flat fee for a company archive. No buyer publishes a per-language rate card. Getphyllo's reported market ranges put commercial licenses at $25,000 to $500,000 or more and research licenses at $5,000 to $15,000, and research licenses rarely permit production training. Our guide to [AI training data pricing models](https://sellmydatatoai.net/guides/ai-training-data-pricing-models) explains per-unit and flat-fee structures, and the glossary defines a [data license](https://sellmydatatoai.net/glossary/data-license). ## Who sells multilingual text? Multilingual text is sold by publishers, news agencies, translation and localization companies, and businesses that serve customers in several languages, such as call centers and ecommerce firms. Support and sales records in a local language show everyday usage that published text rarely contains. ## How should multilingual text be delivered? Multilingual text should be delivered as UTF-8 JSONL or plain text, with an ISO 639 language code on every record and dialect or region labels where they matter. Parallel corpora should keep alignment explicit, in TMX or tab-separated files, with document IDs so each segment can be traced to its source. Ask for word or token counts per language in the [data manifest](https://sellmydatatoai.net/glossary/data-manifest), and for the share of each language that came from each source. A language drawn mostly from one source reflects that source's style and subject matter. ## What should you check before buying multilingual data? Before buying multilingual data, have native speakers review a random sample for each language and run a language identification tool across the set to confirm the labels. Check for boilerplate too. Menus, legal notices and repeated templates add little to training in any language, and they can inflate word counts in a corpus built from websites. Confirm who owns translations. Translation memories can contain client documents, and an agency may hold only a license to use them for that client's work. Personal information in correspondence must be removed or de-identified to the standard of the US state privacy laws that apply, such as the CCPA. If you hold UK or EU correspondence, UK GDPR or EU GDPR also applies unless the data is anonymized. Publishers can read about [selling publisher and education data](https://sellmydatatoai.net/sell-your-data/publishers-and-education), and buyers can learn how to [evaluate a dataset sample](https://sellmydatatoai.net/guides/how-to-evaluate-dataset-samples). To source vetted multilingual corpora, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer). ## Frequently asked questions ### Why is multilingual training data hard to find? Multilingual training data is hard to find because the open web is concentrated in a few languages. English makes up about 42.6% of pages in Common Crawl's CC-MAIN-2026-43 crawl, and Meta's No Language Left Behind project states that reliable translation data is not available for many languages. ### Can translation memories be licensed for AI training? Translation memories can be licensed for AI training if the agency or company holds the rights to both source and target text. Source documents usually belong to clients, and the agency's contract may limit reuse. Check each client agreement and remove personal data before any sample is shared. ### Is machine-translated text useful for training? Machine-translated text is of limited use for training because a model learns the translation system's patterns rather than native usage. Buyers looking for language quality should ask sellers to disclose how much of a corpus was machine translated and to label it separately. ### How is multilingual text measured for pricing? Multilingual text is measured for pricing in words or tokens per language, with documents counted where structure matters. Token counts depend on the tokenizer used, so agree the counting method in the license and check it against the data manifest before the final payment. ## Sources - [Statistics of Common Crawl monthly archives: distribution of languages](https://commoncrawl.github.io/cc-crawl-statistics/plots/languages), Common Crawl, accessed November 2026 - [No Language Left Behind](https://ai.meta.com/research/no-language-left-behind/), Meta AI, accessed November 2026 - [Best public data providers for AI training and grounding](https://www.getphyllo.com/post/best-public-data-providers-ai-training-grounding), Getphyllo, accessed November 2026 - [17 U.S. Code section 101: definitions](https://www.law.cornell.edu/uscode/text/17/101), Legal Information Institute, Cornell Law School, accessed November 2026 - [17 U.S. Code section 103: subject matter of copyright, compilations and derivative works](https://www.law.cornell.edu/uscode/text/17/103), Legal Information Institute, Cornell Law School, accessed November 2026 - [California Civil Code section 1798.140 (CCPA definitions)](https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?sectionNum=1798.140.