Your data
To describe your AI systems to us, you have to tell us about them. Here is exactly where that text goes and which companies see it — named, not summarised.
Documents you upload
They are processed on our own machines and are not sent to any third party.
To search a document we first convert it into a mathematical form. Many products send your text to an outside service to do that. We do not. The model that reads your documents runs inside our own deployment, so the contents never leave it.
That model is multilingual-e5-large, published by Microsoft Research under an MIT
licence. We run it ourselves. Microsoft receives nothing.
Your questions, and the answers
Answering a question means sending it to a language model, because a model has to read the question in order to answer it. There is no way around that, so the question becomes which company reads it.
| What | Who processes it | Where |
|---|---|---|
| Documents you upload | Nobody outside | our own machines |
| Your questions and our answers | Mistral | France |
| If Mistral is unavailable | EU region |
Mistral is a French company and it is the default for anything containing your text. Google is the fallback that exists so the service keeps working if Mistral has an outage. If both are unavailable, we stop and tell you, rather than quietly sending your text to a third company. That third company used to be American — we removed it on purpose, and we would rather tell you that than let you assume it is still there.
What we are changing
One part is not where we want it yet, and we would rather write it down than wait until it is fixed.
When you ask a question about the regulation, we search the legal text to find the relevant passages. That search currently uses an American service to convert your question into a searchable form. Your uploaded documents never go there — but a question you type does.
There is a second place this happens: the free assessment on our home page, where you describe your AI system before signing up. That description — up to 5,000 characters — is also converted by the same American service, for the same reason: to search our reference material for what applies to your case.
We are moving that step onto our own machines, using the same model that already handles your documents. It requires rebuilding our copy of the legal text first, which is work in progress rather than a plan.
Yes, a model writes the documents
A sophisticated buyer works this out quickly, so we will say it first: the reasoning and the documentation are produced by a language model. Three things make that safe to rely on.
Every claim is traceable
Statements are tied to a specific Article, Annex or Recital. You can open the regulation and check. A model that has to cite is constrained by what it cites.
We never assert compliance
The output describes what the regulation requires and what your answers say. It does not conclude that you comply — nobody can conclude that on your behalf.
You review it
It is your documentation, submitted under your name. We produce the draft and the reasoning; the judgement remains yours, and the record shows how each conclusion was reached.
What we have not built yet
AIMS Agent is not yet generally available, so this page describes how the system is designed and how it behaves today — not a service you can currently sign up to. When it is available, this page will say so and will list every processor and location in force at that time.
If your security or procurement team needs this in a different form — a questionnaire, a processing agreement, a data flow diagram — write to us and you will get an answer from one of the two people who built it.