In short
- Two questions to ask: where does your AI agent run, and where does the AI model read your data. The two answers can be different.
- Claude, the model we use, can process your data without leaving the EU through Amazon Bedrock or Google Cloud. Anthropic's direct API and Microsoft Foundry do not allow this (checked on 12 October 2026).
- You choose where the agent runs: your cloud in an EU region, your servers or a dedicated server with us. Our supervision will only receive indicators and alerts.
Data protection holds back many businesses. In 2025, among EU businesses (10 employees or more) that had considered AI without using it, 48.83% cited the fear of breaching data protection and privacy (Eurostat). Saying “your data stays in Europe” is not enough. Here is how, in practice, with the limits.
Two separate questions: where the agent runs, where the model reads your data
An AI agent is a small programme. It reads an email, an invoice or a client record, then prepares an action. Depending on the autonomy level chosen with you, it waits for your team to approve, or it acts and reports back. This programme runs on a server. First question: which one, and where?
To understand and write, the agent sends an extract of text to an AI model, here Claude, built by Anthropic. The model reads this extract and returns a reply. This computation takes place in a data centre. Second question: where?
An agent installed in Paris can send your texts to a model that computes in the United States. Both answers matter.
Claude in Europe: two possible routes, three dead ends
We re-read the official documentation on 12 October 2026.
Anthropic's direct API does not offer Europe. The API is the entry point through which a programme talks to the model. Processing there is “global” (anywhere in the world) or limited to the United States. The data the workspace keeps is stored in the United States only (Anthropic documentation).
Claude Platform on AWS does not either. Do not confuse it with Bedrock: this offer is sold by Amazon, but it is Anthropic that runs the model. Processing there is also global or limited to the United States (Anthropic documentation).
Microsoft Foundry does not either. Microsoft's AI platform on Azure offers Claude in global mode or in a “US data zone”. No Europe zone (Anthropic documentation).
Amazon Bedrock offers an “EU” profile. A profile is an Amazon setting that spreads your requests across several of its regions (groups of data centres) in the same zone. If the agent calls it from an EU region, the EU profile sends requests to Claude (Opus 5.5, Sonnet 5.5 and Haiku 5.5) to six regions: Frankfurt, Stockholm, Milan, Spain, Ireland and Paris. A trap: called from London, the same profile adds London; called from Zurich, it adds Zurich. These two cities are outside the EU. So we will never call it from these two regions. Another limit: for these models, Bedrock does not offer processing limited to a single EU region (Opus 5.5 page on AWS). Amazon advises checking the actual list of destinations from each starting region, without relying on the profile name alone (AWS documentation). We will do so at every installation.
Google Cloud offers an “eu” endpoint. An endpoint is the address to which the agent sends its requests. According to Google, this endpoint “strictly” covers EU member states (Google Cloud documentation). Opus 5.5, Sonnet 5.5 and Haiku 5.5 are available there (Opus 5.5 page on Google Cloud). The global endpoint, on the other hand, can process a request anywhere in the world. Staying in Europe costs a little more: Anthropic states a 10% surcharge compared with the global endpoint (Anthropic documentation).
An open model at your site: the hardware needed
Another route: an open model. Its maker publishes it, and you install it on your own machines. No text then goes to a model provider. Mistral AI publishes several under the Apache 2.0 licence, a free licence.
The hardware is the real issue. An AI model runs on graphics processing units (GPUs): the computing chips of graphics cards. A few benchmarks published by Mistral:
- Mistral Small 4: at least “4x NVIDIA HGX H100, 2x NVIDIA HGX H200 ou 1x NVIDIA DGX B200”, in Mistral's wording. This is data centre equipment (Mistral).
- Mistral Large 3: a server with 8 A100 or H100 graphics processors, in a compressed version of the model published by Mistral (Mistral).
- Ministral 3 (3, 8 or 14 billion parameters: that is its size): NVIDIA offers versions adapted to PCs and laptops fitted with RTX cards (Mistral).
The smaller the model, the simpler the hardware. But these models do not all do the same things: we will test them on your real cases before choosing. This route mainly suits businesses that require it and can invest in this hardware.
Option 1: your cloud account, in an EU region
The agent runs in your Amazon Web Services or Google Cloud account, in a European region. The model goes through Bedrock (EU profile) or Google's “eu” endpoint, in that same account.
- What you gain: you keep the keys to the account. You see what is running and can withdraw our access. With Amazon, Claude usage appears on your AWS bill.
- What it requires: a cloud account and someone to administer it, at your company or your service provider.
- A case to anticipate: for example, an SME with 40 employees set up on Microsoft. Microsoft Foundry does not offer Claude processing limited to Europe. It can open an AWS or Google Cloud account for the model alone, choose option 3, or use an open model.
Option 2: your own servers
The agent runs on your servers, on your premises or with your host. For the model, two cases:
- Claude through Bedrock or Google Cloud in Europe: the text extracts leave your servers, but stay in the EU.
- An open model installed at your site: nothing leaves. You then need the GPU servers described above.
You maintain the hardware. Agent updates and remote access will follow terms defined with you and set out in your quote.
Option 3: a dedicated server with us, in Europe
You do not want to manage anything. The agent will run on a server reserved for your company alone, which we will rent in Europe; the host and the country will be specified in your quote. The model will go through a Bedrock or Google Cloud account in our name, set to Europe.
- What you gain: nothing to install or maintain.
- What it implies: we process your data on your behalf. The GDPR then makes us a “processor” (article 4) and requires a contract. It sets out in particular the subject matter, duration, nature and purpose of the processing, and the type of data concerned. Amazon or Google become our own sub-processors: for this we need your prior written authorisation (article 28, text on the CNIL website).
- The question to ask us: how to recover your agents and your data if you leave. These terms are defined with you and set out in your quote.
- Cost: on quote.
In options 1 and 2 too, we may need to look at your data to install or troubleshoot an agent. But the GDPR counts simple consultation among data processing operations (article 4). As a precaution, we then provide for the same contract.
What will reach our supervision, and what will not
Whichever option you choose, we will follow your agents from our CRM, our client tracking tool. Only indicators and alerts will arrive there.
What will reach it:
- each agent's indicators: number of actions prepared, share approved as is by your team, errors;
- the model's usage, in “tokens”, the unit of text that the provider bills;
- reports: the agent will itself report a bug, an error that keeps repeating or a change of AI model. A ticket will open, and you will follow it in your future client area.
What will not reach it:
- the content of your emails, invoices, quotes and accounting documents;
- the names and contact details of your clients and suppliers;
- the detailed log of each action (what the agent read, what it proposed, who approved): it is planned to stay in your installation (with option 3, on the dedicated server).
A report will contain a technical detail: which agent, which step, which error message. We plan to filter it to remove personal data. And in the stop cases planned for each agent, the agent will stop and hand over to your team, at your site.
These offers change quickly. This article reflects the documentation read on 12 October 2026. We will re-check it before every installation.
Going further
- Data and security: our commitments on hosting and access to your data.
- Supervision: the planned monitoring of your agents, the alerts and the monthly report.
- Our agents: what each agent will prepare, and what your team will approve.
- The audit: we review your tools, your data and your constraints, then recommend an option. It is paid, on quote, and its price is deducted from the set-up quote if you sign it within 30 days. Request an audit