AI companies are looking beyond public material for examples of how work gets done. Emails, documents, internal messages, coursework and other records can show the detail and unpredictability of a working organization, making them valuable to companies building or testing AI models.
A growing data acquisition business is arranging access to those records. The deals raise practical questions about privacy, consent and the future use of workers’ contributions.
From company records to AI training data
Data brokers have approached media companies and other organizations with offers to buy or license collections of past work. The material described in these pitches includes documents, emails, Slack messages, tickets and call transcripts. One writer was offered $1,000 for work on a single story and declined.
Handshake says it licenses professional documents across fields so labs can test and train models on real knowledge work. The company also says graduate students’ papers and coursework make up a significant part of its business. Other firms seek data from companies that are closing, offering to scrub and anonymize records before arranging a sale to AI labs.
The appeal is that records from an operating business capture the complicated details of real work. AI companies have used experts to simulate workflows, but industry participants say those exercises do not reproduce the messiness of an actual company. Access to existing records offers another way to provide models with examples of how people handle professional tasks.
A market growing around data demand
Handshake, Micro1 and Mercor are among the firms competing to acquire data. Mercor began as an AI recruiting company, and its CEO, Brendan Foody, says the market could grow as AI companies seek more data. The industry’s need for computing power is highly visible; collecting data is another important part of building AI systems.
Micro1 CEO Ali Ansari says labs, enterprises and robotics companies pay about $40 to $50 billion a year for data. The article says that figure could not be verified. Ansari predicts the market will exceed a trillion dollars a year in the next two to four years. Micro1 uses purchased information largely to improve its own AI models and also sells to large AI firms.
The companies’ growth reflects a broader rush to supply data. Ansari says many startups are entering the field, and the article notes that 40 of the 65 launched companies in the most recent YC batch were classified as data startups. These claims point to how quickly data acquisition has become a business in its own right, though they do not settle what standards should govern the work.
Privacy claims meet concerns about real records
Data companies say they remove personally identifying information before sharing records. Mercor’s Foody says that process can be verified. But critics argue that removing identifiers may not address all the risks, especially when source material consists of informal, unstructured conversations.
The proposed sale of Spirit Airlines’ digital assets illustrates the dispute. The collection included email accounts, internal messages, Sharepoint files and time-card records. After competing bids, the sale was put on hold following a flight attendants’ union charge that it involved members’ private information.
Mayu Tobin-Miyaji, general counsel of the Electronic Privacy Information Center, has questioned whether anonymization can fully remove privacy concerns. Emails and Teams messages might contain details about medical appointments or family matters. Those kinds of personal disclosures can be mixed into work records, making the boundary between professional data and private information difficult to draw.
Workers’ consent and the value of their work
Privacy is only one part of the debate. Workplace records are created by employees, but companies may control the records and decide whether to sell them. That creates questions about whether workers know their contributions could be used to train AI, and whether they have a meaningful opportunity to object.
There is also a tension in using human work to improve systems that may automate parts of that work. A flight attendants’ union leader described the data as valuable in the age of AI because it can train models to replace jobs and commodify aspects of working life. The concern is that employees may help build tools that could reduce the need for human labor, without directly choosing to participate.
Gordon Ritter, a venture capitalist focused on the issue, says norms and standards for data acquisition have not been established as quickly as the AI industry is moving. Ansari expects companies to sell workflow data more commonly as a revenue source. Whether that becomes routine may depend on how firms address privacy, worker consent and the expectations people have for the records they create at work.