What data do AI companies buy?

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.

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.

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, 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 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.

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 to CCPA, state privacy laws and GDPR 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.

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 for published ranges. If you want an honest view of whether buyers will be interested, including when they will not, book a free consultation.