Frequently asked questions on sovereign AI

No, that is not its role. The assistant decides nothing: it searches, summarises and suggests a first draft. Your staff review, correct and approve. The time saved on repetitive tasks goes back into public service: helping residents, following up on cases, projects that had been waiting.

Yes. You ask it a question in your own words, as you would a colleague. Getting started takes 30 minutes, and trained champions in your departments support their colleagues. During the trial, real usage is measured: active users and message volume, using indicators agreed with your DPO.

It can happen: no AI is infallible. That is why every answer drawn from your documents cites the document it comes from, so staff can check it, and why the assistant takes no decisions: it prepares, the staff member reviews and remains responsible for what they send or sign.

It stays in the European Union. Accounts, histories and document bases are hosted on a server you control; GPU processing is separate from account, document and history storage. No reuse of your data to train a model: the open weight model is used for inference only, and neither your questions nor your documents change its weights. Erasure procedures and their frequency must be verified in configuration and contracts.

Yes, in particular through your inter-municipal body. The tool can be pooled across the community of municipalities or the urban community: the town clerk of a village has access to it just like the departments of the main town, and the cost is shared.

No. The three-week trial is there precisely so that you can decide on facts: at the end, you choose to roll out, adjust or stop, with no obligation to continue.

A sovereign AI is a generative AI that an organisation keeps under its own control from end to end: an open weight model running on infrastructure it governs, hosted in the European Union or on premises, with its data staying in its hands and no dependency on a vendor subject to an extraterritorial law such as the US CLOUD Act.

EU hosting alone does not determine the applicable jurisdiction. Assess contracting entities, subprocessors, support access and key control across the full processing chain. See our Sovereignty page.

Le Chat by Mistral is a product that runs on Mistral's servers. We deploy and operate for you a sovereign infrastructure, connected to your documents (RAG), with a daily wipe of the model's GPU server and integration with your SSO. You do not depend on a closed service: your data and your index can be exported on request, and you can change model at any time.

Albert (DINUM) is an excellent initiative, but it is aimed primarily at central government departments, and you need the in-house skills to operate it yourself. We provide a turnkey solution: deployed, maintained and supervised for you, with dedicated support, advanced document features (RAG with cited sources, summaries, briefings; podcasts and meeting transcription being rolled out) and live references in local authorities.

They are online services operated by two US groups, Microsoft and OpenAI. PixelRaise, a French company, instead deploys an open weight model on infrastructure you control. What the vendors and the legal texts state (see the sources):

• Hosting and data residency: for EU customers, Microsoft Copilot keeps data at rest within the EU Data Boundary, apart from limited pseudonymised data, but inference may be processed in the United States, Canada or Australia at peak times if “flex routing” is allowed, which is the default setting for eligible tenants created after 25 March 2026; the optional Anthropic models are currently excluded from the boundary. ChatGPT Enterprise lets new customers store content at rest in Europe (EEA and Switzerland) and, if eligible, run GPU inference there; authentication, routing and other processing may still take place outside the region.
• Training: Microsoft states that prompts, responses and Microsoft Graph data are not used to train foundation models; OpenAI states that by default it does not train its models on business plan data unless the customer explicitly opts in.
• Jurisdiction: the CLOUD Act requires a provider subject to US law, when served with lawful process by US authorities, to disclose data in its possession, custody or control, even when it is stored outside the United States. On 10 June 2025, questioned under oath by a French Senate committee of inquiry, the director of public and legal affairs of Microsoft France said he could not guarantee that French citizens' data entrusted to Microsoft through UGAP, the French public purchasing agency, would never be handed over following an order from the US government without the explicit agreement of the French authorities, while noting that this had never yet happened.
• Model: Copilot relies on models that Microsoft updates regularly, including those of OpenAI and Anthropic; ChatGPT Enterprise relies on OpenAI's proprietary models. In both cases, the model runs inside the service of the vendor or its subprocessors, not on your infrastructure.
• Reversibility: when you leave these services, the model stays with the vendor. With PixelRaise, the open weight model can be reinstalled on other infrastructure, and your data and document index can be exported on request (commitment ENG-04).

Document features, cost and support depend on the plan and contract chosen; PixelRaise provides local support and scoped operations.

The solution is designed to make your compliance easier: data hosted in the EU, structural minimisation (daily wipe of the model's GPU server), no reuse of your data to train a model (the open weight model is used for inference only, and neither your questions nor your documents change its weights), experimental pseudonymisation of personal data, in testing. Your organisation nevertheless remains the data controller under the GDPR: we provide the documentation and support your DPO, in particular on the data protection impact assessment (DPIA) and the transparency required by the EU AI Act. See also our AI & GDPR page.

The model is fixed cost: the LLM's GPU server is billed for working hours only and there is no per-user licence, so the bill does not grow with the number of users. Estimate your own case with our ROI simulator.

Three weeks: one week of scoping and installation, then a two-week pilot with around twenty users, ending in a go/no-go milestone. You decide whether to roll out on the basis of measured results, with no blind commitment.

Exclusively in datacentres located in the European Union. User data (accounts, histories, knowledge bases) stays on a server you control; GPU servers process documents during inference. Effective erasure depends on verified procedures; stopping the GPU does not purge histories, logs or backups.

Yes. Our stack is model agnostic: Mistral, Llama, Qwen or other open source LLMs, adapting to the GPU resources available. You are never locked into a single model. Details on our LLM models page.