It can be legal, but only if your customer contracts, privacy notices and applicable law allow it. GDPR, the CCPA, HIPAA, recording-consent rules and FTC guidance may each apply. De-identification reduces risk but does not replace permission. This is general information, not legal advice. Review your position with counsel before licensing any data.
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.
The short answer: it depends on rights, not possession
There is no single law that permits or forbids selling customer data for AI training. Whether a company may do it depends on what the data is, where the people in it live, what the company promised them, and what its contracts say. In many cases some data can be licensed and some cannot.
This guide is general information, not legal advice. It describes the questions that usually matter in the United States and the European Union. Laws differ by state, country and sector, and they are changing. Review your own position with qualified counsel before you license anything.
Who owns the data and who it is about
A company usually controls the records in its systems. It may own copyright in documents it wrote and hold the database as a trade secret. That is different from having free use of the information inside. A support transcript sits in your system, but it contains a customer's words, and possibly their name, account and problem.
Privacy laws protect the people the data is about, whoever holds the file. Contract law protects business customers whose confidential information you received. So the working question is not whether you own the data. It is what rights other people and companies have in it, and what you have agreed with them.
Customer contracts, DPAs and NDAs
For business-to-business companies, contracts are often the tightest limit. Customer agreements commonly say that customer data may be used only to provide the service. Data processing agreements, or DPAs, typically bind a vendor to process personal data only on the customer's instructions. NDAs restrict disclosure of confidential information to third parties.
If you hold data as a service provider or processor for your customers, licensing it for AI training is likely to fall outside those terms unless the customer agrees. Some agreements allow use of aggregated or de-identified data, and the exact wording matters. Counsel should read the actual clauses, including older contract versions that still apply to long-standing customers.
Privacy notices and the FTC's position
A privacy notice is a public statement of what you do with personal data. In the United States, the Federal Trade Commission can treat a practice that contradicts that statement as deceptive. If your notice says you do not sell personal information or share it with third parties, licensing that information for AI training may conflict with it.
Changing the notice is not a simple fix. In a February 2024 business blog post, FTC staff wrote that it may be unfair or deceptive for a company to adopt more permissive data practices, such as using consumers' data for AI training, and to inform consumers only through a surreptitious, retroactive amendment to its terms of service or privacy policy. A company that wants to change how it uses existing data should expect to need clear notice, and in some cases consent.
GDPR: lawful basis and purpose limitation
Where the GDPR applies, every use of personal data needs a lawful basis under Article 6, such as consent, performance of a contract or legitimate interests. Supplying personal data to a third party for AI training is a new use that needs its own basis. Legitimate interests requires a documented balancing test, and regulators may not accept it for every kind of data.
Article 5 adds purpose limitation. Personal data collected for one purpose may not be further processed in a way that is incompatible with that purpose. Data gathered to deliver a service was not obviously collected to train another company's models. Truly anonymous data falls outside the GDPR, but the standard for anonymization is high, and pseudonymized data remains personal data.
CCPA and CPRA: sale, sharing and opt-out
California's CCPA, as amended by the CPRA, defines a sale broadly. It covers disclosing personal information to a third party for money or other valuable consideration. Sharing has its own definition, tied to cross-context behavioral advertising. A paid license of personal information to an AI developer may well count as a sale under this definition.
A business that sells personal information must generally say so in its notices, offer consumers a way to opt out, and honor those opt-outs. Service providers face restrictions on selling data they hold for a business. Information that meets the statute's definition of deidentified is treated differently, but the business must meet specific conditions, including a commitment not to reidentify it. Other states have their own privacy laws with similar but not identical rules.
HIPAA, employee communications and call recordings
Health data held by HIPAA covered entities and their business associates has its own regime. Selling protected health information generally requires the individual's authorization. HIPAA does allow use of data that has been de-identified by one of two methods. Safe Harbor requires removing a specified list of identifiers and having no actual knowledge that the remaining information could identify a person. Expert Determination requires a qualified expert to conclude that the risk of identification is very small. A business associate also needs permission in its agreement with the covered entity to de-identify data at all.
Employee communications, such as email and chat, raise separate issues. Employers often have rights over work systems, but employees may still have privacy rights in the content. This is especially so in the European Union, where consent from employees is difficult to rely on and works councils may need to be involved. Messages also contain third parties' information and privileged or confidential material.
Call recordings depend on consent to record. Federal law in the United States generally allows recording with one party's consent, while some states require the consent of everyone on the call. A notice that calls are recorded for quality and training was written with staff training in mind. Whether it covers licensing recordings to train AI models is uncertain and should be checked.
De-identification reduces risk, but it is not a magic key
Removing personal and client identifiers before data leaves the company is the single most useful risk control. It can take data outside some privacy laws, and it limits harm if something goes wrong. Outrich Data de-identifies supplier data before any buyer sees it and documents the method.
It does not solve everything. De-identification can fail, especially in free text and in small or unusual populations where details point to one person. Different laws set different standards for it. And it does nothing about contract limits. A confidentiality clause that bars disclosure of customer information can still apply after names are removed.
A sound process runs in this order: confirm contract rights, check notices and applicable privacy law, exclude what is not cleared, de-identify what remains, and document each step. Have counsel review the result before any license is signed.
Common questions
Do we need customer consent to license data for AI training?
Is de-identified data exempt from privacy law?
Can we update our privacy policy to allow AI training?
Do these rules apply to business data with no personal information?
Sources
- GDPR, Articles 5 and 6 (lawful basis and purpose limitation)2026-11-16
- California Attorney General, California Consumer Privacy Act (CCPA)2026-11-16
- HHS, guidance on de-identification of protected health information under HIPAA2026-11-16
- FTC, AI (and other) companies: quietly changing your terms of service could be unfair or deceptive2026-11-16