// FIELD TOOL · AI AGENT GOVERNANCE
AI Agent Governance Readiness Checklist
Not a policy maturity survey. This is a project-level Go / No-Go scorecard you can bring into a launch review. Score every control 0 / 1 / 2 and require evidence for every high score.
0
No evidence
Verbal agreement, heroics or “we will add it later”
1
Partially verifiable
A control exists but coverage, ownership or testing is incomplete
2
Verified
Named owner, evidence, threshold and validation in test or real operation
Coverage
10 governance domains to expose launch blind spots
Organizational goals
Business outcome, accountable owner, boundaries and stop conditions
Data & access
Least privilege, separate identities, sensitive-data boundaries and credential handling
Model / agent risk
Hallucination, overreach, prompt injection, tool misuse and cascading failure
Human approval
Human gates for irreversible, high-value and external-facing actions
Auditability
Inputs, outputs, tool calls, approvals, versions and decision evidence
Evaluation
Offline evals, regression sets, launch thresholds and online quality monitoring
Cost
Per-task budgets, model routing, spend anomalies and stop-loss controls
Recovery
Timeouts, fallback, retries, idempotency, kill switch, rollback and takeover
Vendors & compliance
Data processing, residency, retention, model terms and exit plans
Launch gates
Convert critical controls into Go / No-Go criteria instead of post-launch debt
How to use
Score alone for 15 minutes. Reconcile differences for 30.
- Have the business owner, agent/platform owner and risk or security owner score independently.
- Discuss only disagreements. A high score must point to logs, policies, eval results or approval evidence.
- If any critical gate scores 0, treat it as a remediation item regardless of the aggregate score.
- Re-score after remediation and reuse the same sheet as the production review baseline.