Self-hosted, and demonstrable
Some data is not allowed to leave. Patient records, public tender documents, contracts under a confidentiality clause. The tools we use install on your own servers, and what matters after that is not claiming it but being able to prove it.
The word sovereignty is made to cover almost anything these days. It appears in specifications far more often than it describes a technical reality, and almost no European organisation is genuinely independent of an American supplier. What companies actually want is more modest and more verifiable: knowing where their data lives, under what legal basis, and who can reach it.
Where the data lives, and under what rule
That is what this page is about. The flow engines we use are open, n8n first among them, and they install on a machine you control. An open language model, Mistral or Llama, can run on your own servers where sensitivity demands it, or with a European host where it does not. The two combine: an innocuous question goes out to a shared service, a question containing a patient’s name stays inside your walls.
Less capable, and we say so beforehand
That choice has a cost and we say so beforehand, not afterwards. A self-hosted model needs a machine, its updates and its monitoring, and at equal budget it will perform less well than an online service. For many internal uses, the gap goes unnoticed. For others, it shows immediately. The trade-off is made use by use, never in one block.
What we establish in writing
- Where each piece of data lives, on which machine and in which country.
- What leaves the company, what never leaves, and the rule that settles it.
- The legal basis for the processing, and its mention in your privacy policy.
- Retention periods, applied and not merely declared.
- Who has access to what, with the technical rights to match.
- A consultable trace, producible for an inspection or a tender.
What the European regulation requires
The European regulation on artificial intelligence comes into force in stages. It classes systems by risk and imposes transparency obligations, in particular where a system addresses a person or contributes to a decision concerning them. An assistant answering your teams about internal procedures is not in the same category as a system sorting job applications. The first question to settle is therefore the classification, before any technical discussion.
Two practical consequences that we apply by default. When an automated system talks to a person, it says so. And when a decision commits someone, a human approves it and their name appears in the trace.
We are neither lawyers nor data protection officers. We build so that your own counsel can do their work: a readable architecture, named processing operations, usable traces. The opposite, compliance retrofitted onto an opaque system, costs several times as much.
If you do not yet know what processing you have, the audit starts by taking stock.