&lawCode=CIV), California Legislative Information, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Buy document and form data for AI training URL: https://sellmydatatoai.net/buy-ai-training-data/document-and-form-data Document and form data is licensed collections of real business documents, such as invoices, contracts, tax forms, reports and scanned records, used to train document understanding, extraction and drafting models. Mercor's published listings value tax returns, workpapers and bookkeeping records at $25,000 to $865,000 per accounting firm, and closed legal files at $50,000 to $1.2 million per law firm. ## Key takeaways - AI teams use document and form data for (1) document understanding and OCR models, (2) field extraction from forms and invoices and (3) contract review and drafting models. - Usually priced per document or page for labeled sets, per archive for company records. - Common delivery formats: Original PDFs and scans in TIFF or PNG, native DOCX and XLSX files and extracted text and layout in JSON. - Check before you license document and form data: (1) US: IRC Section 7216 limits on tax return information, (2) US: HIPAA de-identification for health records and (3) US: GLBA notice and opt-out rules for financial institutions' customer records. - Every seller we shortlist passes 8 vetting checks before you see a sample. ## Buyer's spec - Used for: Document understanding and OCR models; Field extraction from forms and invoices; Contract review and drafting models; Agent training on document-heavy workflows - Formats: Original PDFs and scans in TIFF or PNG; Native DOCX and XLSX files; Extracted text and layout in JSON; Field-level labels and key-value pairs - Usually priced: Per document or page for labeled sets, per archive for company records - Check before you sign: US: IRC Section 7216 limits on tax return information; US: HIPAA de-identification for health records; US: GLBA notice and opt-out rules for financial institutions' customer records; Personal and financial data removed from text and images; Client and counterparty confidentiality; Scans legible and complete; Labels checked against source documents; Document counts and types match the manifest ## What is document and form data? Document and form data is the paperwork of real organizations in its original layout: invoices, purchase orders, contracts, tax forms, insurance claims, inspection reports and scanned records. Layout matters as much as text, because models must learn where information sits on a page. Sources include accounting and law firms, insurers, logistics companies handling shipping documents, manufacturers with quality records and organizations with large scanned archives. ## What do AI teams use document data for? AI teams use document data to train models that read, extract from and produce business documents. The same archive can feed several uses. | Use | What the data needs | | --- | --- | | OCR and layout | Scans and PDFs at original quality, across varied templates | | Field extraction | Labeled key-value pairs checked against the source | | Contract review | Executed agreements with clause-level annotations | | Drafting | Final documents paired with drafts and instructions | | Workflow agents | Documents linked to the decisions they triggered | ## How much is document data worth? Document data has no list price, but published payouts for document-heavy company archives give a guide. Mercor's listings, compiled by ddiy.co in October 2026, include these. - Accounting and tax firms: $25,000 to $865,000 for tax returns, workpapers and bookkeeping records - Law firms: $50,000 to $1.2 million for complete files from closed matters and completed deals - Consulting firms: $150,000 to $300,000 for client deliverables, models and the drafts behind them - Manufacturers: $100,000 to $1.5 million for production, maintenance and inspection records Mercor's listings are payouts to the firms, not prices to buyers. Sellers in these sectors can read our pages for [accounting and tax firms](https://sellmydatatoai.net/sell-your-data/accounting-firms), [law firms](https://sellmydatatoai.net/sell-your-data/law-firms) and [insurance companies](https://sellmydatatoai.net/sell-your-data/insurance). ## What legal rules apply to document data? Document data is governed by the rules for whatever the documents contain, and those rules differ by record type. - Tax records: in the US, IRC Section 7216 prohibits tax return preparers from knowingly or recklessly disclosing or using tax return information - Health records: HIPAA permits de-identification by expert determination or by removing 18 identifiers under 45 CFR 164.514 - Financial records: