AI companies are buying up work emails and documents

AI companies and intermediaries are offering organizations and individual professionals money for archives of work documents, emails, messages and other records, to train models on examples of real work. Executives at Mercor and Micro1 say demand for this data is growing quickly. Micro1 estimates that labs, enterprises and robotics companies now spend around $40–50 billion a year on it, and forecasts that spending will exceed $1 trillion within the next two to four years. But selling archives can expose employees’ personal information, even when the data is de-identified in advance.

AI companies are buying up work emails and documents

What intermediaries pay for

A Wired writer began receiving offers to sell work archives. One email claimed that Anthropic and OpenAI pay media companies between $350,000 and $1.8 million upfront for copies of old documents, emails, Slack messages and support tickets. Another offer included call transcripts and proposed a payment of up to $2 million. The sender, Cora, said public data had already been exhausted and that labs were spending more than $45 billion on private data.

The market turned out to be broader than the writer had expected. Companies are collecting not only organizations’ documents but also material from individual professionals. Handshake offered a freelance writer $1,000 for one article; she declined. A company representative confirmed the offer and said Handshake licenses professional documents from different industries so labs can evaluate and train models to handle real professional tasks. Journalism is one such field. The company says a significant part of Handshake’s business consists of buying research papers and graduate students’ assignments.

An entrepreneur known to the writer had also received emails from companies that help startups shut down. In one, a company called Sunset described its service this way: clean and de-identify the data, remove company information and any details that could identify individuals, then arrange to sell the material to labs.

New intermediaries are trying to replicate the success of market leaders Handshake, Micro1 and Mercor. These companies buy up large datasets. Founded in 2023, Mercor has become the industry’s biggest player: last year, its three co-founders became billionaires at the age of 22. Like other companies, Mercor says it partners with nearly every major AI lab. Anthropic acknowledges that the data market is large, but told Wired it does not buy work archives from active companies such as Wired itself. OpenAI had not responded by publication time.

Why demand is growing

Mercor CEO Brendan Foody said the company began as an AI hiring service, but later discovered that labs were dissatisfied with the quality of their data. Companies were hiring specialists to imitate work processes, but those exercises failed to capture the messiness and complexity of a functioning business. In Foody’s view, the AI industry has two insatiable needs: computing power and data. Investment in the first gets widespread attention, but data collection is just as important.

Foody believes privacy concerns can be addressed: when a company gives data to Mercor, the service removes personal information, or PII. He says the de-identification process can be audited. Representatives of other data-collection companies also believe technical solutions can manage the risks.

Micro1 CEO Ali Ansari expects spending on data to approach the cost of AI computing infrastructure. He estimates that labs, enterprises and robotics companies currently spend around $40–50 billion a year on data. The writer could not verify that figure. Ansari, who founded Micro1 in 2022, forecasts that spending will exceed $1 trillion a year within two to four years. Micro1 uses the data it buys primarily to improve its own models, but also sells it to major AI companies.

Ansari says almost everyone is moving into data collection now because demand is high. He claims that around half of the startups in Y Combinator’s accelerator work with data. In the latest cohort, the writer notes, the number was even higher: 40 of the 65 startups launched were classified as data companies.

The sale of Spirit Airlines’ archives

The risks and opacity of the market came into view after Spirit Airlines went bankrupt and its digital assets were put up for sale. The archive included 80,000 email accounts, 500 million internal messages, 20.6 million Microsoft SharePoint files and 1 million timekeeping records.

Google apparently won the bidding with an offer of $10 million. Micro1 then raised the bid by $2.5 million, triggering a dispute over who had won. The sale is now on hold: the flight attendants’ union said their personal information was included. The dispute calls the practice of selling records of people’s work into question.

Mayu Tobin-Miyaji, chief counsel at the Electronic Privacy Information Center, which filed a court statement in support of the flight attendants, warns that de-identification does not necessarily eliminate every risk. Emails and Teams messages contain unstructured information that is harder to filter out. Employees may discuss doctor visits and family problems there.

Venture investor Gordon Ritter, who opposes current practices, says no rules or standards for data collection have yet been established. In his view, the AI industry is moving faster than society can agree on rules, and that is happening in part by design.

Whose data and whose work

One of the main concerns is that employees’ records may be used without their explicit consent to develop technology that could automate their jobs. Mercor describes its goal as organizing human intelligence to advance the AI economy. Yet, as the writer notes, the eventual result could be less work left for people to do.

The president of the flight attendants’ union said in an interview that data had become valuable in the AI era because it can train models to replace jobs and turn other parts of working life into commodities.

Ansari believes companies will eventually sell data about their work processes to major AI developers on a regular basis. He predicts this could become a more common source of business revenue than advertising. When asked whether employees might object to everything they do at work being fed into large models, Ansari said he understood the concern. He believes employees may accept the practice if it improves AI and makes their lives easier. The writer observes that promises of future abundance through AI are often used to justify concessions in the present.

An offer for Wired

The writer also investigated one of the emails, which came from the domain gorexen.com. “Gorexen” is a fictional creature made up of two Pokémon; it drains fluids from its prey through a thin, tubular mouth. The sender arranged a meeting with Tenkara, a company founded two years ago. The writer guesses that its name refers to a Japanese style of fly fishing.

At the meeting, the writer spoke with Jonah Stillman, head of commercial operations. Tenkara was collecting data to improve a model for managing office processes at manufacturing companies. How Wired’s archive could help with that remained unclear. Stillman confirmed that a significant part of his work involved calls like this, aimed at buying data. The writer compared the meeting to the film Glengarry Glen Ross, but Stillman did not get the reference.

The writer is not authorized to sell Wired’s editorial work documents to AI companies and was unwilling to risk his job by offering the archive to Stillman and Tenkara. He would also have been unhappy if management had agreed to a deal.

In Dataist’s view, the market for work data is growing faster than rules for consent and personal-information protection are taking shape. Even when companies remove direct identifiers, the Spirit Airlines case shows that disputes can arise over the contents of an archive itself and employees’ rights to have it sold. And the promise that AI will make work easier does not resolve the question of who decides how workers’ records are used and who benefits.