The offer came as an ordinary cold email. A sender named Nick wrote that AI labs including Anthropic and OpenAI were paying media companies $350,000 to $1.8 million as a one-time fee for a copy of their old work records: documents, emails, Slack messages, support tickets. A second pitch raised the ceiling to $2 million if call transcripts were included, and mentioned that public text is running out, with labs already spending more than $45 billion on private data.
WIRED's Steven Levy followed the thread, and found an industry larger than the pitches suggest. A group of young companies now buys the internal paperwork of operating businesses, strips out names, and resells the result to model builders. The best known are Mercor, Micro1 and Handshake. All three say they work with most major labs. Anthropic told WIRED it does not buy workflow data from live companies like the magazine, and OpenAI did not answer before publication.
The pitch arrives in your inbox
The offers are specific enough to price. Handshake, which licenses professional documents across several fields, offered a freelance writer $1,000 for her work on a single story. She declined. A company representative confirmed the offer and said a large part of the business is buying graduate students' papers and coursework.
A third category is emerging around failure. One company, Sunset, emails founders of shutting-down startups and offers to run the wind-down, scrub identifying details from the company's records, and broker the sale to labs. Others go directly to employees, which is how the pitches reached a journalist who cannot sell anything.
Mercor's chief executive, Brendan Foody, explained the demand in practical terms. The labs had been hiring experts to act out workplace tasks, and the simulations did not capture how messy a real business is. His company started in 2023 as an AI recruiter and moved into data when it saw how unhappy buyers were with fabricated workflows. Its three cofounders became billionaires at 22 last year. Anthropic, for its part, draws a line at live companies, which suggests the market is segmented by how close the material sits to a working business.
Spirit's bankruptcy sold 500 million messages
The clearest window into the trade opened when Spirit Airlines went through bankruptcy and put its digital assets on the block. The corpus included 80,000 email accounts, 500 million internal messages, 20.6 million SharePoint files and a million time-card records. Google appeared to win with a $10 million bid. Micro1 then bid $2.5 million more, and the parties began arguing over who had actually won. The sale is now on hold.
The flight attendants' union objected, and the Electronic Privacy Information Center filed an amicus brief on its behalf. "It's not clear that anonymizing can fully get rid of privacy issues," EPIC general counsel Mayu Tobin-Miyaji told WIRED. Emails and Teams messages are unstructured, she said, and people discuss medical appointments and family problems in them.
That objection runs against the industry's central claim. Every buyer says personally identifying information is stripped before training, and that the process can be verified. Gordon Ritter, a venture capitalist who has taken up the issue, put the gap more bluntly: the norms for data acquisition have not been set, and the pace is partly deliberate. Our own reporting on what Claude can read from two years of bank transactions made a similar point from the consumer side, where consent is a checkbox and the data is somebody's actual life.
A revenue line bigger than ads
Micro1's chief executive, Ali Ansari, expects the market to keep compounding. Labs, enterprises and robotics companies pay $40 billion to $50 billion a year for data today, he said, a figure he did not provide documentation for, and he expects that to pass $1 trillion within two to four years. He also pointed at startup flow as evidence: 40 of the 65 companies in the most recent Y Combinator batch are classed as data businesses.
His prediction for how this lands inside ordinary companies is the part worth arguing with. Asked whether selling workflow data would become a default corporate practice, he said yes, and that it would eventually outrank advertising as a revenue line. Workers, in his account, will accept it because the models get better and their own jobs get easier.
The Spirit case is the test of that argument. A bankrupt airline's messages were priced at roughly $10 million, then contested by the people whose appointments and shift swaps were inside them. If the sale resumes, the precedent is that a company's records can be liquidated without the consent of the people who wrote them. If it does not, the market moves to the next wind-down, and the next inbox.
Regulators have already shown they will act when the story is about concealment rather than collection, as with Meta's $25 billion fine over how it handled data disclosures. Buying files from a company that is still trading is a different question, and no one has answered it yet.