under GLBA, at 15 USC 6802, a financial institution may not disclose nonpublic personal information to a nonaffiliated third party unless the consumer has had notice and a chance to opt out - Payment records: PCI DSS states that sensitive authentication data, including card verification codes, must not be stored after authorization - Personal data in the UK and EU: UK GDPR and EU GDPR apply unless the data is anonymized We are a broker, not a law firm, so take independent advice on the rules that apply to the records you plan to license. ## How should document data be delivered? Document data should be delivered in original form, as PDFs or scans at the resolution they were captured, with native files where they exist. Extracted text and layout should come alongside in JSON, keyed to page and position, so you can check the OCR against the image. Labels for field extraction should name the field, its value and its location on the page, with the labeling guide attached. Ask for counts by document type and template, because a set dominated by one template teaches a model little about variety. Handwritten forms need their own check. Ask whether handwriting was transcribed by people or by a model, and compare a sample of transcriptions with the scans. ## What should you check before buying document data? Before buying document data, check redaction on the images as well as the extracted text. A name removed from the OCR output but still visible in the scan has still been disclosed. Check legibility on a random sample, confirm document types and counts against the [data manifest](https://sellmydatatoai.net/glossary/data-manifest) and verify labels against source documents. Our guide to [de-identification for AI training data](https://sellmydatatoai.net/guides/de-identification-for-ai-training-data) covers redaction methods. [Book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer) to source vetted document archives. ## Frequently asked questions ### What documents are most useful for document AI training? The documents most useful for document AI training are real business records in their original layout, across many templates: invoices, contracts, tax forms, claims and inspection reports. Variety of layout matters as much as volume, and labeled fields checked against the source add value for extraction models. ### How are scanned documents redacted for AI training? Scanned documents are redacted for AI training by removing identifiers from the image itself as well as from the extracted text. Burned-in redaction on the image is needed, since masking only the OCR layer leaves names visible. Check a random sample of scans after redaction. ### Can accounting firms license client documents? Accounting firms can license client documents only within professional and legal limits. In the US, IRC Section 7216 restricts preparers from disclosing or using tax return information without consent. Mercor publishes payouts of $25,000 to $865,000 for tax returns, workpapers and bookkeeping records from firms of 5 to 200 employees. ### Should document data include drafts as well as final versions? Document data that includes drafts as well as final versions is more useful for drafting models, because it shows how a document was improved. Mercor's consulting listing, for example, covers client deliverables, models and the drafts behind them, with payouts of $150,000 to $300,000. ## Sources - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 - [Section 7216 frequently asked questions](https://www.irs.gov/node/3756), Internal Revenue Service, accessed November 2026 - [Guidance regarding methods for de-identification of protected health information](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html), US Department of Health and Human Services, accessed November 2026 - [15 U.S. Code section 6802: obligations with respect to disclosures of personal information](https://www.law.cornell.edu/uscode/text/15/6802), Legal Information Institute, Cornell Law School, accessed November 2026 - [Why is storage of sensitive authentication data after authorization not permitted?](https://www.pcisecuritystandards.org/faq/articles/Frequently_Asked_Question/For-PCI-DSS-why-is-storage-of-sensitive-authentication-data-SAD-after-authorization-not-permitted-even-when-there-are-no-primary-account-numbers-PANs-in-an-environment), PCI Security Standards Council, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Buy call recording data for AI training URL: https://sellmydatatoai.net/buy-ai-training-data/call-recording-data Call recording data is recorded phone or video conversations, mostly from contact centers, licensed with transcripts and outcome data to train speech, support and sales models. Mercor's published listing puts call centers at $20,000 to $250,000 for recorded customer conversations from operations