← All articles

Mikhail Polyakov

Shut-down startups are selling their correspondence to train AI: how this market works and what to decide in advance if your company has work chats

Level: BeginnerReading time: 8 min

Topics: AI, Security, Small business

An empty office with open cardboard boxes, from which glowing threads and message-like cards stretch towards a window

The short answer: according to an MIT Technology Review article of 13 October 2026, a market has formed in the US where startups that are shutting down sell their archives to companies that train AI: source code, email, documents, conversations in work chats. One founder received about $25,000 for his code and has not yet been paid for the workspace data, likely around $15,000 by the article's estimate. The broker SimpleClosure has handled more than 150 such deals, and according to its CEO, 47% of its deals now include a sale of code or data. A lawyer quoted by the publication calls such data "legally radioactive": the correspondence contains personal details of employees. The proceeds, according to the article, typically go to creditors and investors, not to employees. Below: what is known, what no one in the market has shown yet, and what is worth deciding in advance if your company has work chats and email.

What happened

Journalist Matthew King describes the case of Keelan Miskell. In 2026 he was winding down his crypto startup, which was only a few years old, and did not think its data was worth anything: the users were anonymous. But he received about $25,000 for the source code, and the amount for the workspace data from Google, Notion, Figma and Slack has not been paid yet; by the article's estimate, it is likely around $15,000. His words in the article: everything the team did over two and a half years is worth something to someone.

The sale was brokered by SimpleClosure, a company that helps startups shut down. It added data sales to its services in April 2026. Its Asset Hub platform helps founders sell code and workspace data: emails, chats, documents. According to MIT Technology Review, it has handled more than 150 deals, and the company expects to return more than $10 million to startups by the end of 2026. SimpleClosure CEO Dori Yona says such deals were once rare and considered "somewhat taboo," and have now become routine.

A competitor, Sunset, does something similar. Its head, Brendan Mahony, explains that the value of a closing company's data is assessed by head count, years in operation, volume of data, and how the team used its work tools. He compares such data to the wood shavings used to make particle board: they were considered waste, but they have a price.

Who buys and why

According to the article, there are three kinds of buyers: large labs that build foundation AI models; vendors that build simulated workspaces for training AI; and mature tech companies building their own models.

Demand, according to the publication, emerged suddenly in 2025, when labs hit the limits of public and synthetic data. An example of a buyer: Surge AI builds training environments from such archives, where clients' AI agents practise tasks such as debugging a complex IT system or drafting an appellate brief. Nick Heiner, who leads this work at Surge AI, says he looks for archives with a range of situations: disagreements, tricky judgment calls. In his words, one company's Slack history, if it has varied and substantive conversations, can be worth more than another company's Slack, email and calendars combined.

The largest single deal described in the article involves Spirit Airlines. In August 2026, as part of the airline's bankruptcy, Google, according to the article, outbid the company Mercor and bought the airline's data for $10 million. The data includes 100 million emails, 500 million Teams messages and 176,000 employee records. MIT Technology Review also cites Bloomberg: in September 2026, SpaceX's AI unit was weighing buying data from struggling startups.

What has not been shown yet

The main thing: how much such data improves AI. The article states plainly that no company has yet published tests comparing training on real workplace data with the alternatives. There is an argument in favour of real data: real archives do not have the errors that occur in synthetic ones, such as references to dates that do not exist.

Hence the spread in prices. Heiner says buyers quote very different amounts, because each one estimates based on its own assumption, for example how many useful training tasks can be made from 100,000 lines of code. The author of the article expects prices to stabilise in the coming year, unless an AI bubble pops first. That is the journalist's forecast, not a measurement.

