AI transparency
What the AI in the ADEXMA platform does, which providers run it, and what it never does. Published under Article 50 of the EU AI Act and because our clients' data-protection officers ask.
Last updated: 2026-08-16 · Version 1.0
You are talking to an AI
1. What the AI does
- Answers questions in a workspace, grounded on that workspace's own data — dashboards, uploaded documents, benchmarks — through a retrieval step; answers cite their sources and abstain when nothing supports them.
- Drafts analyses, dossiers and follow-up notes for the human advisor to review.
- Extracts text from uploaded documents (OCR) so they can be indexed and cited.
- Classifies and enriches business contacts in ADEXMA's own relationship management.
2. What it never does
- It does not make decisions with legal or similarly significant effects on you or anyone else. Recommendations are decision-support; humans decide.
- It is not used for the purposes Annex III of the EU AI Act lists as high-risk — recruitment or worker management, creditworthiness, education, essential services, biometrics, law enforcement, migration or justice — and clients agree not to use it for those purposes.
- It does not learn from your data: no client input or output is used to train or fine-tune any model, by ADEXMA or by the model providers under the commercial terms we hold with them.
- It does not read identity documents into its knowledge index; clients are asked to keep passports, ID cards and special-category data out of it.
- It does not reach across workspaces: each engagement's data is isolated by database policy, and the models see only what the requesting user may see.
3. Which models run, and where
| Lane | Inference provider | Location | Notes |
|---|---|---|---|
| US commercial | Anthropic (Claude) by default; Microsoft Azure AI Foundry or Google Gemini as configured | United States | Providers' commercial terms — no training on inputs/outputs |
| EU commercial | Scaleway Generative APIs (Mistral models) or Mistral AI directly | France (Paris) | Platform-enforced: an EU-resident engagement is answered only by an EU-resident model — or not at all |
| Document OCR | Mistral AI (EU engagements); Microsoft Azure AI Foundry (US engagements) | France / United States | EU engagements holding personal data cannot be routed through US OCR |
| Higher compliance tiers | Client-operated accredited boundaries only (for example the client's own AWS GovCloud) | Client's boundary | No ADEXMA-operated model may answer; the platform refuses rather than falls back |
The full provider list, with transfer mechanisms and assurance, is on the Sub-processors page.
4. What is recorded
For each AI interaction the platform records metadata — which advisor, which model and provider, timing, token counts, outcome — in an audit table that does not contain the question or the answer. The conversation itself is stored in your workspace under your engagement's retention rules and is visible only to its members. ADEXMA staff access conversation content only for support at your request, for security investigation, or as the law requires — and every such access is logged.
5. Limitations
Language models can be confidently wrong, can miss context that is not in the workspace, and reflect the data they were grounded on. The platform reduces this by grounding answers on your own numbers first, by citing sources, and by abstaining when it cannot; it does not eliminate it. Verify before you act, especially on prices, contracts and anything a third party will rely on.
6. Questions and concerns
Ask the human advisor on your engagement, or write to clerouge@adexma.com. If you believe an AI output has affected you unfairly, tell us — a person will review it.