with 50 to 5,000 agents. Consent to record and to reuse are the main checks. ## Key takeaways - AI teams use call recording data for (1) speech recognition on real telephone audio, (2) contact center assistants and agent coaching and (3) sentiment and intent classification. - Usually priced per hour of audio, or per archive for a contact center. - Common delivery formats: WAV or FLAC with agent and caller on separate channels, time-aligned transcripts with speaker labels and call metadata such as queue, duration and outcome. - Check before you license call recording data: (1) US: lawful recording where each party was located, including all-party consent states such as California, (2) US: privacy notice covering reuse beyond quality monitoring and (3) US: BIPA and other biometric laws for voice cloning or speaker identification. - Every seller we shortlist passes 8 vetting checks before you see a sample. ## Buyer's spec - Used for: Speech recognition on real telephone audio; Contact center assistants and agent coaching; Sentiment and intent classification; Call summarization and quality scoring models - Formats: WAV or FLAC with agent and caller on separate channels; Time-aligned transcripts with speaker labels; Call metadata such as queue, duration and outcome; Quality scores and dispositions - Usually priced: Per hour of audio, or per archive for a contact center - Check before you sign: US: lawful recording where each party was located, including all-party consent states such as California; US: privacy notice covering reuse beyond quality monitoring; US: BIPA and other biometric laws for voice cloning or speaker identification; Card details removed from audio and transcripts; Outsourcing client contracts permit reuse; Personal data spoken aloud redacted; UK and EU: callers' voices are personal data under UK GDPR or EU GDPR ## What is call recording data? Call recording data is audio of real conversations between customers and agents, with transcripts and the metadata that says what happened on each call. It comes from contact centers and business process outsourcers, and from sales and support teams that record their calls. ## What do AI teams use call recordings for? AI teams use call recordings for speech recognition on telephone audio and for models that support or score contact center work. - Speech recognition trained on narrowband, noisy and accented telephone audio - Real-time assistants that suggest answers to agents during calls - Intent and sentiment classification for routing and analytics - Summarization and quality scoring after each call The broader [speech and audio data](https://sellmydatatoai.net/buy-ai-training-data/speech-and-audio-data) page covers non-call sources. Telephone audio differs from studio speech in bandwidth, background noise and overlap between speakers, so speech recognition teams that serve contact centers need it as well as cleaner recordings. ## How much is call recording data worth? Call recording data is worth $20,000 to $250,000 per contact center on Mercor's published listing, for recorded customer conversations from operations with 50 to 5,000 agents, as compiled by ddiy.co in October 2026. That listing measures size in agents rather than employees, so the number of agents and the years of retained audio are the first questions to ask a seller. Recordings that carry outcome data, such as whether the issue was resolved or a sale closed, give a model a target to learn from. ## What consent rules apply to call recordings? Consent rules for call recordings depend on where each party was when the call took place. Federal law at 18 USC 2511(2)(d) permits a recording where one party to the call consents, but California Penal Code section 632 makes it an offense to record a confidential communication without the consent of all parties, with fines of up to $2,500 for a first violation. Lawful recording for quality purposes does not by itself permit licensing for AI training. Frankfurt Kurnit Klein and Selz notes that companies must comply with the privacy statements in effect when data was collected, and lists call recording consent laws and biometric laws among the rules that can apply. FTC staff warned in February 2024 that retroactive changes to privacy policies to allow AI training may be unfair or deceptive. Voice adds a biometric layer. Illinois' Biometric Information Privacy Act lists a voiceprint as a biometric identifier, and section 15(c) bars an entity holding one from selling or otherwise profiting from it, so speaker identification and voice cloning uses need separate attention. If you hold UK or EU recordings, callers' voices are personal data under UK GDPR or EU GDPR, and voice processed to identify a caller is special category biometric data. > **Card details** PCI DSS states that sensitive authentication data, including