Is It Legal for a US Company to Sell Its Data to AI Companies?

Usually yes: no general federal law bars a US company from licensing its own business records to an AI developer, but several layers of law decide which records can go, to whom, and in what form. The first is the company's own word: the FTC treats breaking a privacy promise as a deceptive practice, and has warned that a company that quietly rewrites its terms to put old data to a new use risks acting unfairly or deceptively. Then come the state privacy laws, where licensing personal information for money is a sale under California's law; the sector laws for health, finance, schools and biometrics; the federal and state rules on employees' messages; the contracts with customers and software vendors; and, when the buyer is foreign, the Justice Department's rule on bulk sensitive data. What passes the privacy layers is the company's own working records, de-identified to the standard the law sets, subject to the contract and buyer checks below.

This guide is for founders, CEOs, CFOs and counsel of a US company, operating or closing, that has been asked about its data or wants to know whether it can sell it. It sets out each layer in the order a company meets it, what it allows and what it requires, and links to our deeper guides on customer data, employees' email and chat and health data. The same question for companies abroad is answered for the United Kingdom and Canada.

AIData.Broker values a company's data free and introduces it to every buyer whose published rules take it, free for the company, because the buyer pays us its published referral payout out of its own money; where a sale run by us earns more, and only for holders based in the United States, the United Kingdom and Canada, we act as the company's agent under a mandate and it pays our fee only if the sale closes. Either way the company itself licenses its data to the buyer, and we do not buy it or license it in our own name. This guide is research from public sources, not legal advice.

Contents

1. What the company can sell
2. Layer one: the company's promises and the FTC Act
3. Layer two: state privacy laws
4. Layer three: sector laws
5. Layer four: employees and their messages
6. Layer five: customers' data and vendors' terms
7. De-identification: the main lawful path
8. Data broker registration
9. Selling to a foreign buyer: the DOJ rule and export controls
10. What the company owns: copyright and trade secrets
11. A US company that is closing or has closed
12. Worked example: a marketing agency
13. Questions people ask
14. What this means for you

What the company can sell

A company owns its working records: the email, chat and documents its staff wrote in the course of their work, its tickets, project files and code. Those are what buyers of company records pay for, because they show how the work was done, and offers to holders for a whole company's record have a median of $300K per company. No general federal statute stops a company licensing its own working records; what the law regulates is what is inside them, and three kinds of content need sorting before anything leaves.

• Information about people. Customers, contacts and staff appear throughout a company's records. The privacy promises, state laws and sector laws below decide whether that information can go identifiable, must be de-identified or must stay out.
• Material that belongs to someone else. A business customer's data held under contract, a client's work product, material received under an NDA and data licensed from a supplier are not the company's to license without that party's agreement.
• Material that is restricted whoever owns it. Health records, financial customers' information, students' records, children's data and biometric identifiers carry rules of their own, and some cannot be sold at all.

The rest, the company's own operating history with the people taken out, is what most sales consist of. Our guide to what AI buyers check before buying data lists what a buyer asks the company to show for each kind.

Layer one: the company's promises and the FTC Act

The first law a company meets is its own privacy policy and terms, enforced by the FTC, the federal consumer protection agency. Section 5 of the FTC Act makes unfair or deceptive practices unlawful, and a company that uses data in a way its privacy policy did not disclose is the commonest deception case. A practice is also unfair when it causes or is likely to cause substantial harm that people cannot reasonably avoid and that benefits do not outweigh, which a sale of data people never expected can be.

The FTC has addressed AI training directly. Its technology office wrote in 2024 that a company that quietly rewrites its privacy policy or terms to start using data it already holds for AI training, or to share it with others, risks acting unfairly or deceptively. So data collected under a policy that promised not to share it is best treated as still bound by that promise unless people were clearly told of the change and agreed (our reading). The remedy the FTC uses frightens buyers most: it has ordered companies to delete not only the data they misused but the models and algorithms trained on it, and in the Rite Aid order the deletion reached third parties that received the data. That is why a buyer asks which policy applied when each record was collected.

