How the market for company data works

Over the past year a new market has opened for something most businesses never thought of as an asset: the records of how their work got done. Email and chat, support tickets, CRM history, project files, SOPs and code are now licensed to companies that build and train AI. This is how that market is put together.

Three tiers: labs, data companies, sellers

At the top are a handful of frontier AI labs. OpenAI, Anthropic and Google are reported customers of the large data companies. The labs rarely buy a company’s records directly. The one documented exception is Google’s winning bid for Spirit Airlines’ business records in the airline’s bankruptcy.

In the middle are the data companies the labs already pay for expert work: Mercor, Scale AI, Handshake AI, Turing, micro1, Surge and others. These are the companies that run programs to license records from businesses. They turn what they license into training environments, evaluations and training data, and sell that work on to the labs. Labeling companies such as Labelbox, SuperAnnotate, Centific and Appen have opened programs of their own.

On the seller side sit the companies whose records are licensed, and the firms that help them. Wind-down specialists such as SimpleClosure and Sunset, whose data arm is Replay, sell the records of startups that have closed. Marketplaces list datasets for buyers. Brokers and advisors match companies to programs, and most are paid by the buyers through referral fees. That includes us: participating buyers may pay us referral fees, and we tell an owner which ones before any introduction.

What the data companies do with the records

Labs are building AI agents that work inside business software. To train them, data companies build simulated workplaces where a model practices real tasks and is scored on the result, and real company records are the raw material. We explain that in a companion piece, “Flight simulators for office work.” The same records also feed evaluations, which test how well a model does professional work.

Three kinds of deal

• A one-off license of a de-identified copy. The most common structure. The company keeps ownership of its records, the buyer licenses a copy with personal information removed, and payment comes once the data is accepted. micro1, for example, says companies keep ownership of their underlying data.
• A recurring partnership. Some buyers license an ongoing feed rather than a snapshot. Scale AI describes its program as a continuous partnership with recurring revenue, and Centific says most partners move to a quarterly refresh.
• A court-supervised sale. In a bankruptcy, a company’s records can be sold as an asset. Spirit Airlines’ sale is structured as a de-identified copy, certified by an independent de-identification agent, with passenger data carved out.

Around those three, the terms vary widely. Replay pays cash up front plus a share of every later license. Labelbox Prism licenses exclusively by default, and Polyshares’ published deals are perpetual exclusive licenses. SuperAnnotate lets the seller approve each buyer and block competitors. These are the clauses worth reading most closely.

How the records change hands

Handover runs along a spectrum. Some programs connect to a company’s tools and extract and de-identify automatically. Others export the data themselves, or work on an isolated copy. Scale AI says it removes personal information inside the company’s own environment, before anything leaves it. Which method a buyer uses decides how much of your raw data ever leaves your hands.

From shut-down startups to the mid-market in six months

The idea that data is licensable came from publishers, which began signing content deals with AI labs in 2024. The supply chain came from expert-data companies, which already sold human judgment to the labs. Demand for company archives followed in 2025, as labs looked beyond public text.

• December 2025: Turing launches Project Lazarus to buy records from companies that have shut down.
• April 2026: SimpleClosure adds data sales to its marketplace for closed startups.
• July 2026: Mercor acquires Deeptune, which builds simulated workplaces for training AI agents.
• August 2026: Google wins the auction for Spirit Airlines’ business records; micro1 later files a competing bid.
• October 2026: micro1 commits to a year of buying company data, financed by Citi and Hercules Capital. Mode opens a program for businesses and nonprofits, and Mercor’s CEO says more than five hundred companies have licensed their data to it.

In about six months the market moved from the archives of closed startups, sold through wind-down firms, to operating mid-sized companies applying directly to buyer programs. Almost every program in our buyer index is under a year old.

How big it is

The data companies in the middle tier have grown quickly, but no analyst firm yet measures the market for companies’ operating records on its own. The firmest signals are the buyers’ own: micro1’s lender-backed twelve-month commitment, Mercor’s count of more than five hundred partner companies, and SimpleClosure’s more than one hundred and fifty completed deals, nearly half of which now include code or data. Each program’s published payout range is on our buyer index.

What this means for an owner

Our view: the market is real, young and still setting its terms. A company that fits one program usually fits several, and the programs differ less on price than on exclusivity, how long the license runs, who removes personal information and when you are paid. Compare at least two before signing anything, and have an attorney read the terms.