An AI agent risk assessment should help a reviewer understand what the agent does and whether the controls match the risk. It should not be a generic questionnaire that treats every AI system the same. Agents require special attention because they can retrieve information, call tools, and act across systems.
Use the template as a structured record for intake, approval, and later audit evidence.
1. System overview
Capture agent name, owner, technical contact, business purpose, users, environment, launch status, vendor or model provider, repositories, and connected systems. Write the purpose in plain language: what workflow does the agent support?
2. Data access
List input data, retrieval sources, stored data, output destinations, personal data, confidential data, regulated records, and vendor retention. Include how permissions are enforced and whether the agent can access data beyond the user's normal rights.
3. Autonomy and tools
List each tool the agent can call, including read, write, send, execute, approve, or trigger permissions. Note which actions require human approval and which are fully automated. Include retry limits and failure handling.
4. Risk scenarios
Identify realistic harms: wrong action, data disclosure, prompt injection, excessive agency, user deception, biased recommendation, regulatory breach, security incident, or operational disruption. Rate severity and likelihood.
5. Controls
Map each material risk to controls: least privilege, human approval, output validation, retrieval filtering, monitoring, rate limits, redaction, incident response, and change review. Identify the control owner and evidence source.
6. Decision
Record risk tier, approval decision, reviewers, rationale, launch conditions, residual risk, review date, and exceptions. The assessment should be updated after material changes.
A good risk assessment does not slow every agent equally. It makes the important differences visible so teams can apply the right controls before production.