MIT Technology Review reveals how failed startups sell workplace data for AI training

A failed startup’s emails, chats and code can now earn money after the business closes. In an October 13 investigation, MIT Technology Review’s Matthew King details a resale market serving AI buyers. SimpleClosure says its marketplace has handled more than 150 deals, while buyers and legal experts describe unsettled questions about training value, privacy and consent.

A shutdown can leave a saleable archive

King’s reporting follows Keelan Miskell, who began winding down his crypto startup this year with little expectation that its records had much value. He received roughly $25,000 for the code base. A separate payment for workspace data from Google, Notion, Figma and Slack remained pending, likely around $15,000.

SimpleClosure brokered the sale. The company added data sales to its shutdown services in April through Asset Hub, a platform for selling code, emails, chats and documents. Its buyers include foundation-model labs, vendors building simulated workplaces for AI training, and established technology companies developing their own models.

CEO Dori Yona told King that code or data sales now appear in 47% of SimpleClosure’s deals. The company expects Asset Hub to return more than $10 million to startups by the end of 2026. That is a forecast, not money already returned, and the deal share describes SimpleClosure’s business rather than the entire resale market.

More records do not always mean more value

The interviews reveal different ways to price an archive. Sunset, another startup shutdown business, considers head count, years in operation, data volume and how a team used its tools, according to founder Brendan Mahony. These factors describe both the size of the archive and the working history behind it.

Surge AI’s Nick Heiner puts particular weight on the situations those records capture. A detailed Slack history containing varied conversations can be worth more than another company’s Slack, email and calendars combined, he said. He seeks archives with disagreements and difficult judgment calls, not simply a large pile of messages.

Surge buys data to build reinforcement-learning environments: practice settings where clients’ AI agents attempt tasks such as debugging complex IT systems or drafting appellate briefs. The archive supplies material for the simulated workplace. Real records can avoid errors found in synthetic worlds, including references to dates that do not exist.

That does not establish better model performance. King writes that no company has released tests comparing training on real company data with alternatives. Heiner says buyers quote widely different prices because valuations rest on research hypotheses about how many useful training tasks an archive can yield—for example, from 100,000 lines of code.

Employers should have express consent before selling workers’ data

Ifeoma Ajunwa, Emory University law professor, speaking to MIT Technology Review

Removing names does not settle consent

Suppliers employ data scientists to remove personally identifiable information and exclude HR records; buyers say they also scrub the data. But attorney Greg Ewing, who advises AI startups, warns that ordinary emails and calendars can contain private details: medical appointments, school pickup arrangements and date-night plans.

Ajunwa draws a separate distinction between collecting workplace data and selling it after shutdown. Legal precedent permits monitoring to help employees do their jobs, she says, but that justification disappears when the employer closes. She argues for express worker consent, while noting that no federal law bans these sales.

The financial benefit also does not typically reach the people whose work produced the records. Proceeds generally flow to senior creditors and investors rather than employees, King writes. Founders face their own information gap: they have little visibility into the buyers or intended uses, and little assurance that the offered price is fair.