Sovereign AI: what it guarantees, and what it doesn’t

“Sovereign AI” has become a catch-all term. Here is what we mean by it, what it allows in practice, and what it does not guarantee.

What we call sovereign AI

The AI model, its tools and your documents are installed on the infrastructure you designate: your servers, or a private cloud you control. We use open-weight models, such as Llama and Mistral, chosen according to your use case, your working language and your hardware.

Inside that perimeter, your documents, your questions and the answers stay with you. Nothing is sent to an external AI API. Only downloads of models and updates come in, and we set them up with you. The access rights and the documentation are handed over to you, and the code stays yours.

Three concrete uses

• An internal assistant. Your teams ask questions in plain language: procedures, standard contracts, internal notes, meeting minutes. The assistant answers from your documents and shows where the answer comes from.
• Search by meaning, not only by keyword. Each result points back to the original document, which you can check.
• Coding help, locally. Your developers get a coding assistant that runs on your side: completion, explanation, tests, review. Your source code does not leave for a vendor.

What it does not guarantee

Three limits, which we state before starting:

1. Performance. An open model hosted on your side can be less capable than the largest cloud models on some tasks. On others, such as answering from your documents, a well-integrated open model can be enough. We do not claim it in advance: we measure it with your users, on your use cases.
2. The cost of hardware. A model of useful size needs graphics cards or a suitable server. The need depends on the size of the model and the number of users; we put a figure on it with you before any commitment.
3. Compliance. A deployment on your side does not, on its own, make your organisation compliant with a regulation. It reduces data flows to the outside. Compliance remains work to do with your security officer and your legal counsel.

How to start

We begin with a proof of concept, in four steps:

1. Scoping: together we choose a use case, the documents involved and the success criteria.
2. Deployment on your side: we install the model and its tools on your infrastructure.
3. Evaluation: your users test on your real documents; we measure quality with them, then adjust.
4. Handover: we give you the code, the documentation and the access rights, so your teams can run the solution.

The proof of concept starts at €8,000. Beyond that, the cost depends on the scope, the hardware and the number of users, and we price it with you once the proof of concept is done.