Sovereign AI, concretely: where the models run, where your data lives.

A sovereign AI is not a 'French' AI by its brand: it is an AI whose execution you control. The useful question is not which logo sits on the tool, but where the models run, where the data lives, and who can switch off what.

Everything we build runs on French infrastructure, or on your own servers. The models used by default are European. These points are written into the contract: sovereignty that is not written down binds no one.

The rest of this page walks through the architecture, brick by brick, and what each one changes for an SME.

The default architecture comes down to three links: your data, French infrastructure or your own servers directly, European models. Each brick is open, standard, replaceable.

The exclusion matters as much as the chain: no client data goes to a US API, including during demonstrations. The engine is demonstrated on public data before it touches yours.

Vendor neutrality is the second half of sovereignty. A provider paid by a software vendor recommends what pays it; we are paid for the project, never for reselling licences: the choice of components follows your situation, not a catalogue.

The same holds for connecting to third-party platforms, e-invoicing first among them: we sell no platform, the recommendation stays ours.

The reversibility clause is in the contract: models used, hosting location, exit conditions. If you leave, nothing locks: you take the installation back, or hand it to whoever you choose.

We do not promise that an AI system 'runs on its own': models evolve, your data changes, an unmaintained infrastructure degrades. Managed operations are an ongoing commitment, and they too are contractual.

That is the promise we can keep: not an absence of upkeep, but an absence of lock-in. You stay free to leave, and that freedom is precisely what obliges us to do the work well.

Our data commitments separate two perimeters. Data entrusted during an engagement: it stays inside the architecture described above, is never used to train a model, and leaves with you at the end. Prospecting data, the kind you leave on this site: a guide to send, a diagnostic, and nothing else; never resold, never an unrequested sales sequence.

These commitments already appear in our engagement contracts and our privacy policy. Their condensed version, a one-page data and AI charter, is in legal review: it will be published on this page once approved, so it can be quoted and held against us as it stands.