AI Agent Accountability Checklist

Use this checklist before deploying an AI agent that can take actions in real business systems — or to review an agent already in operation. Work through each section. Any unanswered item is a governance gap.

The IT Club AI Accountability Chain

Every AI agent deployment should have clear, documented answers to eight questions: Owner (who owns the outcome?), Authority (what can it do?), Boundary (what must it never do?), Approval (which actions need a human?), Monitor (who watches it?), Record (are actions logged?), Stop (who can disable it?), Review (who investigates incidents?).

Ownership

• □ Named business owner — who owns the outcome this agent produces
• □ Named technical owner — who manages the deployment and configuration
• □ Named risk owner — who is accountable if something goes wrong

Purpose and Boundaries

• □ Defined objective — what the agent is for
• □ Defined success criteria — how you know it is working correctly
• □ Defined prohibited actions — what it must never do

Identity

• □ Dedicated identity where practical — not a shared account or employee credentials
• □ No shared administrator account
• □ Credentials are revocable
• □ Separate test and production access

Permissions

• □ Least privilege — only the access needed for the defined task
• □ All systems the agent can reach are documented
• □ External actions are limited and understood
• □ Spending limits applied where relevant
• □ Read-only scope used unless write access is specifically required

Human Approval

• □ High-impact actions require human approval before execution
• □ Exceptions escalate to a named person
• □ Approval gates are implemented in the workflow — not only in the prompt
• □ The approval process has been tested

Monitoring

• □ Agent actions are logged
• □ Alerts exist for unexpected or out-of-scope behaviour
• □ Someone reviews agent behaviour regularly — not just at deployment
• □ Logs are recoverable and retained appropriately

Supplier and Legal

• □ AI and platform provider terms reviewed
• □ Data-processing obligations understood
• □ Insurance position considered
• □ Compliance impact assessed for the relevant regulatory context

Kill Mechanism

• □ The agent can be disabled quickly
• □ Credentials can be revoked
• □ The process has been tested before it is needed
• □ The person authorised to stop it is identified and available

Incident Response

• □ Response owner is identified
• □ Evidence retention process is understood
• □ Escalation process is defined
• □ AI-agent incidents are included in the existing incident response plan

Review

• □ Permissions reviewed regularly — not just at initial deployment
• □ Operational Heartbeat review scheduled
• □ Supplier changes and regulatory developments monitored
• □ Incidents and near-misses reviewed and fed back into controls