Five kinds of buyer, and what each wants
Public text taught models to write. It didn't teach them how a month-end close gets reconciled, how a contract gets redlined or how a roof claim gets approved. That working knowledge sits in operating records, and these are the buyers who want it:
- Data labs. Suppliers that turn real workflows into training sets, tasks and evaluations for AI developers. They want records with a clear sequence and an outcome, such as a ticket from bug report to fix, a claim from first notice to payout or a job from lead to invoice, because those convert cleanly into tasks. They also bring the staff to review and package the material.
- AI labs building agents and evaluations. Teams training and testing AI that does multi-step work in business software. A good task needs an outcome that can be checked. Operating records arrive with outcomes attached: booked or lost, approved or denied, merged or reverted, paid or not.
- Vertical AI companies. Software companies building for one industry, such as legal and contract tools, accounting and finance automation, insurance, logistics, the trades, construction or property management. They want the domain's vocabulary, its paperwork and the way work unfolds, so their product can handle a real redline, reconciliation or permit packet instead of a textbook one.
- Insurtech and claims. Companies working on underwriting, damage assessment, estimating and claim handling. They want submissions tied to quotes, and photos tied to scopes, estimates, supplements and decisions, because a labeled outcome is what turns documents and pictures into training data.
- Voice AI. Companies building phone agents for booking, intake and customer service. They want calls, usually as transcripts, matched to what happened next: the intent, the objections, whether the job was booked or the problem solved.
The records come from most industries, not one. What industries AI companies buy data from covers which functions and industries buyers ask for, and which need special handling.
Why buying straight from an operating business is hard
A buyer can call a roofing company or an accounting firm and ask for its files, and some sales do happen that way. It's hard going, for ordinary reasons.
- Neither side can find the other. A buyer doesn't know which company holds ten clean years of records, and the company doesn't know which buyer is serious or will protect its customers.
- Every purchase needs a rights review, de-identification and a contract. Doing that one small company at a time is slow for both sides.
- Much of the work of building AI training environments goes to specialist suppliers, and some AI developers build in-house too. That's one reason data labs sit between operating businesses and the largest buyers.
- With one buyer in the room, that buyer also sets the terms.
A broker's job is to close those gaps: find the buyers, prepare the record, and make them compete.
How we qualify buyers
A buyer starts by requesting access and telling us who they are and what they're building. We check that, along with how they'd use licensed records. If it's a fit, they sign an NDA before they see anything.
After that, a buyer sees anonymized teasers for packages that match its interests and bids in rounds. Every license carries the same use limits, with more available if you want them:
- No re-identification of any person or company in the records
- No contacting your customers or anyone named in the records
- No voice cloning from call audio or transcripts
- No marketing to the people or businesses in the records
- No resale or sublicensing beyond what the license allows
You can also rule out named competitors, whole industries or uses before bidding opens, so some packages are never offered to some buyers. The signed license, not this page, is what binds a buyer.
How your identity stays private until you sign
A teaser shows sector, region, bands for years of history, revenue and headcount, and rounded volumes by kind of record, such as jobs, tickets or call hours. It leaves out your company name, your website, any person and any address. Your identity goes only to the buyer you choose to sign with.
Anonymized isn't the same as invisible. A very small niche can be easy to guess, which is one reason the competitor block exists. Tell us who you don't want looking.
Which categories are a fit depends on what you have. A company with years of call recordings looks different to buyers than one with a deep archive of tickets or claim files. We can't promise any bids or any price. What we can describe is the process: qualified buyers, your terms set first, and nothing delivered until you sign.
Follow-up questions
Sources
- Epoch AI, Gradient Updates, "An FAQ on reinforcement learning environments" (Jan. 12, 2026): who builds training environments, including specialist suppliers and in-house teams, and what makes a gradable task (opens in a new tab)
- MIT Technology Review, "AI labs are buying data from dead startups to train their models" (Oct. 13, 2026): sales by closing startups through wind-down firms, and in a bankruptcy (opens in a new tab)
This is general information about how data licensing works. It is not legal advice, and we are not your lawyer. Have your own counsel review any license before you sign.