For a company, this layer has a simple test: read the privacy policy, and the terms of service, that applied when the records were created. A policy that said personal information would not be sold or shared with third parties means that information goes only de-identified. A business-to-business company whose records hold few consumers' details meets this layer lightly; a consumer app meets it on every record.

Layer two: state privacy laws

The United States has no general federal privacy law, so the comprehensive privacy laws are state laws, and they bind only the businesses they cover, for the residents of their state. California's is the model. It binds a for-profit company doing business in California that meets any one of its thresholds, and selling personal information can itself bring a smaller company inside it. Licensing Californians' personal information to an AI buyer for money or anything else of value is a sale. A transfer where the buyer takes over all or part of the business is one of the definition's exclusions, and a licence of data alone is not.

A sale under the CCPA carries duties. The company must have told people in advance and offered the right to opt out; it may not use information for an additional purpose incompatible with the one disclosed when it was collected without notice, and any use must stay reasonably necessary and proportionate; it must sign an agreement that limits the buyer to specified purposes and binds it to the Act; and sensitive information, such as the contents of messages the company was not the intended recipient of, carries a separate right to limit its use. Employees are covered too: the CCPA's exemption for staff information expired on 1 January 2023, so a covered company owes its California staff the same rights over what is in their work email and chat.

The other states with comprehensive laws follow a model first enacted in Virginia and refined in Colorado, and they differ in what counts as a sale: Colorado counts a transfer for any valuable consideration, as California does. Our guide to selling customer data tables each state's law, what it calls a sale, how it treats sensitive data and whether its de-identified standard has the conditions California's has.

Layer three: sector laws

Some kinds of data are governed by a federal or state law of their own, whoever holds them, and some of them forbid a sale outright.

The health route is the one most often used in practice: a provider's or health company's records can be sold once they are de-identified to HIPAA's standard, either by removing the listed identifiers or by a qualified expert's written judgment that the risk of identifying anyone is very small. Our HIPAA de-identification guide explains both routes. Biometric data tied to Illinois residents is the opposite case: with no consent exception for selling, it is close to unsellable, and voice or video archives that include Illinois residents need that checked first.

Layer four: employees and their messages

A company's email and chat are the records buyers of company data ask for most, and they are written by employees about customers, suppliers and each other. Federal law allows a company to use its own systems' contents more than people expect. The Stored Communications Act's bar on disclosing stored messages applies only to services provided to the public, which a company's own internal email or chat generally is not, and access authorised by the entity providing the service is outside its offence. The Wiretap Act, which governs intercepting messages rather than licensing stored ones, allows interception where one party has given prior consent, such as an employee who signed a monitoring and data-use policy, unless the purpose is criminal or tortious; that consent speaks only for the person who gave it, and some state wiretap laws require every party's consent.

State law adds notice. New York requires an employer that monitors email to give acknowledged written notice on hiring and post it; Connecticut requires prior written notice of the types of electronic monitoring and where on the premises it occurs; Delaware allows monitoring only after a daily electronic notice or a one-time acknowledged notice. These are rules about collecting, not licensing: a monitoring notice is a floor, and it does not by itself tell staff their messages may be licensed for AI training. A company that plans a sale updates its data-use policy so the next sale rests on a disclosed use.

Two more groups need care. Outside correspondents on the same threads, customers and suppliers, consented to nothing, so their names and details come out in de-identification. And privileged material, legal advice and anything under an NDA, comes out whole. Our guide to employees' email and chat covers each step, and the guides to Slack and email archives cover the exports.

Layer five: customers' data and vendors' terms

A business-to-business company holds two kinds of data that look alike and are not. Its own records about its work belong to it. The data its customers put into its product belongs to the customers, under the contract, and a company that processes it as a service provider may use it only for the purposes the rules list, which licensing it to an outside AI buyer is not. A software company that wants to sell what its customers' users created needs each customer's agreement, and most will not give it.

The software that holds a company's records has terms too. The workplace tools, help desks and CRMs a company uses generally leave the data with the customer and let it export, but some reserve rights of their own or restrict what an export may be used for. The rights pages read each system's terms, and our readiness check asks which systems hold the records a company would sell.

