ARTICLE

AI governance

Where does your AI data really go?

Where does your AI data really go?

Where does your AI data really go?

Your AI assistant reads your documents. The question is where they are processed, which law applies to them and who can access them. A plain-language overview.

9 min

read

·

A glass façade seen from below under a cloudy sky

Want to talk about it?

We look at your situation in 15 minutes, with no commitment.

The question almost always comes at the end of the meeting, once everything else is settled: “and our data, where does it go?”. It is asked by the CIO, sometimes by the CFO, and it deserves better than a reassuring answer.

The answer depends entirely on the architecture you choose, and that architecture is decided at the start of the project, rarely at the end.

Three places your documents can end up

The consumer subscription

An employee opens an account, pastes an extract from a contract and gets a summary. Fast, free, and outside any framework: you do not know where the document was processed, how long it is kept, or who can access it. Today this is the default situation in most companies that have not decided anything.

Two recent facts show what “kept” means. Since 8 October 2025, Anthropic keeps for up to five years the conversations of free and paid Claude accounts whose users have agreed that they may be used for training (Anthropic’s terms). And in May 2025, in its lawsuit with the New York Times, OpenAI was ordered by a US judge to keep the ChatGPT conversations that users had deleted, for its free and individual plans (analysis by the National Law Review). Enterprise plans were out of scope: their retention is set by contract.

The enterprise cloud service

The professional version of the same tool. You get a contract, commitments that your data will not be reused for training, account administration and access logs. It is a real step up from a personal account, and it is enough for the vast majority of office uses. To choose the tool itself, our guide Which AI should your business choose? starts from where your documents already live.

What you do not get: control over where the processing takes place, or over which law applies to it.

The private architecture

Your documents stay within your perimeter. A model is called, but the index of your content, the access rights and the audit trail live with you. It takes more effort to set up, and it is not worth it for writing emails. It becomes worth it as soon as the documents involved are the ones that make the value of the company: studies, client files, health data, industrial property.

“Your data is not used for training” only answers a third of the question

It is the sentence every provider puts forward, and it is generally accurate. Yet it leaves three questions open.

  • Where does the processing take place? A contractual commitment on use says nothing about location. The two are negotiated separately.

  • How long are your requests kept? Most providers keep conversations for a few weeks to fight abuse. That is legitimate, and it means your documents exist somewhere other than your own systems during that time.

  • Which law does the provider fall under? An American company remains subject to American law, including for data hosted in Europe. It is the question asked least often, and the one worth stopping on.

The CLOUD Act: what hosting in Europe does not settle

The CLOUD Act, passed by the United States in March 2018, allows US authorities to require a US provider to hand over the data it holds, wherever that data is stored (US Department of Justice). A server in Frankfurt or Dublin does not change the nationality of the company that runs it.

European authorities analysed this conflict of laws as early as 2019. In their joint response to the European Parliament, the European Data Protection Board and the European Data Protection Supervisor consider that a provider has very few ways to comply with such a request without breaching the GDPR (EDPB and EDPS response).

On 10 June 2025, the question was asked under oath in the French Senate. Questioned by the inquiry committee on public procurement, the director of public and legal affairs of Microsoft France answered that he could not guarantee that data of French citizens entrusted to Microsoft would never be handed over on an order from the US government without the agreement of the French authorities, while pointing out that this had never happened (French Senate minutes).

The framework for transfers is fragile too. Data flows to the United States now rely on the EU-US Data Privacy Framework adopted in July 2023. Its predecessor, the Privacy Shield, was struck down by the Court of Justice of the European Union in July 2020 (Schrems II ruling), overnight for the companies that relied on it.

In practice, for ordinary files this risk remains theoretical and a well-configured enterprise plan is enough. For sensitive files (litigation, health, defence, strategic public contracts), the provider’s nationality is part of the risk analysis, just like the server’s address.

The Copilot case, a good illustration of the trade-off

Microsoft has built a European data boundary, the EU Data Boundary, which commits processing to stay within the Union. Since Anthropic’s models (Claude) arrived in Copilot, its documentation states two things: these models are excluded from that boundary, and they are turned off by default for customers in the European Union, EFTA and the United Kingdom. It is up to the administrator to turn them on, user by user or for everyone.

The same page flags a second case: some advanced models are only offered with conversations kept by Anthropic, for up to 30 days, under Anthropic’s terms rather than the Microsoft contract.

In other words: with the same provider, accessing the latest models and staying within the European framework can pull in opposite directions. It is a trade-off, and it is yours to make. You still need to know it exists, and who in your company has the authority to decide it.

Reading a privacy policy in five minutes

An AI provider’s terms often run to more than twenty pages. Five passages are enough to know what you are committing to.

  1. The list of subcontractors: who other than the provider sees your content, and in which countries.

  2. The retention period, plan by plan: it often differs between the free version, the paid individual version and the enterprise plan.

  3. The training clause: on or off by default, and for which plans.

  4. Requests from authorities: does the provider commit to informing you, and does it publish a transparency report?

  5. Reversibility: what becomes of your data and your index if you leave.

If one of these five passages cannot be found, that is already an answer.

The four questions to ask a provider

They fit in one email, and written answers are worth more than answers given in a meeting.

  • In which countries is my content processed, and by which subcontractors?

  • How long is it kept, and in what form?

  • Who on your side can access it, and in what circumstances?

  • What happens if I terminate: what becomes of the index built on my documents?

What it concretely changes in an architecture

On projects where confidentiality is the first criterion, we always separate three things: where the documents live, where the index used to find them lives, and the model that writes the answer. The first two stay with the client. The third becomes a replaceable component.

This split settles compliance, and it also protects your freedom of choice: if a better model comes out in six months, you switch without rebuilding your uses.

What a private architecture looks like

This is the approach behind Celor, the AI assistant we install at our clients’ premises. Here is how it answers the four questions above.

  • Where is the content processed? On your servers, or on a dedicated server opened in your name with a European host. By default, the model that writes the answers is an open-source model running on that same infrastructure: nothing leaves your network.

  • How long is it kept? As long as you decide, since it never leaves your environment. Celor consults your documents when it answers; it learns nothing from them and trains no model on them.

  • Who can access it? Each employee only sees what they are allowed to see: SharePoint or Drive permissions are carried over document by document and applied before the model reads anything. Every answer cites its sources, and a log keeps track of the questions asked and the documents consulted.

  • What happens if you stop? The infrastructure and the index are yours. They stay with you.

If a use case calls for a more powerful model than an internal server can run, Celor can call a provider, chosen with you and in the region of your choice. The Copilot trade-off then arises, but it arises in writing, for that specific use, and you are the one who decides.

Frequently asked questions

Does the GDPR forbid using ChatGPT at work?

No. It requires you to know which personal data is processed, where, by whom and for how long. A supervised professional use, with an enterprise plan and a clear policy, is entirely defensible.

Should everything be hosted in-house?

Rarely. The right split is made by type of document, not on principle. Meeting notes and sales emails do not need the same treatment as an engineering study or a patient file.

Is hosting in Europe enough?

It is a good start. The provider’s nationality and the chain of subcontractors matter as much as the server’s address, as the CLOUD Act shows.

How long does a private architecture take?

A few weeks for a first useful scope, provided you keep it narrow. The projects that go off track are the ones that want to index everything from day one.

In short

The sovereignty of your data is decided in the architecture of your uses, far more than in your provider’s terms and conditions. Three questions are enough to frame the subject: where the documents live, where the index lives, who writes the answer. The rest follows.

If you would like to take stock of your own uses, with the constraints of your sector, let’s talk.