Sell your company's data to AI labs, the safe way

Why AI labs want company data now

The public internet is running out as a source of new training text. Epoch AI estimates the stock of public human-written text at about 300 trillion tokens and expects it to be fully used between 2026 and 2032. Its estimate leaves out private data, such as company messages, because that data is fragmented and legally harder to reach.

That private data is exactly what labs now seek. OpenAI's data partnership programme asks for data that is not already easily accessible online, and especially data that "expresses human intention", such as long conversations rather than disconnected snippets. In April 2026, TechCrunch reported that Meta would record some employees' keystrokes and clicks to train its models, because they need real examples of how people actually use computers.

A market has formed. Forbes reported that one wind-down firm had handled about 100 sales of shut-down startups' Slack, email and Jira archives in a year, at $10,000 to $100,000 each. In 2026, Google bid $10 million for Spirit Airlines' internal data in its bankruptcy, according to a law firm briefing that also set out the objections it drew.

What labs pay for: decisions, not output

Finished work is common, because the internet is full of it. What is scarce is the trail of reasoning around the work, which only exists inside companies. OpenAI's research found that feedback on each step of a solution trains a model better than feedback on the final answer alone, and builders of training environments for labs stress real workflows, with their exceptions and escalations.

That is why a Slack thread where your team argued over three fixes, rejected two and explained why, then linked to the ticket and the change that shipped, can be worth more than the code itself.

What makes your data worth more, and what rules it out

Raises value

• Depth: years of unbroken history on the same tools. A switch of tools three years ago means three years of usable history, however old the company is.
• Linked systems: one piece of work you can follow from email to chat to ticket to change to result.
• Decisions in writing: remote, written teams keep their reasoning. If the work happens on calls, the record only shows that a ticket was opened and closed.
• Outcomes: whether the fix worked or the customer stayed.
• Rarity: fields where public data is thin. What buyers want shifts over time.

Rules data out

• Data you hold for clients under their contracts, unless each client agrees.
• Privileged legal advice, which generally needs each client's informed consent.
• Health records you cannot de-identify to HIPAA's standard, and anything about children.
• Promises in your privacy notice, customer terms or handbook that rule out new uses.

Who buys company data

• AI labs, which train general models and take data through partnership programmes and licensing deals.
• Training environment builders, which turn real workflows into practice worlds where AI agents learn to do work. The shut-down startup archives in the Forbes report went to this kind of use.
• Data marketplaces and brokers, which package data from many sources for buyers.

You don't need to approach a lab directly. Most deals run through intermediaries who know what each buyer is looking for at the moment.

Why you shouldn't de-identify it yourself

Removing names is the easy part. The hard part is everything else in free text: signatures, quoted replies, nicknames, project codenames, customer names and numbers. Even widely used detection tools say plainly that they can't guarantee to find all sensitive information, and a 2026 study showed that AI can link pseudonymous writing back to real people at scale.

• Keep the relationships. Each person should become the same stand-in, such as "employee 632", across every system, so who-said-what survives. Ad hoc redaction breaks it.
• Keep the reasoning. In-house clean-ups often delete the threads that look messy, which are the ones buyers value most.
• Keep it independent. A specialist that processes the raw files, deletes them afterwards, and documents the method is easier to stand behind than a script your team ran.

The FTC fined Avast $16.5 million in 2024 after it sold browsing data it had described as anonymous. "Anonymous" is a claim you can be held to.

The legal ground in the US, in brief

• De-identified, properly. Under California's privacy law, data counts as de-identified only if you take reasonable measures against re-identification, publicly commit not to re-identify it, and contractually bind every recipient. Virginia, Texas, Oregon and others use the same three steps.
• Employees. Most state laws exclude employees, but California's has covered them since January 2023.
• Your promises. The FTC says a company "cannot then unilaterally renege" on the privacy commitments it collected data under.
• Your tools. Slack's API terms ban using its API data to train large language models, and Google's Workspace API policy bans transfers for general AI models. Check how your data can be exported before you plan anything.
• Sector laws. HIPAA for health records, GLBA for financial institutions, and Illinois' biometric law for voice and face data.

This is general information, not legal advice. Every deal needs its own legal review.

How it works with HerWorkCircle

1. Take the free check above, then a free call. You share record counts and the tools you use, not files.
2. We review what you could license: which systems, how many years, what has to stay out, and what your contracts and tools allow.
3. An independent specialist de-identifies a sample, and our trained data-checking team reviews it by hand for anything software missed. You approve it before anything goes further.
4. We take the prepared dataset to buyers looking for that kind of data, and help you weigh the terms: scope, length, exclusivity and deletion at the end.
5. You sign the licence you choose, the data is delivered, and you keep your business and your records.

Questions

Sources

1. Will we run out of data? Limits of LLM scaling based on human-generated data (June 2024) Epoch AI
2. OpenAI Data Partnerships (November 2023) OpenAI
3. Meta will record employees' keystrokes and use it to train its AI models (21 April 2026) TechCrunch
4. Failed companies are selling old Slack chats and email archives to train AI (17 April 2026) Gizmodo, reporting Forbes
5. Selling company data for AI training: what to know before you say yes (1 October 2026) Frankfurt Kurnit Klein & Selz, law firm
6. Let's Verify Step by Step (May 2023) Lightman et al., OpenAI, on arXiv
7. RL environments for enterprise workflows: a buyer's checklist (14 July 2026) Invisible Technologies
8. California Consumer Privacy Act, section 1798.140(m), deidentified California Privacy Protection Agency
9. AI (and other) companies: quietly changing your terms of service could be unfair or deceptive (13 February 2024) US Federal Trade Commission
10. FTC order on Avast's sale of browsing data (22 February 2024) US Federal Trade Commission
11. Slack API Terms of Service (29 May 2025) Slack
12. Google Workspace API user data policy protections (3 February 2024) Google Workspace
13. Presidio: data protection and de-identification SDK Microsoft (open source)
14. Large-scale online deanonymization with LLMs (February 2026) Lermen et al., on arXiv