Is It Legal to Sell Your Company's Data to AI Companies?

Is it legal to sell your company data to AI companies?

Selling or licensing your company data to AI companies is legal in the United States for records your company owns and is free to license, and no federal statute prohibits it. What makes a particular deal unlawful is almost never the AI use. It is that some of the material was never yours to license, or that your own published promises said you would not do this.

Frankfurt Kurnit Klein & Selz set out why in an October 1, 2026 article on selling company data for AI training. The firm's starting point is ownership: email and chat routinely contain other companies' confidential information, covered by NDAs and commercial agreements, and licensed material, vendor code and third-party copyrighted content carry restrictions or competing ownership claims.

MIT Technology Review's October 13, 2026 report makes the other half of the point. Ifeoma Ajunwa, a law professor at Emory, told it that employers should have express consent before selling workers' data, and the publication noted there is no federal law banning such sales. She described the proceeds of these sales as "captured capital", because they go to senior creditors and investors rather than to the workers whose records were sold. An attorney advising AI startups, Greg Ewing, called this kind of data "legally radioactive" in the same piece.

So the honest answer has two parts. Licensing is lawful. Licensing everything is not, and the gap between the two is work you do before you sign rather than a risk you accept afterwards.

Do you have the right to license the data you hold?

Holding data and having the right to license it are different things, and this is where most of these deals shrink. Go through four categories before you talk to a buyer: material held under customer or partner contracts, data you hold as a processor for somebody else, licensed and vendor-supplied content, and the platform terms of the systems the records live in.

Customer and client material is the biggest one. Frankfurt Kurnit's October 2026 article advises reviewing customer, partner, vendor and license agreements, and data held as a processor, before anything moves. Polyshares, one of the buyers in this market, states the same limit on its own site: the license "does not cover what you hold on behalf of your own clients".

Platform terms are the one owners never think of. Frankfurt Kurnit flags that Slack's API terms prohibit third-party application providers from using API data to train a large language model, which matters because the buyer's extraction tooling may be exactly such a provider. Your right to your own messages is not the same as a vendor's right to pull them through an API for training.

Two further risks in that article deserve an owner's attention because they are not obvious. A sale may waive privilege over legal communications, and it may jeopardize trade secret protection for pricing logic, strategy documents and source code, since the protection for a trade secret depends on the steps you take to keep it secret. Both are permanent in a way a bad price is not.

Does your privacy policy allow you to license company data?

Your own privacy notices bind you, and most of them were not written with this in mind. Frankfurt Kurnit's position in its October 2026 article is that companies must follow the privacy promises in place when the data was collected, and that most privacy policies, employee notices and handbooks do not contemplate selling workplace communications for AI training.

Quietly rewriting the policy to fix that does not work retroactively. The same article cites the Federal Trade Commission's treatment of retroactive changes to privacy promises as potentially deceptive, and its caution that quietly adopting more permissive practices through a policy amendment may cross into unfair or deceptive conduct. For data you already hold, prior commitments generally bind unless you obtain consent. For data you collect in future, notices can be updated prospectively, with opt-in consent where it is appropriate.

That asymmetry is worth sitting with, because it changes the shape of a deal. The archive is the valuable part, and the archive is the part your old notices govern. A company that updates its notices today has improved its position for next year's records, not for the ten years of history a buyer wants.

So read what you published, in the versions that were live across the period the records cover. Most companies have never done this. It is also the cheapest piece of diligence in the whole process, and the same gaps tend to surface in later acquisition diligence anyway, so the work is not wasted if no license happens.

Which privacy laws apply when you license company data?

Privacy law applies to the personal information inside your records, and in the United States the California Consumer Privacy Act is the one that reaches furthest into business records. On Frankfurt Kurnit's reading in its October 2026 article, it covers personal information about California workers, job applicants, contractors and business contacts, which is to say the people inside your email, your CRM and your HR systems.

The mechanism matters. A paid transfer of workplace data generally counts as a sale of personal information unless the data qualifies as de-identified under California Civil Code section 1798.140(m), a standard the firm says requires technical measures, a public commitment not to re-identify, and contractual duties on recipients. Treating a transfer as a sale brings notice, opt-out and risk assessment duties under California Code of Regulations title 11, section 7150.

Other regimes attach to particular content rather than to the deal. The same article lists HIPAA for health information, the Gramm-Leach-Bliley Act for financial data, children's privacy laws, biometric laws and call recording consent laws, and mentions the GDPR and other international laws.

Buyers do handle some of this. Scale AI states on its data partnership page that it anonymizes everything inside your environment before any data is extracted, and Mercor states on its company data page that no AI lab buyer or third-party vendor ever sees your raw data. Both were read on their own program pages on November 3, 2026. Neither claim decides your obligations, because the duty under these rules sits with you.

Is de-identification enough to make a data license legal?

De-identification is not the test, and treating it as the test is the most common mistake in this market. Frankfurt Kurnit's sentence in its October 2026 article is the one to keep: "De-identification also addresses whether a record can be linked to an individual, not whether its content is confidential."

