AI Agent Governance Checklist

Authority

List every action the agent can perform. Separate read-only access, recommendations, draft creation, external communication, system updates, approvals, financial actions, and irreversible changes. Give the first release the minimum authority it needs.

Human review

Define which actions always require approval and which can proceed automatically under specific conditions. Document the reviewer role, information presented for review, and what happens when a reviewer rejects or edits the proposed action.

Data access

Identify the data sources the agent may access, the fields it actually needs, and whether sensitive categories require additional controls. Use least-privilege credentials and avoid broad access simply because it is technically convenient.

Knowledge sources

Define which repositories are authoritative, who owns them, how stale content is identified, and what the agent should do when sources conflict or no approved answer exists.

Evaluation

Create test cases for normal work, edge cases, ambiguous inputs, missing data, adversarial or malformed inputs, integration failures, and policy-sensitive scenarios. Measure appropriate behavior and escalation, not only task completion.

Logging and auditability

Record enough information to understand what the system received, what tools or sources it used, what action it proposed or took, what rules were applied, and whether a person reviewed the result. Avoid logging sensitive content that is not necessary for the audit purpose.

Failure behavior

Define safe outcomes when a model, API, credential, data source, or downstream system fails. The agent should be able to stop, retry within limits, queue work, or escalate rather than improvising around missing dependencies.

Change control

Treat prompt changes, model changes, tool permissions, routing rules, knowledge-source changes, and integration changes as production changes. Define testing and approval requirements appropriate to their impact.

Operational ownership

Name the business and technical owners responsible for quality, access, incidents, source maintenance, exception policy, and periodic review. Governance needs an operating owner after launch.