Control objective

Record the responsible-AI decision file for one governed Foundry agent, add a human approval gate to one consequential tool, and confirm approved and rejected tool-call paths.

Why it matters

Problem. A governed agent can pass deployment checks and still lack a clear record of why it is acceptable, what users must be told, and who can stop a consequential action.

Solution. The module records those decisions and makes one write or irreversible tool wait for a human decision before execution.

EU AI Act. Supports Articles 5, 6 and Annex III, 9, 13, 14, 26(2), 27, and 50. Engineering mapping, not legal advice.

Architecture at a glance

Decision file shape

The module keeps three records:

• ai-system-inventory-entry.json for use case, risk class, Article 5, and Article 50.
• impact-assessment.md for harms, affected groups, data, and human role.
• transparency-note.md for user notice, limits, and review cadence.

The schema rejects drift before runtime deployment.

Human oversight gate

The selected tool must be write or irreversible.

The gate returns awaiting_approval, then resumes with the approver's decision:

• approved calls the configured tool route once.
• rejected returns without tool execution.

The requester and approver come from App Service authentication headers and claims, not from the request body.

No rejection can become a tool call.

Implementation tradeoffs

What preflight checks

Preflight rejects:

• unresolved __REQUIRED_*__ decisions;
• target scope other than one-approved-foundry-agent;
• invalid enum values or a prohibited risk class;
• a blocked Article 5 screen;
• oversight settings that allow rejected execution.

It also states that no read-only deployment preview exists for the approval code.

Confirm the result

Operating state