As soon as you hand a model a customer email, a file or a contact list, you are processing personal data. The GDPR then applies exactly as it does for any other tool, with no AI exemption.
This entry does not replace advice from a legal professional. It aims to give you the vocabulary and the right questions, not ready-made answers about your situation.
What we are talking about
The GDPR governs the processing of personal data, meaning any information relating to an identifiable person. A name, an email, a phone number, but also a message written by a customer or a recording of a conversation.
The most common confusion is thinking a quick test does not count. Pasting a customer file into an assistant to "see what it does" is processing personal data, with everything that entails.
The four questions to ask
Do I have a legal basis?
Processing must rest on a ground: consent, a contract, a legal obligation, or legitimate interest. Using AI creates no new ground, you rely on the one already covering that data.
Where does the data go?
Most large providers process outside the European Union. That is not forbidden, but it requires a framework, and many now offer European hosting. It is the first thing to look for in their documentation.
Is the provider a processor?
If you entrust data to them on your behalf, yes, and a data processing agreement is needed. Business offers provide one, free offers rarely do.
Is my data used for training?
The most concrete question. Business offers generally commit to not reusing your content, consumer offers far less often. See Training data.
Habits that reduce the risk
Minimise. Send the useful paragraph, not the whole file. The simplest principle to apply and the most effective.
Anonymise where possible. Many analyses work just as well without names.
Write an internal rule. What may be given to an assistant, what may not. Without a written rule, everyone decides case by case.
Keep the decision human. A fully automated decision producing a significant effect on a person is specifically regulated: see Human in the loop.
A model does not forget what it was given the way a database does. If content was used in training, removing it is not a simple row deletion. One more reason to verify the non-reuse commitment before sending, not after.
Frequently asked questions
Does self-hosting settle the question?
It simplifies part of it, since the data does not leave. The other obligations remain, and you become directly responsible for security. See Self-hosting.
Must the people concerned be informed?
Information about processing and its recipients is part of the general obligations. Using a new tool may therefore require updating your notices, which is worth checking with a professional.
Does an assistant handling my emails fall under this?
Yes, and it is one of the most common and most sensitive cases, because an email thread often contains far more personal information than people realise when handing it over.
Where do you find reliable references?
From your national data protection authority, which publishes AI-specific guidance, and from legal counsel for your situation. Our Claude Cowork course covers these hygiene habits when deciding what to hand a model.