How to sell company data to AI companies

AI developers are building models that can do office work: answer tickets, reconcile accounts, triage bugs, draft contracts. To train them, they need real examples of that work being done, with all the mess and judgment calls left in. Your company's records are exactly that.

Buyers have noticed. OpenAI and a training data vendor reportedly asked contractors to upload real documents, spreadsheets and code from their past jobs (TechCrunch, reporting on Wired, January 10, 2026), and a service that helps startups shut down said it handled about 100 sales of old Slack, email and Jira archives to AI companies in a single year (Gizmodo, reporting on Forbes, April 17, 2026).

If you'd rather be paid for your records than find them in someone's training set by accident, here is how a sale usually works.

1. Take stock of what you have

Start with a plain inventory. For each system your team works in, write down:

What kind of record it holds (tickets, SOPs, CRM activity, project history, code, email)
How many years it covers and roughly how many records there are
Who owns it internally and who would need to approve sharing it
Whether it contains customer, employee, health or financial details

You don't need exact counts yet. "Six years of Zendesk tickets, about 400,000, full of customer names" is enough to tell a buyer whether it's worth a conversation. Our guide to what data AI companies buy covers which of these are in demand.

2. Check that you're allowed to share it

Owning a system doesn't always mean you can license everything in it. Before you go further, look at:

Customer contracts and data processing agreements. Some restrict how you may use customer data, even in anonymized form.
What you've told employees. Your handbook and privacy notices describe how work communications may be used. Surprising people is a bad idea even when it's legal.
Your own privacy policy. The FTC has warned that quietly changing terms to allow AI training can be unfair or deceptive (Federal Trade Commission, February 2024). If the policy doesn't cover this, that's a problem to fix openly or a reason to leave that data out.
Privacy laws. GDPR, state privacy laws like the CCPA, and HIPAA for health information all shape what can be shared and how it must be cleaned.
Third-party material. Vendor documentation, licensed content and open-source code come with their own terms.

This is the step to bring your lawyer into. It is much cheaper to scope things out now than to unwind a signed deal.

3. Decide what's in scope

Start narrower than you think. One or two well-documented systems usually beat a dump of everything, and they are far easier to clean. Leave out HR files, privileged legal advice, board and M&A material, and anything involving health information unless you have a specific plan for it.

4. Find the right buyer

Labs rarely buy directly from ordinary operating companies. Most deals go through vendors that assemble training data for them, and closing companies often sell through shutdown services. Each works differently, pays differently and treats your data differently. See who buys company data for how to compare them.

5. Agree the terms

The license decides who can use the data, for what, for how long, and what happens if something sensitive slips through. The terms that matter most are use restrictions, exclusivity, deletion, liability and the payment schedule. Our data licensing agreement checklist goes through each one.

6. Clean the data and hand it over

Personal and confidential details come out before transfer, and you review samples of the result before anything is used. The cleanup is usually the longest part of the project. Our guide to anonymizing company data covers what to remove and how to check it.

Mistakes to avoid

Letting employees sell their own work. When individuals upload old files, the company gets nothing and still carries the confidentiality risk. An IP lawyer quoted on the OpenAI program called that approach risky for exactly this reason (TechCrunch, reporting on Wired, January 10, 2026).
Sending raw exports. Once unredacted data leaves your systems, you're relying entirely on someone else's controls.
Signing an exclusive deal cheaply. Exclusivity stops you licensing the same records to anyone else. It should come with a clear premium and a time limit.
Treating it as a one-off. If your records keep growing, a refresh agreement can be worth more than the first delivery.

Common questions

Can a small company sell its data to AI companies?

Sometimes. Most buyer programs look for around 30 employees or more, because smaller teams rarely have enough records. Closing startups are the exception: several have sold their archives with much smaller teams.

Do we need our customers' permission?

It depends on your contracts. If a customer agreement or data processing agreement limits how you use their data, those limits still apply after redaction. Check them before you decide what's in scope.

How long does a deal take?

Plan for several weeks. Most of the time goes into internal approval, the contract and the cleanup, not into finding a buyer.

Sources

TechCrunch, reporting on Wired, January 10, 2026: OpenAI is reportedly asking contractors to upload real work from past jobs
Gizmodo, reporting on Forbes, April 17, 2026: Failed companies are selling old Slack chats and email archives to train AI
Federal Trade Commission, February 2024: AI (and other) companies: quietly changing your terms of service could be unfair or deceptive