Data licensing guide

Why AI labs want ordinary company data

Quick answer

AI models learned mostly from the public web, which shows what people write, not how work gets done. Labs now want AI agents that can carry out real tasks, and those need examples of real work: a ticket from problem to fix, a contract from draft to signature, a month-end close from start to finish. Those records sit inside companies. Public text is also getting harder to use, as more websites restrict AI crawlers. That's why programmes now pay for ordinary business records.

How, not what
records show work being done
2026–2032
Epoch AI's 2024 estimate for using up public text
28%+
of key web sources restricted for AI by 2024
Ordinary
boring records are the point
From the video (Ade): AI labs want real records of how work gets done, not more public internet text.

If you run a recruitment firm or a logistics company, it's fair to ask why an AI lab would pay for your old tickets and process documents. The answer comes down to the difference between knowing about work and doing it.

01The public web shows what people say

Today's models learned mainly from public text: articles, forums, books and code. That's why they write and summarise so well. But the public web holds few complete records of a job being done. Almost nobody publishes how they closed last month's books, or the 40 messages it took to fix a customer's problem.

02Agents need to see work being done

Labs are now building AI agents: software that takes a goal and works through the steps, using tools the way a person would. To do that well, an agent needs examples of:

  • Sequences. What happened first, next and last.
  • Judgement. Why someone escalated, approved or changed course.
  • Outcomes. Whether it worked.

A resolved ticket has all three. So do a project history, a contract with its redlines, and a code change with its review. Handshake puts it simply: "Business records help AI models learn how companies work."

03Public data is getting harder to use

Two trends push the same way:

  • The supply is finite. In a paper updated in June 2024, Epoch AI estimated about 300 trillion tokens of quality public text, with 80% confidence it would be fully used between 2026 and 2032 if trends continued.
  • Websites are closing the door. A 2024 audit of 14,000 web domains found 28% or more of the most actively maintained sources in one widely used dataset were fully restricted from AI use, most of that within a year.

Both are projections or snapshots from 2024, so treat them as direction, not fact. See AI training data statistics.

04So which records are valuable?

The ones that show a process with a result, over several years. That's why the programmes' own lists look so ordinary: SOPs, knowledge bases, tickets, CRM records, project histories. See what AI labs buy.

What makes records useful also makes them sensitive, because real work involves real people and clients. So check what's yours to license first.

05Is it worth anything to you?

Possibly. The public deals and published ranges are the evidence. Our free fit check shows which programmes you're likely to match. The programme decides.

06Questions people ask

Why would an AI company want my SOPs?

Because they show how a task is really done, step by step. That's hard to find on the public web.

What is an AI agent?

Software that carries out a series of steps towards a goal, rather than answering one question. To do that well, it needs examples of how those steps work in practice.

Will this demand last?

Nobody knows. Programmes' terms change often, which is why we log every change in our updates page.

Sources

Our links to these programmes carry our referral code, so the programme can see the introduction came from us. The programme may pay us a referral fee, and you pay nothing.

Check which programmes you’re likely to matchA few short questions, about a minute. Free, no sign-up. The programme decides who qualifies.trackimpact.io/data →