You sell company data to AI companies by licensing a de-identified copy of the records you already keep, after confirming your contracts allow it. The work runs in five steps: check your rights, choose the systems, scrub the data, find a buyer you approve, and sign a license that limits what the buyer can do with it.
Most companies that do this never hand over a database. They sell a cleaned extract of one or two systems, such as a ledger, a ticket history or a set of lab notebooks, under a license with an end date.
Why AI labs buy business records
AI labs have trained on most of the public internet. Epoch AI estimates that, if trends continue, language models will fully use the stock of human-generated public text between 2026 and 2032. What labs can't scrape is how real work gets done inside a company: the decisions people made in real tools and what happened afterward.
Labs now pay for that kind of material. Reddit disclosed $203 million in data licensing contracts signed in January 2024. Forbes reported in April 2026 that shut-down startups were selling old Slack messages, Jira tickets and emails, with one wind-down firm handling nearly 100 such deals in a year at $10,000 to $100,000 per company. Operating companies with years of records and steady new activity have more to offer than a closed startup's archive, because their data keeps growing.
Step 1: Check that you can sell it
Rights come first because they decide everything else. If you own the data outright, the question is short. If you hold data on behalf of customers, the answer usually depends on their written permission.
Read three kinds of documents:
- Customer agreements. A clause saying data is "used only to provide the service" blocks a sale, even after scrubbing.
- Data processing agreements. If you are a processor for your customers, you generally can't sell their data without their consent. Our guide on selling data you hold for customers covers this in detail.
- Confidentiality clauses and professional rules. Accountants, lawyers and regulated firms have extra duties on top of contracts.
Privacy law applies too. In the EU and UK, selling to a lab is usually a new purpose that needs its own legal basis, which our GDPR guide explains. In California, a transfer for money counts as a "sale" unless the data meets the statute's definition of de-identified, covered in our CCPA guide. This is general information, not legal advice, and you should have a privacy lawyer look at your contracts before any data moves.
On our first call we ask about rights before anything else. When we can't get a clear answer for a system, we leave it out of the deal.
Step 2: Choose the systems worth selling
Not every system is worth the effort. Labs pay most for records that show decisions and outcomes over time, and least for generic fields that look the same at every company. Good candidates include:
- Accounting ledgers with reconciliation history
- CRM records with deal stages and win or loss outcomes
- Support tickets with resolutions
- Lab notebooks that include failed experiments
- Claims files with decisions and payouts
- Plans, tickets and pull requests from a software team
Years of history matter, and so does an audit trail showing who changed what and when. Audit trails let us build episodes, which are step-by-step records of a person handling a situation. Our guide on what makes data valuable to AI labs goes deeper on the qualities buyers look for.
Step 3: De-identify the records
Removing names is not enough. People can often be picked out from combinations of ordinary fields, such as a company domain plus a job title, or an unusual amount plus a date. Research by Latanya Sweeney found that 87% of the US population was likely unique on just ZIP code, gender and date of birth.
Our scrubbing process does the following:
- Removes names, contact details, account numbers and secrets such as API keys, tokens and passwords
- Rounds rare values, so amounts become ranges and exact dates become a month or quarter
- Drops free-text fields unless someone has reviewed them
- Tests whether any record can still be singled out, and keeps the report
You review a sample of the output before any buyer sees it. Our de-identification vs anonymization guide explains the difference between the two terms and why it matters legally.
Step 4: Find a buyer you approve
Buyers include AI labs that train models for business, science and industry, plus companies that supply those labs with data. A buyer usually reviews a redacted sample before making an offer.
You approve every buyer before a license is signed, and you can rule out any company, including competitors. Nothing goes to a buyer until you say yes.
Step 5: Sign a license
A data license sets what the buyer may do with the data. A sound license names:
- Permitted uses, such as training, fine-tuning or evaluation
- The length of the license and what happens at the end
- Whether it is exclusive (ours are non-exclusive by default)
- A ban on re-identification, resale and combining the data with other sources
- Deletion of raw data, audit rights, and the same limits for anyone the buyer shares it with
The license also sets the price. Ours have two parts: a price for the history and a price for each monthly refresh. Our guide to data licensing terms walks through each clause.
Get paid more than once
A one-off sale pays for your history and stops. A license with a refresh term keeps paying as your company keeps working. After the first license, our SDK runs on your machine on a schedule, monthly by default. It pulls only the records created since the last run, scrubs them locally and uploads them as a new batch. Each accepted batch pays the refresh price for every license with an active refresh term, and you do nothing once the schedule is set. Raw data never leaves your machine through the SDK. See how recurring data revenue works.
What it costs you
We take a commission on each license. If nothing sells, you pay nothing. Your time goes into a 30-minute call, a contract review with us and an export or connection for the systems you choose. We prepare the data, find buyers and negotiate.
Find out what your records are worth
Our calculator gives an estimated range from your industry, systems, years of records, team size and rights. It takes about a minute and needs no access to your systems. Value your data, or read how the process works on our seller page.
Frequently asked questions
Is it legal to sell company data to AI companies?
It is legal when your contracts and privacy law allow it, which is why rights are the first thing to check. Data you hold on behalf of customers usually needs their written permission, and personal data has to be properly de-identified or have a lawful basis for the new use. Get a privacy lawyer to review before any data moves.
Who buys company data for AI training?
AI labs that train models for business, science and industry buy it, along with companies that supply those labs. Each buyer usually reviews a redacted sample first, and you approve every buyer before a license is signed.
What data sells best?
Records that show decisions and outcomes over several years sell best, especially when an audit trail shows who did what and when. Generic fields that look the same at every company are worth little on their own.
Can I keep earning after the first sale?
Yes, if the license includes a refresh term. Our SDK sends new de-identified records each month, and every accepted batch pays the refresh price. Read more on recurring data revenue.