AI companies are fueling a rapidly expanding market for private workplace data, with startups offering to acquire documents, emails, Slack messages, support tickets, call records and other business materials for model training. Emails seen by the author claimed that labs such as Anthropic and OpenAI pay media companies between $350,000 and $1.8 million in one-time fees for old work records, while another pitch suggested deals could reach $2 million and that AI labs are spending more than $45 billion on private data.
The messages were presented as sales pitches. “I don’t know if you’ve heard, but AI labs like Anthropic and OpenAI pay media companies like Wired $350,000 to $1.8 million as one-time fees for copies of old work records (documents, emails, slack messages, tickets, etc.) to train better models,” wrote someone named Nick.
In a separate email containing similar claims, with call records added to the list of requested information, someone named Cora wrote: “Public data is running out, and labs are spending over $45 billion on private data to build better models.”
AI companies have been buying expert data from many sources, but the practice has received relatively little attention. Further research indicated that the business is larger than expected. People working in the industry say it could affect workers and society on a scale comparable to the data-center expansion that has dominated much of the recent AI debate.
A growing market for specialized data
New data-collection companies are quietly acquiring the workflows of large enterprises and continuing to seek more material from workers’ inboxes and other internal systems.
A freelance writer recently described how a company called Handshake offered her $1,000 for the work behind individual stories. She did not sell the material, and a Handshake representative confirmed the offer.
“We license specialized documentation across many disciplines, so the lab can test and train how models handle real-world knowledge tasks,” the representative said. “Journalism is one of those fields.” Handshake also says that a substantial part of its business involves buying graduate students’ papers and classes.
An entrepreneur told the author that he receives emails from several companies offering to “help startups downsize.” One such company, Sunset, described its service this way: “We handle the entire process, scrubbing and anonymizing the data to remove company associates and personally identifying details, and brokering contracts with laboratories.”
Startups pursue an AI data gold rush
The data-collection startups are seeking rapid growth by selling to major AI companies, many of which were founded only a few years ago. Industry participants include Handshake, Micro1 and Mercor, which are building large databases in what resembles a modern-day gold rush.
Mercor, founded in 2023, appears to be the leading company in the sector. Last year, its three co-founders became billionaires at 22 years old. Like other companies in the field, Mercor says it has partnerships with almost all major AI institutes.
The AI companies offered more measured responses. Anthropic acknowledged that there is a large market for this type of data, but said it had no intention of buying the workflows of a living company such as WIRED. OpenAI had not responded before publication.
Why AI labs want real workplace workflows
Mercor CEO Brendan Foody said in an interview that the market is poised to expand sharply as the AI industry demands far more data. Mercor began as an AI recruitment company, but its team later learned that major research institutions were dissatisfied with the quality of their training data.
AI companies have typically hired experts to simulate real-world workflows. Foody said those simulations fail to capture the complexity and disruption of actual business operations. His view is that the AI industry has two effectively bottomless needs: computing power and data. While investment in computing infrastructure is highly visible, he argued that collecting high-quality data is equally important.
Source: www.wired.com