How to sell your company's data to AI companies

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.

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 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 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. 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. 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. 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. 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.

5. 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. 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. 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. 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. 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. FTC staff wrote in February 2024 that adopting more permissive data practices, including AI training, through a surreptitious, retroactive change to a privacy policy may be unfair or deceptive. The CCPA also treats a transfer of personal information for value as a sale. Our guide on whether you can legally sell company data goes through each law.

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.

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, or try the data value estimator.

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.

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.

Our comparison of AI data marketplaces, brokers and direct deals 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 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 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.