Your company data is worth what an AI lab will pay to license a de-identified copy of it. That price depends less on how much data you have than on whether you can legally license it, how many years it covers, and whether it records decisions and their outcomes.
There is no public price list for business records. Deals are private, and the few public figures come from large media companies, press reports and interviews with vendors that sell to labs.
The public price signals
A handful of published figures give a sense of scale. None of them is a price for your data, but they show what labs pay for different kinds of material.
| Signal | Figure | Source |
|---|---|---|
| Shut-down startups selling Slack, Jira and email archives | $10,000 to $100,000 per company | Forbes, April 2026 |
| Expert-written RL task for training AI agents | $200 to $2,000 per task, rarely $20,000 | Epoch AI |
| Replica of a website or app for agents to practice in | About $20,000; a complex product about $300,000 | Epoch AI |
| Reddit's data licensing contracts | $203 million in aggregate, two to three year terms | TechCrunch |
The Forbes figures are the closest match for an ordinary company. Those were archives from businesses that had closed, so they were one-time sales with no new records coming. An operating company has history plus a steady supply of new records, which can be licensed again each month. Our guide on what AI labs pay for data covers more deals and what they included.
What makes the number go up
Clear rights
Data you can't legally license is worth almost nothing, however good it is. Labs need a statement that you have the right to sell, and they check it in diligence. If you hold the data on behalf of customers, you may need their written consent first. See can I sell data I hold for my customers?
Years of history
Ten years of ledgers or tickets show how practice changed, how rare cases were handled and what happened over the long run. A few months of data rarely justifies the work of a deal.
Decisions and outcomes
Labs pay most for records that link a situation to what a person did and how it turned out: the invoice that was paid late and chased twice, the experiment that failed at 80 °C, the claim that was denied and then reversed on appeal. Audit trails let us turn those records into episodes, which sell at prices closer to the task figures above than to bulk data. Our guide on episodes, records and environments explains the difference.
Rarity
Data that a lab can't get elsewhere is worth more. Published papers mostly report experiments that worked, so full lab notebooks with failures are scarce. Semiconductor yield histories and insurance claims files are hard to find for the same reason. Generic records with standard SaaS fields are easy to find and worth less.
Volume and team size
More people working in a system means more records, more variety and bigger monthly refreshes. Volume helps, but it multiplies the factors above rather than replacing them.
Exclusivity
A license that lets only one buyer use the data costs more. Two founders interviewed by Epoch AI put exclusive deals for training environments at roughly 4 to 5 times the non-exclusive price. Our licenses are non-exclusive by default, which lets the same dataset earn from more than one buyer.
What makes the number go down
- Unclear rights or restrictive customer contracts
- Free-text fields full of personal details that have to be removed
- Data that is mostly boilerplate, templates or default fields
- Short history or gaps in the record
- Content that falls under sector rules, such as health records, consumer financial data or export-controlled process data, which need their own review or are excluded
How our calculator estimates a range
Our calculator asks five questions: your industry, the systems you would consider sharing, years of records, team size, and whether you can license the data. Each system has a base value for about five years of history at a company of 11 to 50 people. The total is adjusted for your industry, the years of history and your team size.
Rights change the estimate directly. Answering "not sure" cuts it by 30%, and "customer contracts restrict it" cuts it by 70%, because a lab won't pay full price for data it may not be allowed to use. We would rather show a lower number than hide the problem.
The calculator gives an indicative range, not an offer. The real number comes from diligence: a sample, a rights review and buyer interest. We confirm the estimate on a 30-minute call.
History pays once, refreshes keep paying
The first license sells your past records. After that, every month your team works creates new records, and a license with a refresh term pays for each monthly batch. So the value of your data has two parts: a one-time price for the history and a recurring payment for each refresh. Read how that works on our recurring data revenue page, or the longer explainer.
Get your estimate
Answer five questions and see a range for your company in about a minute, with no access to your systems needed. Value your data, then book a call if the number is worth pursuing. Industry-specific signals are on our price pages, such as accounting and software.
Frequently asked questions
Is there a price list for company data?
No public price list exists for business records such as ledgers, CRM exports, claims files or lab notebooks. Deals are private, so estimates rely on the few published figures, comparable task prices and the specifics of your data.
Why does the calculator lower my estimate when I answer "not sure" about rights?
A lab can only pay for data it is allowed to use. When rights are unclear, the estimate drops by 30% until we review your contracts on the call. If contracts restrict the data, it drops by 70%.
Is more data always worth more?
Volume helps, but rights, years of history and recorded outcomes matter more. A smaller dataset with a clean audit trail can be worth more than a large export of default fields.
Do I get paid once or every month?
The first license pays for your history. If the license includes a refresh term, each accepted monthly batch of new records pays again, with no added work once our SDK is scheduled.