AI Brokers Are Turning Workplace Files Into a New Data Market

AI data brokers are trying to buy companies’ internal work files, emails and messages so they can train models on private workplace records, and the deals can reach millions of dollars, WIRED reports. The push marks a new frontier for AI development as public web data becomes harder to use and corporate “knowledge work” turns into a commodity.

AI’s new appetite is private workplace data

For years, the easiest source of training material for large language models was the public internet. That source is no longer enough on its own, according to WIRED’s reporting, and AI labs are now looking deeper into the ordinary digital exhaust of businesses: documents, email threads, Slack messages, support tickets and call transcripts.

The market is already starting to price that data like a scarce asset. WIRED says some pitches to media outlets offered one-time licenses worth between $350,000 and $1.8 million for old work records, with one proposal suggesting the price could rise to $2 million if call transcripts were included. Another pitch claimed AI labs spend $45 billion on private data.

One freelancer was reportedly offered $1,000 for a single story, underscoring how uneven the market can be depending on the source, the volume and the sensitivity of the records involved.

Who is buying and selling the data?

According to WIRED, a small group of startups is emerging as the center of this market: Mercor, Micro1 and Handshake. They are described as leading data-acquisition companies working between AI developers and the organizations that control large archives of workplace information.

Handshake told WIRED that it licenses professional documents so labs can “test and train” models on real knowledge work. The company also buys graduate students’ papers and coursework, extending the same logic beyond corporate records into academic material.

Mercor, founded in 2023, is another major player in the space. WIRED reports that its three cofounders became billionaires at 22 last year. The company says it strips personally identifiable information from company data before brokering deals, an important claim in a market that is growing faster than the rules around it.

Micro1 is pushing the same broader thesis with even bigger numbers. Its chief executive, Ali Ansari, says current annual spending by labs, enterprises and robotics firms is already about $40 billion to $50 billion. He predicts that figure could rise to more than $1 trillion a year within two to four years.

“Big AI has two bottomless needs: compute and data.”

That was how Mercor CEO Brendan Foody framed the business in WIRED’s report, placing data alongside computing power as one of the core inputs of modern AI.

Why it matters

This is not just a new procurement category. It is a shift in what counts as valuable corporate property.

For AI firms, workplace records may be more useful than public text because they reflect how people actually write, decide, negotiate and solve problems inside organizations. Internal files can contain patterns that never appear in the open web: policy debates, customer service exchanges, product planning, management decisions and the back-and-forth of day-to-day operations. If labs can buy or license that material at scale, the quality of training data may improve even as the privacy risks intensify.

For businesses, the market creates a new line item: the possibility that email archives, chat logs and internal documents are no longer just records to be stored or deleted, but assets with a resale value. That creates incentives that may collide with employee expectations, labor rules and confidentiality obligations.

For workers, the stakes are more personal. Messages written for colleagues, managers or customers can contain sensitive details, opinions and context that were never intended for model training. Even when data is stripped of obvious identifiers, the practical risk of re-identification or accidental leakage may remain.

For regulators and courts, the question is whether existing privacy standards are enough for a market built on unstructured workplace communications. The answer, based on the reporting so far, is not clear.

How private data became the next frontier

The backdrop is a familiar one in AI: public web data has been heavily mined for training, and companies are now searching for new sources. That search has already pushed the industry toward licensed material and purchased datasets. What is different in WIRED’s reporting is the type of data being sought. Instead of polished public documents, the market is moving toward raw internal exhaust.

That shift matters because internal workplace material is often messy, conversational and rich in context. It can include drafts, unresolved threads, human mistakes and private references. In other words, it may be precisely the kind of data a model can learn from—but also the kind that raises the most concern when it leaves a company’s control.

Bankruptcy sales highlight the issue in its starkest form. When a company goes under, its digital records can become saleable assets, even if those records contain highly sensitive material. WIRED points to Spirit Airlines as a case in which a bankruptcy asset sale included 80,000 email accounts, 500 million internal messages, 20.6 million Microsoft SharePoint files and 1 million time-card records.