card verification codes, must not be stored after authorization even if encrypted. Any call where a customer reads out card details must be cut or redacted before licensing. ## How should call recordings be delivered? Call recordings should be delivered as lossless audio with agent and caller on separate channels where the system recorded them that way, plus time-aligned transcripts, speaker labels and call metadata. Separate channels make diarization and redaction easier. Redaction must cover the audio as well as the transcript. Personal data spoken aloud, such as names, addresses and account numbers, should be muted or replaced with a tone in the audio and masked in the text, with the method documented. Ask for the date range, total hours, number of agents and languages in the [data manifest](https://sellmydatatoai.net/glossary/data-manifest), and check them against a random sample of calls. ## What should you check before buying call recordings? Before buying call recordings, ask for the recording notice callers heard, the privacy notice in force at the time and, if the seller is an outsourcer, its client contracts. An outsourcer records on behalf of its clients, and the client contract decides who can license the recordings. Contact centers on the supply side can read about [selling call center data](https://sellmydatatoai.net/sell-your-data/call-centers). Our [due diligence checklist](https://sellmydatatoai.net/guides/ai-training-data-due-diligence-checklist) covers the contract. To source vetted call archives, [book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer). ## Frequently asked questions ### Can a call center sell its call recordings for AI training? A call center can license its call recordings for AI training if the calls were lawfully recorded, its privacy notices allow the reuse and its client contracts do not prohibit it. Outsourcers usually need their clients' agreement. Mercor publishes payouts of $20,000 to $250,000 for recorded customer conversations. ### Does a quality monitoring notice cover AI training? A quality monitoring notice does not automatically cover AI training. Frankfurt Kurnit Klein and Selz notes that companies must comply with the privacy statements in effect when data was collected, and FTC staff warned in February 2024 that a surreptitious, retroactive change to a privacy policy to allow AI training may be unfair or deceptive. ### How are payment details handled in call recordings? Payment details in call recordings must be removed before licensing. PCI DSS states that sensitive authentication data, including card verification codes, must not be stored after authorization even if encrypted. Calls where card details were read out should be cut or redacted in both audio and transcript. ### What metadata should come with call recordings? Metadata that should come with call recordings includes queue, duration, language, outcome or disposition, quality scores and speaker channel labels. Outcome fields matter most, because they tell a model whether a call succeeded. Ask for separate agent and caller channels where the recording system supports them. ## Sources - [How to sell company data to AI companies (2026 guide)](https://ddiy.co/sell-company-data-to-ai/), ddiy.co, accessed November 2026 - [California Penal Code section 632](https://law.justia.com/codes/california/code-pen/part-1/title-15/chapter-1-5/section-632), Justia, accessed November 2026 - [18 U.S. Code section 2511: interception and disclosure of wire, oral, or electronic communications prohibited](https://www.law.cornell.edu/uscode/text/18/2511), Legal Information Institute, Cornell Law School, accessed November 2026 - [Selling company data for AI training: what to know before you say yes](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes), Frankfurt Kurnit Klein and Selz, accessed November 2026 - [AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive](https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/02/ai-other-companies-quietly-changing-your-terms-service-could-be-unfair-or-deceptive), Federal Trade Commission, accessed November 2026 - [Biometric Information Privacy Act, 740 ILCS 14/10 definitions](https://lig.ilga.gov/legislation/ilcs/documents/074000140K10.htm), Illinois General Assembly, accessed November 2026 - [Biometric Information Privacy Act, 740 ILCS 14](https://www.ilga.gov/legislation/ilcs/ilcs3.asp?ActID=3004&ChapterID=57), Illinois General Assembly, accessed November 2026 - [Why is storage of sensitive authentication data after authorization not permitted?](https://www.pcisecuritystandards.org/faq/articles/Frequently_Asked_Question/For-PCI-DSS-why-is-storage-of-sensitive-authentication-data-SAD-after-authorization-not-permitted-even-when-there-are-no-primary-account-numbers-PANs-in-an-environment), PCI Security Standards Council, accessed