Insights · 6 min read

How Much Is My Company Data Worth to AI Companies Today?

Reported deals run from about ten thousand dollars to ten million. Where you land depends on a handful of things you can check today.

Bearish AI · October 21, 2026

If you run a company that has been writing code, chatting in a team workspace and filing tickets for years, you have probably wondered whether any of it is worth money to an AI lab. The honest answer is: sometimes, and the range is enormous. This article gathers what has been publicly reported, so you can place your own archive on the map without guessing.

What has actually been reported

There is no price list for company data. What exists is a handful of press reports and a few research interviews. Treat every figure below as a data point, not a quote for your archive.

The airline figure is the outlier on purpose. It covers a very large company: lawmakers describing the sale cited 100 million emails and 500 million Teams messages. Almost no startup looks like that. The wind-down range and the single founder sale are far closer to what a small company should expect.

What buyers say they look at

In MIT Technology Review’s reporting, the chief executive of a wind-down firm said a company’s data is valued on headcount, years in operation, data volume and how the team used its tools. A buyer’s environments lead described a different method: “research theses” that estimate how many useful tasks a body of data can yield, for example per 100,000 lines of code. The article notes that buyers quote widely different prices as a result.

In plain terms, buyers are not paying for files. They are paying for the number of training tasks they can build and trust. That is why a ten-person team with eight years of history and a ticket trail can be more interesting than a hundred-person team with one year of scattered documents.

Who is buying, and what they want it for

Gizmodo’s report says the archives are used to train agentic AI models in simulated workplace settings, sometimes called reinforcement learning gyms. Epoch AI’s interviews point the same way: reinforcement learning is the main use, and one startup employee said they receive perhaps 10 to 20 times more requests for it than for benchmarking. That shapes what is valuable. A buyer building a practice workplace wants the texture of a real one: the arguments, the tickets, the fixes, the half-finished decisions. A folder of final documents has less of that than a team’s working history.

Three archives, three rough answers

The following is our own rough framing from the reported data points, not a set of reported prices. Use it to calibrate your expectations, not to set a number.

• A young, small team. Two years, five people, little linked history. Expect a small figure, and possibly nothing worth selling on its own.
• An established team. Several years, a few dozen contributors, code with history plus chat or tickets. The reported range of tens of thousands of dollars is the realistic neighborhood.
• A very large organization. Hundreds of millions of messages, as in the airline case. This is rare, it needs a legal and privacy process of its own, and it is nothing like a startup sale.

Five things that move your number

• Years of recorded work. A long, continuous history shows how decisions, bugs and fixes unfold. A single snapshot does not.
• People and output. More engineers over more years means more episodes of real work. The same arithmetic is why the figure for a tiny, young team is small.
• Linked sources. Code alone is one thing. Code plus the chat that explained it plus the tickets that tracked it lets a buyer rebuild the story of a change.
• Rights. If you cannot show that you own or can license the material, it is worth far less, because buyers usually ask for a warranty that you have the right to license it.
• Cleanliness. Suppliers scrub personal information and leave out HR records, and buyers scrub again. Data that arrives clean costs the buyer less to use.

What pulls it down

Be skeptical of anyone who promises a seven-figure number without looking at your material. Small archives, short histories, heavy personal data, unclear ownership and thin documentation all pull the figure toward the low end of the range, and some archives are worth little on their own. An attorney quoted by MIT Technology Review warned that even emails and calendars can contain private details such as doctor appointments, which is the kind of content that has to be found and removed before anything is shared.

Timing matters too. The same article says prices will likely stabilize in the coming year, unless an AI bubble pops first. Nobody can promise today’s prices will hold.

A fair way to get a first number

Write down the number of people who contributed, the years covered, which kinds of records you have (code, reviews, tickets, chat, design files, recordings, support conversations, business documents), what language the work is in, and whether anything involves customers’ personal data. Then ask a lawyer who knows data licensing who owns what. That is the minimum a buyer will ask for anyway.

If you want a quicker start, the free estimate at /estimate asks for exactly those inputs and returns a conservative range for your archive as it is and after preparation. It is deliberately modest, so there is room to do better in a real review. For the market side of the picture, see /insights/how-much-do-ai-companies-pay-for-training-data.

This is general information, not legal or financial advice. Ownership, customer contracts and privacy law decide what you may sell, so consult a lawyer before you share or license anything.