Skip to content
Outrich Data

AI labs are buying data from closed startups. What should an operating company take from it?

Published 2026-11-17 · Updated 2026-11-17 · 7 min read · By the Outrich Data team

MIT Technology Review reported on October 13, 2026 that AI labs and data vendors are buying code, chat archives and documents from startups that have shut down. One founder reported about $25,000 for a code base and about $15,000 for workspace data. The reporting also shows why rights and employee privacy have to be settled before any sale.

This page is general information, not legal advice. Whether your company can license a given dataset depends on your contracts, your privacy notices and the law where you and your customers are. Review it with your own counsel before signing anything.

What the reporting says

On October 13, 2026, MIT Technology Review published a report on AI labs buying data from closed startups. Its subject is a trade that barely existed a short time ago: when a startup winds down, its code, email, chat history and documents are sold to AI labs and to vendors that build training material for them.

The article describes wind-down services that have added data sales to the work of closing a company. One of them told the publication that 47% of its deals now include a sale of code or data, and that such sales had gone from rare to routine. These are that company's own figures as reported, not an audited market statistic.

The article is clear about why the buyers want this material. Records of real work show how people handle real tasks, including the disagreements and judgment calls, in a way that invented examples do not. That matches what we describe in what data AI labs buy and why AI agents need real business data.

The figures, and how much weight they carry

The article gives few prices, and each one is a single data point. We repeat them here as reported, with the date, and nothing more.

What was soldFigure as reportedContext in the article
A company code baseAbout $25,000One founder winding down a startup, sale arranged through a wind-down service
Workspace data in Google, Notion, Figma and SlackAbout $15,000The same founder; the sale was still pending and the amount was described as likely
An airline's email, Teams messages and employee records$10 millionBought by Google in the airline's bankruptcy in August 2026, after outbidding an AI data company

Why those numbers are not a price list

A small startup's archive and a large carrier's decades of records sit at opposite ends of a wide range, and the article itself says prices vary widely. It also reports that founders have little to no visibility into who is buying the data and what the buyer plans to do with it, and little assurance that the offer is fair.

That is what a market looks like before it has standard terms. For figures from larger licensing deals, each with its source, see our table of publicly reported AI data licensing deals. For how buyers arrive at a number, see AI training data pricing.

The privacy problem, stated plainly

A company's chat and email archive is not only a record of work. It holds what employees said to each other when they assumed a small audience. An attorney quoted in the article called this kind of data legally radioactive and pointed to the private details that sit in emails and calendars.

The article reports that in the airline sale, the flight attendants' union objected on privacy grounds. It quotes a law professor at Emory University saying that employers should have express consent before selling workers' data, while adding that no federal law bans such sales. It also reports that suppliers employ data scientists to scrub personally identifiable information and exclude HR records, and that buyers say they scrub too.

Three separate groups of people are inside a typical workspace archive, and each raises its own question. The points below are general information, not legal advice, and the answers differ by jurisdiction and by contract.

  • Employees. Direct messages, private channels and personal remarks in public channels. Workplace privacy law varies by state and country, and what staff were told about monitoring and reuse matters. See employee data and AI licensing.
  • Customers and their end users. Names, contact details, account numbers and the content of support conversations. Customer contracts and the company's own privacy notice often limit what can be shared, and privacy laws such as the CCPA and GDPR may apply.
  • Clients whose data the company only held. Records processed on a client's behalf are generally the client's to license, not the vendor's. See client data vs company data.

Who has the right to license a closed company's records

This is the hard question in a shutdown, and it is a legal one. In general terms, the right to deal with a company's assets during a wind-down sits with whoever controls the company at that point. In a solvent dissolution that is usually the board and officers. In an insolvency process it may be a trustee, an assignee or another court-supervised party. Counsel handling the wind-down should confirm which applies.

Control of the files is not the same as the right to license them. The contracts the company signed while it operated do not disappear because it closed. A confidentiality clause in a customer agreement, a privacy notice that promised a limited use, or an earlier exclusive license can each restrict a sale, whoever is signing.

The article also notes where the money goes: proceeds typically flow to senior creditors and investors, not to the employees whose messages are in the archive. That is a fairness question as much as a legal one, and it is a reason to treat consent and notice seriously. Our guides on closing company data and the startup shutdown data checklist go through the steps in order.

What a company that is still operating should take from it

The reporting is about companies that have closed, but the lesson for a company that is still running is more useful. Everything that makes a shutdown sale risky is easier to handle while the company is alive.

  • The people who know the contracts are still there. Someone can say which customer agreements restrict reuse and which do not, before a data title search rather than after.
  • Employees can be told. A company can give notice and explain what is in scope and what is excluded. A starting point is the employee notice in our templates.
  • Scope can be chosen instead of dumped. An operating company can license public channels and leave direct messages out entirely. Outrich Data does not buy direct messages, as the Slack page explains.
  • There is time to de-identify properly. Removing personal and client details is slower than a wind-down timetable usually allows. See how to anonymize company data.
  • The company knows who the buyer is. With a direct buyer, the counterparty is the party that signs the license. See direct buyer vs data broker vs marketplace.
  • A license leaves ownership in place. An operating company can license a copy, keep the data and license it again. See data licensing vs selling data.

A short checklist before anyone exports a workspace

Whether the company is closing or thriving, the same preparation applies. None of it requires a buyer to be involved yet.

  • List the systems and what each holds, using the data inventory worksheet in templates.
  • Pull the customer contracts, the privacy notice and the employee handbook, and read what each says about reuse. The data rights pre-check builds a tailored list.
  • Decide exclusions first: direct messages, private channels, HR and legal channels, and anything held for clients.
  • Convert a small sample with the Slack export parser, which pseudonymizes users in the browser, and review what is left in the text.
  • Run a secrets scan on any code before it leaves the building. Shared credentials should be rotated regardless of the result.
  • Have counsel confirm who has authority to sign and what notice or consent is needed.

What we do not know

The article is one report. It does not publish contracts, and most of its figures come from the companies involved. It says prices vary widely and that no company has published tests showing how much real-world data improves a model compared with the alternatives. We have not seen the underlying deals and do not describe them beyond what the article states.

What can be said with confidence is narrower: buyers exist for internal work records, the public figures so far are few, and the privacy and rights questions are not solved by a company closing its doors.

Common questions

How much do AI labs pay for a closed startup's data?
Public figures are few. MIT Technology Review reported on October 13, 2026 that one founder received about $25,000 for a code base and expected about $15,000 for workspace data. The same article reported a $10 million sale of an airline's records in bankruptcy. These are single data points as reported, not a price list.
Is it legal to sell a closed company's Slack archive?
It depends on the company's contracts, its privacy notices, the laws where its employees and customers live, and who has authority over the assets in the wind-down. There is no single answer. This is general information, not legal advice, and counsel handling the closure should review it.
Do employees have to consent before their messages are licensed?
The rules vary by jurisdiction. A law professor quoted in the article argues employers should obtain express consent, and noted that no US federal law bans such sales. Whatever the legal minimum, excluding direct messages, removing personal details and giving notice lower the risk. Ask counsel what applies to your workforce.
Does Outrich Data buy direct messages?
No. Outrich Data does not buy direct messages. People treat them as private even on a company workspace, and the consent picture is rarely clean. Public channels are where the licensable value sits.
Is it better to license data while the company is still operating?
It is usually easier. The people who know the contracts are available, employees can be given notice, exclusions can be chosen with care, and there is time to remove personal and client details before anything is delivered.

Find out if your data qualifies.

Eleven questions, about five minutes, no call required. You get an answer on screen and a written follow-up.

esc