De-identification: the main lawful path

Most of the layers above stop applying to information that no longer identifies anyone, and that is why de-identified records are what most company sales consist of. California sets the clearest test: information is deidentified only when it cannot reasonably be linked to a person and the company takes reasonable measures, publicly commits not to re-identify it, and binds every recipient by contract to the same, and records that meet it fall outside the Act's personal information rules, the rules on sale included. Colorado's definition has the same three conditions, and HIPAA has its own standard for health data.

De-identifying a company's email, chat and tickets is harder than de-identifying a table, because names, addresses and account numbers sit inside free text and attachments. The courts have seen how it is done well: in the Spirit Airlines records sale, the court's privacy ombudsman backed the sale once passenger systems were excluded and an independent agent was hired to de-identify and certify each transfer. Our de-identification page explains the methods, and the de-identification tool shows what a scrub finds in a sample.

Data broker registration

A company that sells data about its own customers is usually not a data broker. California's registry covers businesses selling data about people they have no direct relationship with, so a company selling data about its own customers is not a California data broker for that data, though the CCPA's sale rules still apply. Oregon excludes a business collecting data about its own customers, subscribers or users, and Connecticut's 2026 law exempts a business collecting data from people in a contractual or similar relationship with it.

New Jersey is the exception. Its 2026 law reaches a company that sells or licenses its own customers' personal data to a data broker, which must register with the Division of Consumer Affairs. A company whose records go to a buyer that resells data, rather than training its own models, and that holds New Jersey consumers' personal information should check this before it signs. Records de-identified to the privacy laws' standard are likely outside these registries too, which register dealings in personal data (our reading).

Selling to a foreign buyer: the DOJ rule and export controls

The Justice Department's data security rule bites only on access by a country of concern or a covered person to government-related data or bulk US sensitive personal data, so a licence to a US buyer that is not owned by a country of concern falls outside it (our reading of the rule). A covered person includes an entity 50 percent or more owned by a country of concern, and a US company may not license bulk US sensitive data or government-related data to a covered person, whatever the contract says. De-identification does not help here: bulk sensitive data is covered whether or not it is anonymised or de-identified.

The rule also reaches buyers that are foreign but not covered. A US company that licenses data within the rule's scope to any foreign buyer must include a clause that bars onward brokerage of the same data to countries of concern or covered persons, and report a known breach of it. Code has a rule of its own: handing source code to a foreign person counts as an export under the Export Administration Regulations, even inside the United States, so code for encryption, defence or other controlled products can need a licence before it goes to a buyer abroad or to a buyer's foreign staff.

What the company owns: copyright and trade secrets

A company owns the copyright in what its employees wrote at work: for a work made for hire by an employee within the scope of the job, the employer is the author and owns the copyright unless a signed writing says otherwise. Work by contractors and agencies is different, since a transfer of copyright must be in a signed writing; code or content an agency built is the company's to license only if the contract assigned it. Material created for a client may belong to the client under its contract, so each contract is checked (our reading).

A company's records are also often trade secrets, and they stay trade secrets only while the company takes reasonable measures to keep them secret and they keep their value from not being generally known. Showing them to a buyer without a confidentiality agreement can weaken that protection, so a company signs an NDA before any sample moves, and licenses rather than sells outright, so that a buyer using the data beyond the licence misappropriates it. Tax withholding on a licence fee is a question for holders abroad: a US buyer withholds 30 percent of a US-source royalty paid to a foreign company unless a treaty claim lowers it, as our withholding guide explains.

A US company that is closing or has closed

A company that is winding down can still license its records, and many sales of company data come from companies that have closed. Who signs depends on where it stands. A dissolved Delaware corporation continues for three years, or longer if the Court of Chancery directs, to wind up, including to dispose of its property; who signs for it, and when a court-appointed trustee or receiver is needed, is covered in our guide to who owns a dissolved company's data with the other cases.

