How Ansvar AI actually operates.
Auditable AI requires honest operating posture. Here's how we handle model access, data retention, training, and the citation pipeline that sits between the model and its sources.
Model access (BYOK)
Inference runs through your own Anthropic, OpenAI, or Azure credentials at the Premium, Team, and Company tiers. Model traffic flows directly from your client to your provider — Ansvar AI does not hold or proxy your model keys. For EU-only posture, use Mistral La Plateforme.
Training on customer data
We do not train, fine-tune, or evaluate any model on customer data. Uploaded documents, query text, and citation metadata are used only to serve the user's own workflow. BYOK providers' retention settings apply per their own terms.
Citation validation pipeline
Every citation that leaves the gateway passes through a deterministic (non-model) validation pipeline: article numbers are resolved against corpus indexes, paragraph anchors are checked, URLs are verified to return the cited text. Failed citations are shown as unavailable, not silently dropped or fallen back to AI guesswork.
Retrieval transparency
Open legal connectors are published under Apache 2.0 where source licensing permits — inspect them at github.com/Ansvar-Systems. You can audit how we reach each source.
Data residency
All infrastructure runs on Hetzner in the EU. No customer data egresses EU borders unless a non-EU BYOK provider is explicitly configured. The audit trail persists in-region with per-workspace envelope keys.
Model versions and determinism
Planner and summariser model versions are pinned per-workspace and rolled forward with notification. Retrieval, cross-referencing, and citation validation are deterministic — the same question returns the same sources across runs.
Questions? Security overview covers procurement-grade architecture. For a written answer to a specific question, contact team@ansvar.eu.