Insights — Sovereignty and AI
Five questions, ten minutes: a partner or a reseller.
The applied-AI market is crowded with successful demonstrations. What distinguishes a partner is not the model quoted in the brochure, but the ability to answer in writing, precisely, about your data, your costs and your freedom to leave.
01 — The problem
The question is not "which model", but "who decides what".
Two providers can offer the same model, the same lead time and the same price, and not sell you the same thing at all: one leaves you owning your index, your logs and your calls, the other installs you on a platform you cannot leave without losing everything.
That difference cannot be seen in a demonstration — it shows in five written answers, and it can be checked in a single meeting.
01The five questions that sort in ten minutes
They are about neither technology nor performance: they are about control. Ask them in writing, and note the form of the answer as much as its content.
- Who owns the document index, and in which format is it exportable?
- Where do queries run, and where are the logs stored?
- Which subcontractors are involved, and which of them see your data?
- What happens the day we want to change supplier, or tool?
- How do you measure that an answer is right, and how often?
02The answers that must raise a red flag
Three answers close the discussion. "The index stays with us": you are renting your own knowledge. "We will look at that later", about data egress: the cost of switching will be your problem, not theirs. And on accuracy: no figure, no measurement method — that is a demonstration, not a tool.
Conversely, a serious provider will talk to you about constraints. They will say what they cannot do, where the limits of their use case are, and what they refuse to put into production. Uncomfortable in a meeting, reassuring three months later.
03Our own checklist, applied to ourselves
We answer the same questions. The index we build belongs to you: documents, vectors and logs stay with you, readable without us. Hosting is in France or in Europe, and we can say when a provider does not guarantee it. Models go through a gateway: changing engine does not require rewriting the application.
Finally, we decline a mission when we cannot answer the fifth question. An AI project whose accuracy cannot be measured is not a project, it is an expense.
02 — Common questions
What we are asked before signing
Is a European provider enough to be compliant?
Not automatically: what matters is the actual processing — where queries run, where logs are stored, who can access them and under which rules. A European provider that processes your data outside Europe offers no additional guarantee.
How do we verify these claims?
By asking for a written document, then testing it. A hosting commitment is verified by looking at the actual addresses of the API endpoints used by the application. That is what we show during an audit.
Should we require a model hosted on our own premises?
That depends on how sensitive the data is. For a strategic internal document base, yes. For everyday use already covered by a solid contract, a mid-sized model hosted in Europe can be enough — provided you know how to switch.
Do you want your AI quotes put under the microscope?
Architecture and contract review: we look at what is actually described, what is missing, and what it will cost to change supplier.