Blog · Public sector

AI in a local authority: where to host citizens' data, and why 'in France' is not enough

Before handing a citizen's request to an assistant, a local authority must answer four questions: where is the data, who can access it, who trains what, and who validates. The server's location is only the first.

By Christophe Delebarre · 4 November 2026 · 4 min read

Article prepared with the help of artificial intelligence from our monitoring, then reviewed and approved by Christophe Delebarre before publication.

A town hall receives citizens' requests every day: civil status, planning, school enrolment, complaints. Sorting them, spotting missing documents, preparing a reply: an AI assistant does it well, and fast. But every request contains personal data, sometimes sensitive. "Where does this data go?" is the first question a chief executive asks, and it is the right one. It is simply not sufficient.

First question: where is the data

The GDPR does not forbid hosting outside France, nor even outside Europe, but it strictly frames transfers outside the Union. In practice, for a local authority, hosting within the European Union is the simple answer, and hosting in France the comfortable one, because it avoids having to justify anything.

But "server in France" says nothing about the operator. A data centre in Paris, run by a company subject to US law, remains exposed to access requests from that country's authorities under the 2018 CLOUD Act. What matters, then, is the nationality of the operator, and of its parent company, as much as the address of the data centre. For a local authority, it is a good selection criterion, and a simple question to put to any provider.

The Isalia platform is hosted by OVH, a French operator, in France. We say it without making it a sufficient argument: it is the starting condition, not the conclusion.

Second question: who can access the data

An AI assistant that reads citizens' requests passes them to a language model for analysis. Where does that model run? At whose premises? Under what terms of use? That is where the real transfer hides. A tool hosted in France that sends the text of every request to a consumer-grade American model has moved the problem, not solved it.

Our rule is simple: whenever citizens' or company data is involved, processing goes through a European model, hosted in Europe, under a processing contract. Each authority has an isolated, encrypted space with its own keys. Processors are listed and disclosed, which the GDPR requires anyway.

Third question: who trains what

Many terms of use of consumer tools allow the publisher to reuse what is submitted to improve their models. For a social housing application or a welfare file, that is unacceptable, and yet it is what happens when a staff member pastes the text into a free chatbot for lack of a framed tool.

The answer fits in one contract line, to be demanded of any provider: submitted data is never used to train a model. That is our written commitment, and it applies to every processing.

Fourth question: who validates

This is the one people forget, and it is the most important. A local authority takes decisions that affect people's rights: allocating a place, a benefit, a permit. The European AI regulation classes these uses among the most regulated, and the GDPR already prohibits fully automated decisions with legal effects without human intervention.

The line to hold is clear: the assistant sorts, prepares, proposes. The staff member decides and signs. Nothing goes to a citizen without a staff member's validation, traced and timestamped: who validated, what, when. That trace is also what protects you in case of appeal, and what you will show the data protection authority or the regional audit chamber if asked.

The four questions, in one sentence. Data hosted in France with a European operator, a European model under contract, no training on your data, validation by a staff member before anything is sent. If a provider does not answer all four clearly, walk away.

What it looks like in a department

A planning department receiving thirty requests a week: the assistant classifies them by type, checks the file's documents, prepares an acknowledgement and, for simple cases, a draft reply based on the department's procedures. The case officer reviews, corrects, validates. The time saved goes to complex files, the ones that need a person. The same pattern applies to meeting minutes and regulatory watch, described on our local authorities and public bodies page.

And public procurement?

An AI project in a local authority goes through public procurement, with its thresholds and documents. Compliance documentation, processing register, list of processors, data reversibility, digital accessibility: we provide them for your tenders, because we know that without them the file does not move. The full framework is described on our trust and compliance page, and we are available for a presentation to your departments.

Related articles