Short answer
Data generally can't be licensed for AI training when the company lacks the rights to it: patient records, client-confidential files, personal data without a lawful basis, privileged legal communications, other companies' confidential information and data licensed from third parties. High-risk categories are usually excluded outright rather than de-identified.
The most useful thing a company can do before talking to an AI data buyer is decide what is off the table. Buyers will often ask for everything. Most companies cannot license everything, and the ones that try create problems for themselves later. This guide lists what generally stays out and why.
Key takeaways
- The test is rights, not possession. Having data in your systems does not mean you may license it.
- Six categories usually stay out: regulated personal data, client-confidential files, privileged communications, third-party confidential information, sensitive employee records and data you licensed from someone else.
- Exclusion is safer than de-identification for high-risk material.
- What remains is still substantial: the operational record of how your team works.
The test: rights, not possession
Every dataset sits under layers of obligation. Laws restrict some kinds of information. Contracts restrict others. Your own privacy policy and employee notices restrict more. A dataset can be licensed only if it clears all of those layers.
A law firm that advises on these deals puts it plainly: most companies cannot license everything a buyer requests. The work is in drawing the line.
What generally stays out
| Category | Examples | Why it stays out |
|---|---|---|
| Regulated personal data | Patient records, financial account data, children's data, biometrics | Sector privacy laws restrict use and disclosure |
| Client-confidential files | A law firm's client files, an accountant's client records | Professional duties and client agreements |
| Privileged communications | Email and chat with lawyers | Disclosure can waive privilege |
| Third-party confidential information | Customer data under NDA, partner pricing | Confidentiality clauses in contracts |
| Sensitive employee records | HR files, payroll, health and benefits, union communications | Privacy law, labor law and trust |
| Licensed or borrowed data | Purchased datasets, vendor content, open-source code with conditions | You hold a license, not the right to sublicense |
Regulated personal data
Some information is protected by laws written for a specific sector.
Health information. In the United States, HIPAA restricts how covered organizations and their business associates use and disclose protected health information. Patient records are not licensable in ordinary form. HIPAA does recognize two formal de-identification methods, but meeting them is a specialist task. See de-identification for AI training data.
Financial information. Customer financial records held by financial institutions are subject to their own federal privacy rules.
Children's data, biometrics and recordings. Children's privacy law, state biometric laws and call-recording consent laws each add restrictions.
International data. Records about people outside the United States may fall under GDPR or other national laws.
This does not shut out whole industries. A healthcare services company still has non-clinical operations: scheduling workflows, vendor management, internal process documentation. The patient records stay out. The operational record may not have to.
Client-confidential files
Professional firms hold information that belongs, in a practical sense, to their clients. A law firm's matter files, an accounting firm's client workpapers and a consultant's client deliverables are covered by duties of confidentiality and by engagement terms. They generally cannot be licensed without the rights to do so.
Privileged communications
Communications with your lawyers are protected by privilege, and that protection can be lost through disclosure. Privileged email, chat and documents should be excluded before any dataset is assembled.
Other companies' confidential information
Your systems are full of information that came from someone else: a customer's specifications, a vendor's price list, a partner's roadmap. NDAs and commercial agreements usually restrict sharing it. Removing the company's name does not necessarily satisfy a confidentiality clause.
Sensitive employee records
HR files, payroll and tax records, health and benefits information and union communications are commonly excluded. There is also a labor law point. Under the National Labor Relations Act, employees' discussions of pay, scheduling and working conditions can be protected activity, whether or not the workplace is unionized. Where there is a union, a collective bargaining agreement may restrict a deal or require bargaining over it.
See employee privacy when a company licenses workplace data.
Data you licensed from someone else
If you bought a dataset, subscribed to a data service or built on a vendor's content, you hold a license with its own limits. Source code deserves a check as well: open-source components and vendor code may carry conditions or competing ownership claims. Reports generated by a software tool may be restricted by that tool's terms.
Trade secrets: licensable, but think first
Pricing logic, strategy documents and proprietary methods are yours to license. The risk is different. Trade secret protection depends on keeping the information secret, and sharing it widely can weaken that protection. A buyer could also use the data to train a system that competes with you, and what a model has learned cannot practically be removed. Many companies leave their most sensitive know-how out.
Why exclusion beats de-identification
It is tempting to include everything and rely on de-identification. Three problems make that a poor plan for high-risk material.
- It is imperfect. Stripping identifiers from millions of free-form messages is much harder than from structured fields.
- It does not remove confidentiality. A sensitive complaint is still sensitive without the name.
- Linked records can re-identify. Buyers want records connected across systems, and connections can point back to a person or a small group.
De-identification is the right tool for ordinary operational records. Exclusion is the right tool for the categories above.
What is left
Quite a lot. The focus is operations: how your team works, your processes and your documentation. That means project discussions, support workflows, sales process, internal procedures, operational planning and the everyday coordination that fills chat and email. See what data AI companies buy.
When you describe your company to DataRelai, leave out customer names and personal data. The intake form asks only which systems you use and what kinds of records they hold. There are no uploads.
Frequently asked questions
Can a healthcare company license any data for AI training?
Its patient records generally not, without the rights to do so. Its non-clinical operational records may be a different matter. Suitability depends on the buyer and on legal review.
Can we license data that contains personal information if it is de-identified?
For ordinary operational records, that is the standard approach. For regulated or highly sensitive categories, exclusion is safer.
Is our customer list licensable?
Buyers of AI training data are not looking for contact lists, and privacy laws and customer contracts often restrict sharing them. The value is in operational records, not in who your customers are.
Who decides what is excluded?
Your company does, with your attorney. A buyer may ask for more. You approve the final scope.
Sources
- Frankfurt Kurnit Klein & Selz, "Selling Company Data for AI Training: What to Know Before You Say Yes" (October 1, 2026)
- U.S. Department of Health and Human Services, guidance on de-identification under the HIPAA Privacy Rule
This guide is general information, not legal, tax or financial advice. Whether a particular dataset can be licensed depends on what it contains and on obligations your company already has.