How to sell your company's data to AI labs

Short answer: AI labs and the vendors that supply them now pay companies for real business records, such as email and chat archives, documents, code and operational data, because AI agents need examples of real work to learn from. Published deals range from about $25,000 for a small startup's code base to a reported $10 million for an airline's records in bankruptcy. Before any sale, a company has to confirm it has the right to sell, exclude high-risk data, de-identify what remains and license it under strict use limits. The safest structure is a license, not an outright sale, because data used to train a model generally cannot be removed from it later.

Last updated November 12, 2026. Published by Data Revenue.

Key takeaways

• Buyers include foundation AI labs and companies that build simulated workplaces where AI agents practise tasks (MIT Technology Review, October 13, 2026).
• Prices vary widely. MIT Technology Review reported that Google paid $10 million for Spirit Airlines' data in bankruptcy.
• Holding data is not the same as having the right to sell it. Contracts, privacy notices, platform terms and employment law can all limit a sale (Frankfurt Kurnit Klein & Selz, October 1, 2026).
• Some data should be excluded entirely: privileged material, HR and payroll records, health information, union communications and data you only license from others.
• Training is permanent. Price and contract terms should reflect that the buyer cannot give the data back.
• Operating companies, not just shutdowns, can license a de-identified copy of their records and keep running the business.

Why AI labs are buying company data now

The first generation of large language models was trained mostly on public web text. The next generation is being built to do work: answer email, run workflows, write code, reconcile accounts. To learn that, models need examples of how real companies actually operate, and that data is not on the public web.

MIT Technology Review reports that buyers include foundation labs and vendors that build simulated workspaces. One training-data vendor buys company archives to build what the article calls "reinforcement-learning gyms," environments where AI agents practise tasks (MIT Technology Review, October 13, 2026). The same article notes that Bloomberg reported in September 2026 that SpaceX's AI unit was weighing purchases of data from struggling startups.

What published deals have paid

Sources: the first three rows are from MIT Technology Review, October 13, 2026. Reddit: TechCrunch, February 22, 2024. News Corp: Variety via Yahoo, May 22, 2024.

MIT Technology Review also reports that code or data sales now appear in 47 percent of that wind-down service's deals. The lesson from the table is that value depends on volume, years of history, how much real work the records capture, and how cleanly the rights can be shown. Two companies with similar headcount can hold data worth very different amounts.

What data AI buyers want

Based on the deals reported so far, buyers are paying for:

• Workplace communications: email, chat and meeting records that show how people coordinate real work.
• Documents and work products: proposals, tickets, design files, specifications and reports.
• Code and engineering history: repositories, commit history and issue trackers.
• Operational records: transactions, orders, claims, service records, logistics events and other structured data from running the business.
• Domain depth: long histories in a specialist field such as healthcare operations, logistics, insurance, construction or financial services, which general web data cannot replace.

Volume matters, but so do years of history, completeness and the presence of outcomes (what happened after a decision), because those let a model learn cause and effect.

License, do not sell outright

Most companies should license a defined, de-identified copy of their data rather than transfer ownership. A license lets you:

• limit permitted use (for example, model training only, no resale)
• set a term, territory and exclusivity, and charge more for exclusivity
• ban re-identification and onward sharing
• keep the right to license the same data to other buyers later, if the deal is non-exclusive

Price the license knowing it is effectively permanent. Frankfurt Kurnit Klein & Selz notes that data used for training generally cannot be removed from a model, and that buyers could use what they learn to build products that compete with yours (FKKS, October 1, 2026).

The legal checks before any sale

The FKKS guide sets out the main risks. In summary:

1. Rights to the data. Company files often contain other companies' confidential information, vendor-owned or licensed material and third-party copyrighted content. Platform terms (for example, a chat tool's API terms) can restrict how exported data is used. A sale can also waive legal privilege or put trade secrets at risk.
2. Privacy commitments. Existing privacy notices, employee notices and handbooks usually do not mention selling data for AI training, and changing them after the fact can be deceptive. Consent may be needed.
3. Privacy law. Under the CCPA, selling personal information triggers notice, opt-out and risk assessment duties unless the data is properly de-identified. HIPAA, GLBA, biometric privacy laws and the GDPR may also apply.
4. The limits of de-identification. Removing names does not make confidential content harmless, and linked records can still reveal information about individuals or small groups.
5. Employment law. Employee messages about pay and working conditions may be protected activity under Section 7 of the National Labor Relations Act, and collective bargaining agreements may restrict a sale.
6. Disclosure. California's AB 2013 and the EU AI Act may require AI developers to disclose where their training data came from, including whether it was bought or licensed. Assume the deal may become public.
7. Vendors holding your data. Your software vendors may hold copies of your data and may be able to sell them if they are acquired, shut down or go bankrupt. Check those contracts now.

Data to exclude entirely: privileged legal material, HR and payroll records, health information, union communications and any data you license from someone else.

This is a summary, not legal advice. Every sale needs review by counsel who knows your contracts and jurisdictions.

How Data Revenue handles an AI data sale

Data Revenue runs the process for established companies, so the management team keeps running the business:

1. See what your data could be worth. The free 15-question Data Snapshot gives an indicative value range in about five minutes.
2. Map the data and the rights. We inventory the systems, record types, volumes and years of history, and review the contracts and notices that decide what can be licensed.
3. You approve the scope. Nothing is offered to any buyer until you sign off on exactly what is included and what is excluded.
4. De-identify and package. The approved data is de-identified to the standard that applies, documented and prepared as a product a buyer can evaluate quickly.
5. We find the best buyer. We take the package to AI labs and data buyers in our network and negotiate price, use limits, exclusivity and security terms.
6. Keep earning. A one-time AI license is often only the start. The same data, managed through a separate data company, can be licensed to other buyers for years, with your company keeping ownership and the majority of the revenue.

Learn how the AI Labs program works

Operating companies, not just shutdowns

Most of the deals reported so far involved companies that had closed or were in bankruptcy, because their records were available in one piece and they had no further use for them. An operating company can do better: it can license a de-identified copy of its records while continuing to run the business, choose its timing instead of selling in a liquidation, and keep the right to license the same data again. For more on long-term programs, see our guide on how to monetize data.

Frequently asked questions

Can I sell my company's data to AI companies if we are still operating?

Yes, if you have the rights to the data and it can be properly de-identified. An operating company usually licenses a defined copy of selected records for a set purpose, rather than selling everything. You keep the original data and keep running the business.

Who buys company data for AI training?

Buyers include foundation AI labs, companies that supply training data to those labs, and vendors that build simulated workplaces where AI agents practise tasks, according to MIT Technology Review's October 2026 reporting.

How much will an AI lab pay for my company's data?

There is no standard price. Reported deals range from about $25,000 for a small startup's code base to a reported $10 million for Spirit Airlines' records. Value depends on volume, years of history, how much real work the data captures, uniqueness and how clean the rights are. An indicative valuation is the best starting point.

Is it legal to sell employee emails and chat messages?

It depends on your notices, contracts, platform terms and the laws that apply, including privacy and labor law. Many sales exclude HR, payroll, health and union-related material entirely and de-identify the rest. Get legal advice before sharing any sample.

Can the data be deleted from the model after a sale?

Generally no. Once data has been used to train a model it cannot practically be removed, which is why the license terms, exclusions and price need to be right before anything is delivered.

Should I sell or license my data?

License in most cases. A license limits how the buyer can use the data, can ban re-identification and resale, and can leave you free to license the same data to other buyers later.

Sources

1. Matthew King, "AI labs are buying data from dead startups to train their models," MIT Technology Review, October 13, 2026. https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models
2. Daniel Goldberg, Frankfurt Kurnit Klein & Selz PC, "Selling Company Data for AI Training: What to Know Before You Say Yes," October 1, 2026. https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes
3. TechCrunch, "Reddit says it's made $203M so far licensing its data," February 22, 2024. https://techcrunch.com/2024/02/22/reddit-says-its-made-203m-so-far-licensing-its-data
4. Variety (republished by Yahoo), "News Corp Inks OpenAI Licensing Deal Potentially Worth More Than $250 Million," May 22, 2024. https://malaysia.news.yahoo.com/news-corp-inks-openai-licensing-211417483.html