A company in bankruptcy sells through the court. A sale outside the ordinary course needs notice and a hearing, and where the privacy policy in force at the filing barred transferring personal information, it can go only consistently with that policy, or after a consumer privacy ombudsman is appointed and the court approves it, finding no showing that the sale would violate applicable law outside bankruptcy. Google's $10 million bid for Spirit Airlines' records is going through that process and had not been approved when this guide was written, and our guide to selling customer data in bankruptcy sets out each step. For a company closing outside court, the estates page describes how a sale is run, and our guide to keeping records after closing covers what must be kept before anything is sold.

Worked example: a marketing agency

An example, not a real company. A 40-person Chicago marketing agency, operating, has seven years of Google Workspace email, Slack, project boards, creative files and campaign reports. It is not a CCPA-covered business but has a few California clients' contacts in its mail, and it never had a privacy policy for consumers because it serves businesses.

First, what is its own. Its internal mail, Slack, project boards and its own process documents are its records. Its client contracts give the campaign creative and client reports to the clients, so they stay out unless a client agrees. Media it licensed from stock libraries stays out.

Second, the people in it. Its staff signed a monitoring and data-use policy, which the agency updates to name AI training before the sale, since the monitoring notice laws cover collection, not licensing. Clients' and suppliers' names, emails and phone numbers inside the threads are removed in de-identification, and so is any face-geometry scan or voiceprint taken from its Illinois clients' video shoots, which Illinois law puts beyond sale.

Third, the paper. The agency signs an NDA before any buyer sees a sample, licenses rather than sells outright, and checks whether each buyer is foreign before it signs, since a foreign licence of data within the DOJ rule's scope needs its onward-transfer clause. Its records are valued against the $300K per company median of offers to holders for a company's record, and it is introduced to the buyers whose rules take an agency's records.

Questions people ask

Is it legal for a US company to sell its business data to AI companies?

Generally yes, for its own records: no general federal law bars a company licensing them, though the DOJ rule limits licences of bulk sensitive data to some foreign buyers. What decides which records can go is the company's own privacy promises, which the FTC enforces as deception or unfairness, state privacy laws where licensing personal information is a sale, sector laws, and the contracts behind customers' data.

Do I need consent to sell data that contains customer information?

Not always. A company covered by a state privacy law must have disclosed the sale and offered an opt-out; a company that promised not to share data risks acting unfairly or deceptively if it quietly changes its terms to do so. Records de-identified to the law's standard fall outside those rules.

Can I sell employees' emails and Slack messages?

Usually, for business messages on the company's own systems: the federal bar on disclosing stored messages applies only to services provided to the public. Give staff a notice that names the use, take outside correspondents' details out, and remove privileged material; California staff have full CCPA rights at a covered company.

Does selling customer data make my company a data broker?

Usually not. California, Oregon and Connecticut exclude a business selling data about its own customers. New Jersey's 2026 law requires registration from a company that sells or licenses its own customers' data to a data broker.

Can I sell company data to a foreign AI company?

Yes to most, with two checks. A company owned 50 percent or more by a country of concern, or organised or based in one, is a covered person, to which bulk US sensitive data or government-related data may not be licensed, and any foreign licence of in-scope data needs a clause barring onward brokerage. Controlled source code can also need an export licence.

Can a closed company still sell its data?

Yes. A dissolved Delaware corporation continues for three years, or longer if the court directs, to wind up and can dispose of its property, and a company in bankruptcy sells through the court, with notice and a hearing and an ombudsman where its privacy policy barred transfers.

What this means for you

A US company can usually sell its own records to AI companies, and the work is in sorting them before anything leaves. These steps follow the layers in order.

1. Read the policies in force when the records were created: the privacy policy, the terms, and the staff data-use policy.
2. Sort the records: the company's own, its customers' and clients', and the sector-restricted kinds.
3. Update the staff notice to name AI training before the sale, where the company monitors or archives mail.
4. De-identify the people inside the records to the standard the law sets, with an independent check where the records are large or sensitive.
5. Sign an NDA and a licence, never an outright sale of what you would keep, and check whether the buyer is foreign.
6. Value the records against published prices with a valuation, and check any offer with the offer check.

For each kind of record and each system, the rights hub has the detail, and the readiness check asks the questions for your own company. AIData.Broker introduces a company to every buyer whose rules take its data, free for it, paid by the buyers' own referral programmes.