The situation is almost always the same. Teams have discovered what artificial intelligence saves them. They use it. And management discovers, often after the fact, that internal documents have been pasted into a service hosted outside Europe.
From there, two options present themselves, and both are bad.
The false choice
Ban it. This is the reflex answer, and the least effective. A memo does not remove a habit that saves two hours a day: it moves it onto personal phones, out of all visibility. You have not removed the risk, you have removed your ability to see it.
Allow it. That is honest, but it means accepting that documents travel to an infrastructure you do not control, under a law that is not yours. For a public authority, a professional practice, or a company handling health, HR or contract data, that is not a decision anyone takes lightly.
For a long time this choice was inescapable, because every usable model was remote. That is no longer true.
What changed
Open models, hostable on your own infrastructure, are now good enough for most office tasks: drafting, rewriting, summarising, extracting from a document, answering from an internal document base.
"Good enough" deserves unpacking, because that is where the credibility of the project sits. A self-hosted open model does not match the best models on the market on the hardest tasks — long reasoning, complex code, fine-grained analysis. It matches them comfortably on what an administrative officer asks of it ten times a day.
So the right trade-off is not "the best model", but the best model that satisfies the constraint. When the constraint is "no data leaves", the field narrows — and there is still plenty left to work with.
What it actually takes
Worth saying plainly: this is not a subscription you switch on. You need:
- a machine, sized for the chosen model and the number of concurrent users;
- a closed network, where access happens from the inside only;
- named accounts, so you know who did what — that is what makes the usage auditable, and therefore defensible;
- maintenance, because a model you install and then abandon becomes a liability within eighteen months.
That last point is the one most often forgotten at decision time, and the one that costs most once forgotten.
What you get back
A verifiable guarantee. "No outbound calls" is not a contractual promise, it is a property of the architecture. It can be checked at network level, by somebody who does not have to take your word for it.
The end of workarounds. When the internal tool is decent and available, unsanctioned use stops on its own. It is not a question of policy, it is a question of service quality.
A predictable cost. You pay for infrastructure, not for per-token consumption that rises with adoption.
The Grand Chambéry case
For the Grand Chambéry agglomeration authority we deployed a fully internal conversational assistant: an open French model, access over the internal network only, named accounts, hosting in France. No outbound calls.
Delivered end to end, infrastructure included, then maintained for twelve months.
What makes the project interesting is not the technical feat — there is none. It is that the constraint was set first, and the whole architecture followed from it. When you start by picking a tool and then try to bring it into compliance, you never get there.
How to tell whether this is you
Ask yourself three questions.
Are your teams already using AI, with or without permission? Would a document you handle routinely cause a problem if it left? Could you say today who sent what, and where?
If the answers are yes, yes, no — you do not have a tooling problem. You have an architecture problem, and it is solvable.
The two-hour audit
We map your flows, and you keep the document.
Remote or on site. No commitment, and no automated follow-up.
Read next
Hire a developer or use a studio: the sum nobody doesPeople compare a salary with a day rate and conclude far too quickly. Here are the lines that sum leaves out — and the cases where hiring is still right.