Spirit Airlines shows how valuable internal records can be

Spirit’s bankruptcy process became a concrete example of how far this market can go. WIRED reports that Micro1 bid $12.5 million for Spirit’s digital assets after a reported $10 million bid from Google. The sale is on hold after the flight attendants’ union objected over private information.

That dispute illustrates a central tension in the emerging market: the same records that are valuable for AI development can also be deeply personal, legally sensitive and difficult to sanitize. The scale alone is striking. A trove of 500 million internal messages is not just a dataset; it is the operational history of a company and its workers.

Micro1’s bid suggests the competition for such archives is not theoretical. There is now an active bidding environment in which major technology companies and specialized startups can see the same corporate records as training fuel.

Can anonymization solve the problem?

Not necessarily. The Electronic Privacy Information Center, known as EPIC, says anonymization may not fully eliminate privacy risks from unstructured workplace messages.

That caution goes to the heart of the business model. De-identification can help, but it is not always enough when data is conversational, context-heavy and full of references to people, projects and events. A message thread may not contain a name, yet still reveal enough about a person, a deal or an internal dispute to make the privacy risk meaningful.

Mercor says it strips personally identifiable information before brokering company data. That may reduce some exposure, but the reporting suggests the broader question remains unsettled: what level of cleaning is sufficient, and who decides?

Gordon Ritter’s view, as reported by WIRED, is blunt: data-acquisition norms and standards have not been set. That leaves buyers, sellers and workers operating in a gray zone where the commercial value is obvious but the market rules are still being written.

What the startups are saying

The companies in this market are making a straightforward economic argument: the demand is real, the supply is valuable and the pricing could get much larger.

Ali Ansari, the Micro1 chief executive, told WIRED:

“In the next two to four years, this is going to be over a trillion dollars a year”

He also said:

“I think it’s going to become more common than ads as a revenue line item for companies”

Those forecasts are ambitious, but they reveal how quickly some players think the market could mature. If internal documents, support tickets and call transcripts become routine data products, companies may begin to treat information sales the way they currently treat advertising, subscriptions or software licenses.

That would be a major shift in the business model of data-rich organizations. It would also create a new incentive to gather, preserve and package internal communications in ways that can be monetized later.

By the numbers

Background: how we got here

The AI industry’s data problem is easy to describe and hard to solve. Large language models are often built on public internet text, but public data is finite, contested and increasingly exhausted as a training source. That has pushed companies toward licensed content, purchased datasets and more specialized material that reflects real-world expertise.

In that context, workplace records are attractive because they are rich in practical language. Emails, support tickets and internal messages show how people actually communicate under deadline pressure, handle customer problems and make decisions. For model builders, that is useful. For companies and employees, it is also deeply sensitive.

The rise of data-acquisition startups shows how the industry has begun to professionalize this trade. Mercor, Micro1 and Handshake are not merely supplying compute or software; they are building a marketplace around the private output of everyday work.

Handshake’s approach extends into academic labor as well, buying graduate students’ papers and coursework. That broadens the ethical questions beyond corporations to universities and the students whose work may also become training material.

Mercor’s rapid ascent, meanwhile, signals how quickly this corner of AI has become capitalized. Founded in 2023, it has already become one of the names shaping the conversation, with WIRED noting that its three cofounders became billionaires at 22 last year.

What happens next

The immediate question is how far this market will spread and who will feel pressure to sell. The reporting suggests demand is already broad, touching media organizations, bankrupt companies and potentially any business with a large archive of digital communications.

Several developments will be worth watching:

• Whether more companies openly license internal records to AI labs.
• Whether courts or unions succeed in limiting sales of workplace communications in bankruptcy cases.
• Whether privacy protections such as de-identification prove strong enough for unstructured data.
• Whether data becomes a standard revenue stream for ordinary businesses, as some executives predict.

The Spirit Airlines case may be especially important because it combines all the tensions in one place: a valuable archive, worker objections and a high-stakes sale process. If that sale moves forward, it could become a template. If it is blocked or narrowed, it could slow the market’s expansion.

For now, the basic fact is clear. AI companies are no longer only scraping the open web. They are trying to turn the private record of work itself into training fuel, and the price tags are already climbing.