Why this is dangerous for the people in the correspondence

Spirit Airlines' flight attendants' union, according to the article, objected to the sale of the data to Google on privacy grounds. Greg Ewing, an attorney who advises AI startups, calls such data "legally radioactive." Sellers, according to the article, hire data scientists to scrub information that can identify a person and exclude HR records; buyers say they scrub too. The article does not say whether this rule applies to all deals: the Spirit Airlines data, as mentioned above, included 176,000 employee records, and the article does not say in what form they were handed over. Still, Ewing lists what remains in email and calendars: doctor appointments, arrangements for who picks the children up from school, date-night plans. These details do not look like passport data, and a formal scrub based on markers such as phone numbers may miss them. This is my reasoning, not a conclusion of the article.

The second question is who gets the money. According to the article, the proceeds typically go to senior creditors and investors, not to the employees whose correspondence is being sold. Ifeoma Ajunwa, a law professor at Emory University, calls this "captured capital." Her argument: monitoring employees is allowed in order to help them do their jobs, and once the company has closed, that justification disappears. She believes an employer should obtain the employees' express consent before selling their data, but adds that there is no federal law in the US banning such sales.

The third question concerns the founders themselves. According to the article, they have little to no visibility into who is buying the data and what will be done with it, and little assurance that the price is fair.

What a company in Russia can take from this

The article is about the US. What follows is my advice, not conclusions of MIT Technology Review, and it is not legal advice.

The main idea: a company's work chats and email are not only a tool, they are also an archive that may contain employees' personal details and clients' trade secrets. It is easier to decide what can be done with it while the company is running than when it is closing.

  1. Write down what is in the work chats and email. Not item by item, but by type: correspondence with clients, contracts, information about employees, personal conversations.
    • Who does it: the manager together with whoever administers the work tools.
    • How to check: a short list of the types of data and the tools where they are kept, with the date it was compiled.
  2. Check the contracts with clients. If clients gave you their data or documents under a confidentiality clause, that clause may prohibit passing them on to anyone else, including as part of an archive when the company closes. This is my assumption about the usual terms of such contracts; what yours say can only be seen in the contract itself.
    • Who does it: a lawyer or the manager who signed the contracts.
    • How to check: in each major contract, the confidentiality section has been found and what it allows has been written down.
  3. Keep the personal data law in mind. In Russia, the processing of personal data is governed by Federal Law No. 152-FZ on personal data, and transferring such data to third parties also counts as processing. Article 6 of the law lists the grounds on which processing is permitted; the person's consent is one of them. Whether selling work correspondence to a buyer who will train AI on it fits these grounds is for a lawyer to decide, not for the broker offering the deal.
    • Who does it: a lawyer, before any deal involving the data is signed.
    • How to check: a written opinion from the lawyer on the legal basis for transferring employees' and clients' data, or a refusal of the deal.
  4. If someone offers to buy your data, ask who the end buyer is and for what purpose. According to the article, founders have little visibility into who is buying their data. Without an answer, it is hard to assess both the risk and the price.
    • Who does it: the manager.
    • How to check: the buyer's name and the purpose of use, in writing, in the contract.
  5. If a deal does go ahead, tell the employees. Ajunwa considers express consent the right approach. Even where the law does not directly require it, an employee who learns from the news that their correspondence has been sold is a reputational risk for the manager.
    • Who does it: the manager.
    • How to check: a written notice to employees before the deal, and the responses of those who object.

If the company is not planning to close and no one is offering to buy its data, it is worth starting with the first two points: they are needed to know what you have in your chats, email and documents and what your contracts say about it. The other points will be needed when an offer of a deal appears.

I wrote about a similar dispute, from the side of those whose texts have already been used for training, in the article on why USA Today is demanding money from OpenAI.

Summary

According to MIT Technology Review, the workplace archives of startups shutting down in the US have become a commodity for companies that train AI: from tens of thousands of dollars for a small startup to $10 million for the Spirit Airlines data. How much such data improves models, no company has yet shown in published tests. A lawyer quoted by the publication warns that personal details of employees remain in email and calendars, and the proceeds typically go to creditors and investors. For a company that has work chats, it is useful to know already now what is in them and what the contracts say about them, and to show any deal involving such data to a lawyer first.

I work on AI agents and automation. If you would like to see my projects or discuss your own task, take a look at my portfolio.

Sources