Insights · 7 min read

How to Sell Data to AI Companies: A Step-by-Step Guide for Owners

Nine steps from “we might have something” to a signed license and a clean delivery.

Bearish AI · October 21, 2026

Selling company data to an AI lab is a project with a sequence, not a one-off email. Public reporting shows buyers include foundation AI labs, vendors that build simulated workspaces for training, and large technology companies building their own models. What they want is data that is lawful to use, clean enough to train on and described well enough to approve. The steps below follow that logic.

Step 1: Confirm that you can sell it

Start with rights, not with buyers. Who created the material, under what agreements, and what did you promise your customers about their information? Third-party code, customer records and anything covered by confidentiality clauses may be off limits or need consent. This is the step where a lawyer earns their fee, and it is cheaper to find a problem now than during a buyer’s review. Everything here is general information, not legal advice.

Step 2: Decide what you are offering

Be specific about the unit. An archive of code, chat and tickets is one product. A set of graded coding tasks built from your bug history is another. Pricing in public reports varies by unit: interviewees in Epoch AI’s study mostly cited about $200 to $2,000 per task for reinforcement learning work, while other reported deals license a whole body of content for a fixed term. Decide too whether you will license non-exclusively or exclusively. The same interviews say exclusive deals cost roughly four to five times as much as non-exclusive ones, which is a reason to price exclusivity deliberately and not give it away.

Step 3: Prepare the data

Rotate secrets, pseudonymize people, remove duplicates, keep the links between records, and leave out what you cannot license. The full order of work is in /insights/prepare-company-data-for-ai-labs. If you cannot do this yourself, this is where a preparation partner fits.

Step 4: Build a sample and a data card

Give buyers something to evaluate: a small slice made with the same process as the whole, plus a one- or two-page description of sources, dates, what was removed and the license basis. Our guide at /insights/data-cards-and-provenance-for-ai-training-data lists what to include. For code, a handful of finished tasks with tests is more persuasive than a folder of repositories.

Pricing models you will meet

• Per task. Graded, verified tasks priced one by one, mostly in the low thousands of dollars or less in Epoch AI’s interviews.
• Per item. HarperCollins’s reported book-licensing deal paid $5,000 per title, split evenly with the author, and a draft document-licensing offer described by the Columbia Journalism Review offered a thousand dollars per case.
• Flat fee for a body of data. Reported archive sales from shut-down startups have been put at $10,000 to $100,000 per company.
• Term license. Reddit’s prospectus disclosed contracts of two to three years with an aggregate value of $203 million.

These are reported figures for different assets, so they cannot be averaged into a price for yours. They show the main choice: sell something once, or license it for a period.

Step 5: Choose a route to buyers

Whatever the route, ask who handles the unscrubbed data and under what contract. Reporting from founders who have sold archives says they had little visibility into who bought their data or how it was used.

In the first conversation, expect to be asked what the data is, how much there is, how it was made safe, whether anyone else has it, and what you want to allow. Having those five answers written down saves a round of emails.

Step 6: Share under NDA and expect diligence

Buyers will test the sample, ask for provenance, and check contamination and privacy. See /insights/ai-lab-data-due-diligence for the questions. Have your rights memo, data card and removal counts ready, so answers take hours and not weeks.

Step 7: Negotiate the license

Read the grant carefully. One draft license reported in October 2026 for documents offered a non-exclusive grant described as royalty-free, perpetual, irrevocable, transferable and sublicensable, which is far broader than many sellers expect. Field of use, term, exclusivity, warranties and what happens at termination all matter. Our clause guide is at /insights/ai-data-licensing-agreement-terms. Bring a lawyer who has seen data licenses.

Step 8: Deliver, get acceptance, get paid

Define delivery format, acceptance criteria and payment timing in the contract. In one reported bankruptcy sale of a company’s business data to Google, the privacy ombudsman asked that a certification be filed before each transfer, including how many individuals’ details were found and removed. You do not need a court to borrow the idea: log what you delivered, what you removed and when.

Step 9: Keep records after signing

Keep the rights memo, data card, hashes of delivered files and the removal log. Licenses in the public record run for fixed terms; Reddit’s prospectus described data licensing contracts of two to three years. Terms end, renewals come around, and a clean file makes both easier.

Before you start, the free estimator at /estimate gives a conservative range for your data as it is and as prepared, which helps you judge whether the effort and the route make sense.