Where it earns its keep
Search over your own corpus. A training organisation's team spends its day finding the right convention, the right procedure, the right past answer to a learner's question. A model that has read those documents, and only those, answers with the passage and the source, in seconds, and the human checks it. This is the use we ship most, because it is the one where the failure mode is a wrong citation the reader can see, not a wrong decision nobody sees.
First drafts the human edits. A convocation, a reply to a routine request, the summary of a file before an instruction meeting: a model writes the draft from the record, a person reads and sends. The time saved is real, minutes per item, hours per week, and the responsibility stays where it was.
Sorting what comes in. Requests arrive by mail, by form, by phone note; a model reads each and routes it to the desk that can answer, with a confidence the desk can see. When it is unsure it says so and a person sorts. That is the whole of the CAF du Var portal's routing problem, and a model does it better than a list of keywords.
Where it must not decide
Anything that decides for a citizen or a member. Eligibility, a grant, a sanction, an application's fate: a model may prepare the file and flag what is missing, and a named person decides, with the reasons written down. This is not caution for its own sake: it is what the law and an auditor will ask, and a tool that cannot show the human in the loop will not pass the review that precedes a public order.
Anything you cannot explain afterwards. A ranking of ideas on a participatory platform, a score on a learner, a priority on a request: if the answer to « why this one » is « the model said so », do not ship it. We build the explanation first, the rule, the fields it read, the threshold, and the model second, or not at all.
How we host it
The models run on European infrastructure, Scaleway's model service in Paris, and no client data is ever used to train one. Each use is a line in the processing register delivered with the instance: what goes in, what comes out, how long it is kept. The client's instance is isolated like everything else on Agora; the model sees one client's documents and nothing of another's.
This is also why we do not build on a US consumer API by default. Not out of principle, but because the buyer's data-protection officer will ask where the prompt went, and « to a server in Paris that keeps nothing » is an answer that ends the meeting.
What it costs, honestly
A search over a corpus is a few weeks of work and a few euros a month to run. An assistant inside a back office is a project: the model is the easy part, the data it may read and the screens it appears on are the work. A fine-tuned model of your own is almost never worth it for a company of your size, and when someone proposes one, the question to ask is what the general model got wrong that made it necessary.
If the honest answer to « where would AI help us » is « nowhere yet », that is an answer too. We have given it.