GDPR and AI training – can you sell customer data?

General information, not legal advice. This page describes how we see the rules in broad terms. Every dataset is unique, and national rules vary. Before a license, you should always get help from your own lawyer or data protection officer.

GDPR and AI training in brief

When data contains personal data, GDPR applies – even if the data is to be used to train an AI model. That covers emails, chats, support tickets, CRM history and recorded calls. Most of a company's operational data contains personal data in some form.

Three questions determine whether AI training is possible under GDPR:

1. Is there a legal basis for the new processing?
2. Is the new use compatible with the purpose the data was originally collected for?
3. Can the data be de-identified sufficiently before it leaves you?

Can you sell customer data?

The question comes up often, and the honest answer is: it depends. You cannot assume customer data may be licensed just because it is in your systems. Customers provided their details for a specific purpose – to get an item delivered, a case resolved or an invoice sent.

In many cases the way forward is to de-identify the data before it becomes relevant for a license. In other cases the conclusion is that the data should not be licensed at all. We go deeper in the guide can you sell customer data?

Legal basis – legitimate interest is not automatic

Many point to legitimate interest as the legal basis. In Opinion 28/2024, the EDPB (European Data Protection Board) made clear that legitimate interest is not automatic for AI training. It requires a documented balancing test in which the company's interest is weighed against the rights and expectations of the people concerned.

Purpose limitation – compatibility with the original purpose

GDPR's purpose limitation principle means that reusing personal data for AI training must be compatible with the original purpose. European data protection authorities have stressed this point, and also that creating synthetic data from personal data is itself processing of personal data.

Anonymization and pseudonymization

It is easy to say data is "anonymized". It is harder to achieve.

• Anonymized data – no person can be identified by any means reasonably likely to be used, not even by combining it with other information. Only then does the data fall outside GDPR.
• Pseudonymized data – names and direct identifiers have been replaced or removed, but people can still be identified indirectly. It is still personal data.

Free text is particularly difficult. An email can point to a person through context, a job title or an event, even when the name has been removed. The anonymization work itself is also processing that needs a basis under GDPR.

Data we don't handle

Health data and other special categories of personal data under Article 9 GDPR are outside what we broker.

Employees' emails and chats

Where employees' mailboxes or chats in Slack and Teams are concerned, workplace privacy rules also apply. Depending on your country, you may also need to inform or consult employee representatives or a works council before a decision is made.

For Slack there is one more factor: Salesforce has reportedly restricted bulk export and use for LLM training through Slack's API terms. Exporting your own archives therefore also needs a legal review of the platform terms.

Copyright and the AI Act

Besides GDPR, there are two more sets of rules to know about:

• The DSM Directive's text and data mining exception (Directive (EU) 2019/790, Article 4) – text and data mining is permitted unless the rights holder has reserved those rights in an appropriate way, such as in machine-readable form for online content.
• The AI Act, Article 53 – providers of general-purpose AI models must respect such reservations and publish a summary of their training content.

How we work

We always start with rights and personal data, not with the price.

• We prioritize data without personal data, such as published content and media archives.
• We carry out an assessment of legal basis and de-identification before a license is on the table.
• Nothing is shared until a non-disclosure agreement and an accepted offer are in place.
• Delivery only takes place after de-identification.
• Our documentation helps buyers meet the AI Act's transparency requirements.

If we see that a dataset cannot be licensed in a defensible way, we say so early. That can apply, for example, to parts of customer service data with a lot of free text. How the license itself is structured is described on our page about data licensing.