Read that against a real archive. A client's pricing strategy in an email thread is confidential whether or not anyone's name is attached. A privileged memo is privileged with the names stripped out. Your supplier's cost structure, written into a quote, is still your supplier's. Removing identifiers does nothing about any of it.

The technical side is harder than it sounds too. The same article notes that removing identifiers from millions of emails and chats is far harder than from structured fields, and that buyers want records kept linked to one another, which makes it possible to draw conclusions about individuals or small groups. Linked records are what make an archive useful for training, and they are also what makes re-identification plausible.

Hence the firm's blunt recommendation, which is the core of the argument here: "Exclude high-risk data entirely rather than relying on de-identification." Its exclusion list names legal-privileged material, personnel and HR files, payroll and tax records, health and benefits information, union-related communications, confidential information of customers and partners, licensed data and data from regulated systems. It also says not to share sample data before your own review is done, because a sample itself could be a disclosure.

An exclusion list is also easier to defend later than a pipeline. It is a decision you made, in writing, before anything moved.

Do you have to tell your employees before licensing company data?

Telling employees is a legal question in the United States, not simply good manners, because some of what sits in your systems is protected activity. Frankfurt Kurnit's October 2026 article notes that employee communications about pay and working conditions may be protected concerted activity under section 7 of the National Labor Relations Act, whether or not your employees are unionized, and that a collective bargaining agreement may also restrict a sale.

Unions do object in practice. Spirit Airlines' flight attendants' union objected on privacy grounds to the sale of the airline's records, as MIT Technology Review reported on October 13, 2026.

Consent is where the commentary is ahead of the law. Ifeoma Ajunwa of Emory told that publication that "Employers should have express consent before selling workers' data", while the report noted no federal law requires it. That is an argument about what employers should do, and it is a reasonable description of what a workforce will expect once it finds out.

The sequence that avoids most of this is simple. Decide the exclusions first, so HR files, payroll and union communications are out before anyone asks. Then tell the team what is being licensed and what is not, before data moves rather than after. Buyers exclude HR records as a matter of course, and Labelbox's Prism program states on its FAQ that human resources records, employee reviews, individual staff identifiers, internal pay data and benefits details are excluded by default. Employees are still in the ordinary records, with their names removed and their judgment calls intact.

What should you exclude from an AI data license?

Exclude anything that is not yours to give, then every category that carries a statute of its own. Exclusions are the part of the deal you control completely, and the list is short enough to write in an afternoon. Start with anything that is not yours: client and customer material held under contract or NDA, partner data, licensed content and vendor-supplied records.

Then the categories that carry their own statutes. Frankfurt Kurnit's October 2026 list is legal-privileged material, personnel and HR files, payroll and tax records, health and benefits information, union-related communications, confidential information of customers and partners, licensed data and data from regulated systems. Recorded calls deserve their own line, because consent rules differ by state and a recording made lawfully in one place may not be lawfully transferable from another.

Buyers expect this and build for it. Scale AI publishes on its partnership page what it will not take: employee and customer identities, raw systems, credentials or source access, and anything outside the signed agreement. A written exclusion list makes your first scoping call faster rather than slower, because the buyer is trying to work out the same boundary.

The part worth deciding deliberately is the grey middle: internal email and chat. It is often the most valuable material a buyer wants, and it is where privilege, employee protections and third-party confidentiality all live at once. You can license documents and tickets while holding email back. That trades some payout for a much smaller review, and for many firms it is the right trade on the first deal.

What happens if the company is closing or in bankruptcy?

A closing or bankrupt company can sell its data, and the process is supervised in a way an operating company's is not. Under section 363(b)(1) of the Bankruptcy Code, if the debtor disclosed a policy prohibiting the transfer of personally identifiable information to unaffiliated parties and that policy was in effect when the case began, the trustee may not sell that information unless the sale is consistent with the policy, or a consumer privacy ombudsman is appointed under section 332 and the court approves the sale after notice and a hearing.

The Spirit Airlines sale shows that machinery working. MIT Technology Review reported on October 13, 2026 that Google paid $10 million for the airline's internal records, a set including 100 million emails, 500 million Teams messages and 176,000 employee records, after outbidding Mercor. The Next Web reported on October 9, 2026 that micro1 also bid and that the court-appointed privacy official backed Google's offer.

Smaller shutdowns go through intermediaries rather than courts. MIT Technology Review's October 2026 report found that SimpleClosure's Asset Hub has handled more than 150 deals, expects to return more than $10 million to startups by the end of 2026, and that 47% of its deals now include code or data sales. One founder in that piece, Keelan Miskell, was paid about $25,000 for his code base and expected around $15,000 for workspace data.

Two things follow for an operating company. The first is that your own vendors may sell your data this way: Frankfurt Kurnit's October 2026 article advises checking whether a vendor contract permits a sale, license or transfer on acquisition or insolvency, and notes that "Contract changes are far easier to negotiate before a vendor receives an offer of its own." The second is that a sale in distress is a worse sale. The owner has no bargaining power, the scope is whatever the estate holds, and the proceeds go to creditors.

Frequently asked questions

Related resources

• Can I sell my company's data to AI companies?
• What happens to my data after I license it to an AI company?
• Which AI company pays the most for business data?
• How BLACKSIG helps companies license their data