The questions from an auditor or regulator tend to be the same. Which decisions does AI touch? Who authorised it? What data does it use? How was it tested? What happens when it's wrong? Can a person step in? A chat rollout struggles with most of these because the decisions happen in people's heads after they read the output.
How a decision service answers them
- Which decisions: a register of decision services, each with a named owner and a written list of allowed actions.
- Who authorised it: the policy version and the approval that released it.
- What data: the state definition for each question, retrieved under the caller's permissions.
- How it was tested: a versioned labelled evaluation set and a threshold report for every model or policy release.
- What happens when wrong: the review route, the error log and the stop rule.
- Human oversight: the reversibility test, which sends irreversible actions to a person every time.
Be precise about the data boundary
In the proposed Microsoft design, everything runs in your tenant except the call to Jev, which sends the state for each question to TypeSafe's API. That call needs a data protection impact assessment, a supplier assessment and a clear retention position, the same as any external processor. Send only the fields each question needs.
Stay clear of the high-risk end
Data protection law restricts decisions made solely by automated means where they have legal or similarly significant effects on people. The EU AI Act places uses such as employment and access to essential public services in its high-risk category. Keeping a person on consequential decisions, and starting with reversible operational ones, keeps early pilots away from those lines. Take legal advice on your specific case; this is a design principle, not legal guidance.
What to hand over
For each decision service: the register entry, the state definition, question and policy versions, the latest evaluation report, the review statistics and the incident log. If you can produce those on request, the conversation with an auditor is about your controls rather than about AI in general.