November 2026 - [UK GDPR Article 4: definitions](https://www.legislation.gov.uk/eur/2016/679/article/4), legislation.gov.uk, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation --- # Buy CRM and sales data for AI training URL: https://sellmydatatoai.net/buy-ai-training-data/crm-and-sales-data CRM and sales data is the history of deals, accounts, activities and pipeline held in a customer relationship management system, licensed to train sales assistants and AI agents. micro1 lists CRM records among the operational data it licenses, with company payouts from $100,000. The value is in activity histories with outcomes, and contact personal data is the main legal risk. ## Key takeaways - AI teams use crm and sales data for (1) sales assistant and email drafting models, (2) agents that update records and plan next steps and (3) forecasting and pipeline analysis models. - Usually priced per company archive, as a flat license fee. - Common delivery formats: CSV or Parquet exports of objects and fields, JSON activity timelines with linked emails and calls and schema documentation for custom fields. - Check before you license crm and sales data: (1) US: CCPA sale and opt-out rules for California residents, (2) US: privacy notice in force when contacts were collected and (3) contact personal data removed or pseudonymized. - Every seller we shortlist passes 8 vetting checks before you see a sample. ## Buyer's spec - Used for: Sales assistant and email drafting models; Agents that update records and plan next steps; Forecasting and pipeline analysis models; Evaluating agent decisions against real deal outcomes - Formats: CSV or Parquet exports of objects and fields; JSON activity timelines with linked emails and calls; Schema documentation for custom fields; Pseudonymized account and contact IDs - Usually priced: Per company archive, as a flat license fee - Check before you sign: US: CCPA sale and opt-out rules for California residents; US: privacy notice in force when contacts were collected; Contact personal data removed or pseudonymized; Customer contracts restricting use of their data; Won and lost outcomes present; Schema and custom fields documented; UK and EU: lawful basis under UK GDPR or EU GDPR ## What is CRM and sales data? CRM and sales data is the record a sales team keeps in its customer relationship management system: accounts, contacts, opportunities, stages, activities, notes, quotes and outcomes. Linked emails and call notes turn it from a table into a history of how each deal was won or lost. micro1's data partnerships page names CRM records among the operational data it licenses, alongside SOPs, project histories, support tickets and internal communications, from companies with 30 or more full-time employees. See our [micro1 profile](https://sellmydatatoai.net/data-buyers/micro1) for its published terms. ## What do AI teams use CRM data for? AI teams use CRM data to train sales assistants and agents that work inside CRM systems, and to test whether an agent's recommended next step matches what actually won deals. - Drafting follow-up emails and call notes in context - Updating records and planning next steps as an agent - Forecasting deal outcomes from stage and activity history - Evaluation sets built from closed deals with known results ## How much is CRM data worth? CRM data has no separate public price. micro1's published payouts, which cover CRM records alongside other operational data, are set in three tiers. | Tier | Published payout | micro1's description | | --- | --- | --- | | Qualified enterprise data partnership | $100,000 to $500,000 | Entry tier for companies that meet the criteria | | Large-scale or ongoing | $500,000 to $3 million | Large-scale operational datasets or ongoing participation across teams | | Highly unique | $3 million and above | Proprietary operational data with significant value for frontier AI development | Those are payouts to data owners. Fortune reported in September 2026, citing SimpleClosure, that payments for data from closed startups typically run from $10,000 to $100,000 per company, which is a better guide for small archives. Ask the seller how many years of records the archive covers and how many deals closed in that time, since the number of complete deal histories is what you are buying. ## What are the legal risks in CRM data? The legal risks in CRM data come from the personal data of contacts and the confidential terms of customers. Every contact record holds names, email addresses and phone numbers, and every opportunity can hold a customer's pricing. - US sale rules: under the CCPA, at California Civil Code section 1798.140(ad), making personal information available to a third party for valuable consideration is a sale, unless the data is deidentified to the standard in section 1798.140(m), and California residents