Sovereign AI: what happens to your document when you give it to an AI assistant

A district nurse pastes a care plan into an AI chat to make it shorter. The result is fine. Only, the care plan is now also on a server that you as privacy officer never said yes to.

Moments like this happen every day, usually with good intentions. This piece explains what happens technically when someone gives a document to an AI assistant. And which questions you ask per service to keep it within Europe.

What happens when you press send

An AI assistant is usually a language model that runs at a provider. What you enter goes there:

1. Your question and the text you send along go over the internet to the provider.
2. The model processes the text on the servers of that provider and writes an answer.
3. The answer comes back. Depending on the terms, the provider keeps your input for a while, for example for abuse monitoring or to improve the model.

With American providers such as OpenAI, Microsoft and Google, that processing falls under US law, even if you choose a European server option. That is the same CLOUD Act question from 2018 that applies to mail and storage.

What sovereign AI means

Sovereign AI means that the data, the model and the processing remain under your control. Or under that of a party that falls under European law. In practice, it is a scale, not a yes or no. A model on your own server is the most sovereign. A European provider is a step in that direction.

What RAG is: an assistant that answers from documents

Many organisations want an assistant that answers on the basis of their own protocols, policies or handbooks. That is called Retrieval-Augmented Generation, RAG for short. It works like this:

1. The documents of the organisation are indexed in their own search index.
2. With a question, the system looks up the passages that matter.
3. Those passages go to the language model together with the question.
4. The model writes an answer on the basis of those passages.

The advantage for privacy: not your whole archive goes to the model, only the pieces needed for one question. The index itself stays with you or with your supplier.

Which rules apply?

For personal data, the GDPR applies. Transfer to a country outside the EU requires a valid legal basis, and health data is a special category of personal data with extra protection.

The European AI Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024 and applies in stages. The prohibited applications have applied since 2 February 2025. AI that decides about citizens, for example on benefits or permits, counts as high risk and gets strict requirements for transparency, human oversight and documentation. An assistant that only summarises a text generally does not fall under that.

Five questions per AI service

1. Where does the model run, and where is the search index?
2. Which law does the provider fall under, and which sub-processors are there?
3. Is your input used for training, and is that in the contract?
4. How long is input kept, and can you turn that off?
5. Can you turn the AI functions on or off per group?

How it works at Noveu

The search happens with us: the search index of the assistant and the conversion of your question into a search vector (bge-m3) run on our own infrastructure in Falkenstein. All AI functions (the assistant, the mail helper, summarising, sheets and presentations) have the text written by Mistral AI, a French company; according to Mistral, that data is stored in the EU by default. For the assistant, what goes to Mistral is your question, up to five earlier turns, the passages found with their source address and what the assistant looks up for you, such as emails, tickets and invoices; for the mail helper, the sender, subject and text of the email. Email addresses, phone numbers, IBANs, IP addresses, BSNs and web addresses in the text that start with http or https are replaced by placeholders beforehand. The parties in the chain are on the list of sub-processors.