can opt out of sale - US privacy notices: FTC staff warned in February 2024 that retroactive changes to privacy policies to allow AI training may be unfair or deceptive - Customer contracts: NDAs and confidentiality clauses can cover pricing and deal terms - UK and EU: UK GDPR and EU GDPR require a lawful basis, and [pseudonymization](https://sellmydatatoai.net/glossary/pseudonymization) leaves data personal Our guide to [UK GDPR, EU GDPR and CCPA when selling data for AI training](https://sellmydatatoai.net/guides/gdpr-and-selling-data-for-ai-training) explains each regime. ## How should CRM data be delivered? CRM data should be delivered as a fixed export of each object, such as accounts, contacts, opportunities and activities, in CSV or Parquet, with a schema document explaining every custom field and picklist value. Keep the relational keys so you can rebuild each deal's history. Activity timelines are where the value sits, so ask whether linked emails and call notes are included and how they were redacted. Exports from Salesforce, HubSpot and similar systems differ, so agree the format and the [data manifest](https://sellmydatatoai.net/glossary/data-manifest) before extraction starts. ## What should you check before buying CRM data? Before buying CRM data, confirm that contacts were pseudonymized consistently, so the same person keeps one ID across records, and that free-text notes were redacted as well as structured fields. Check that won and lost outcomes are present and that the schema is documented, since custom fields are hard to interpret without it. Our [due diligence checklist](https://sellmydatatoai.net/guides/ai-training-data-due-diligence-checklist) covers the contract. [Book a free consultation](https://sellmydatatoai.net/book-a-consultation?side=buyer) to source vetted CRM archives. ## Frequently asked questions ### Do AI companies buy CRM data? AI companies do buy CRM data as part of wider operational archives. micro1 names CRM records among the data it licenses from companies with 30 or more full-time employees, with published payouts starting at $100,000 to $500,000 for a qualified enterprise data partnership. ### Is a contact list worth selling to AI companies? A contact list on its own is rarely worth selling to AI companies, because names and email addresses teach a model nothing about how work is done, and they are personal data that carries legal risk. Activity histories with outcomes are what AI buyers want from CRM systems. ### Does selling CRM data count as a sale under the CCPA? Selling CRM data can count as a sale under the CCPA if it includes personal information made available to a third party for monetary or other valuable consideration, the definition in California Civil Code section 1798.140(ad). Data deidentified to the standard in section 1798.140(m) is not personal information. That standard requires reasonable measures, a public commitment not to re-identify and contractual obligations on recipients. ### How should CRM contacts be de-identified? CRM contacts should be de-identified by replacing names, email addresses and phone numbers with consistent IDs and removing identifiers from free-text notes. That is pseudonymization, which the CCPA does not treat as deidentified unless the conditions in section 1798.140(m) are also met. Under UK GDPR, pseudonymized data is still personal data. Full de-identification needs a re-identification test across linked fields. ## Sources - [Monetize your company's operational data](https://www.micro1.ai/data), micro1, accessed November 2026 - [Little-known AI startup Micro1 tries to trump Google's bid for bankrupt Spirit Airlines' data](https://fortune.com/2026/09/14/micro1-tries-to-beat-googles-bid-to-buy-bankrupt-spirit-airlines-data-for-12-5-million-highlighting-value-of-data-to-ai-companies/), Fortune, accessed November 2026 - [California Civil Code section 1798.140 (CCPA definitions)](https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?sectionNum=1798.140.&lawCode=CIV), California Legislative Information, accessed November 2026 - [California Consumer Privacy Act (CCPA)](https://oag.ca.gov/privacy/ccpa), State of California Department of Justice, Office of the Attorney General, accessed November 2026 - [AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive](https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/02/ai-other-companies-quietly-changing-your-terms-service-could-be-unfair-or-deceptive), Federal Trade Commission, accessed November 2026 - [UK GDPR Article 4: definitions](https://www.legislation.gov.uk/eur/2016/679/article/4), legislation.gov.uk, accessed November 2026 - [Anonymization](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/), Information Commissioner's Office, accessed November 2026 --- Book a free consultation: https://sellmydatatoai.net/book-a-consultation