ICMD Agent Governance Checklist (2026)

Use this checklist to operationalize AI-first leadership without creating a bureaucracy. Do an initial pass over a few weeks, then iterate.

1) Inventory
- List every agent/automation that can take actions (not just draft text).
- For each: name, version, owner, vendor/model (OpenAI/Anthropic/Google or self-hosted), and where it runs (Slack, web app, CI, CRM).
- Document systems it can read from and write to (GitHub, Stripe, Salesforce, Zendesk, production databases).
- Identify shadow agents: personal accounts, browser plugins, unofficial Zapier/Make workflows.

2) Assign accountability
- Assign one human DRI (directly responsible individual) per agent.
- Assign a security partner for high and very-high risk agents.
- Define on-call and escalation for agent-caused incidents (who gets paged and what the response path is).

3) Risk classify
- Classify each workflow: Low / Medium / High / Very High.
- High/Very High signals: money movement, customer commitments, legal/compliance exposure, production changes, identity/permissions.
- For each workflow, write down what failure means (wrong refund, incorrect email, broken deploy, leaked data, policy breach).

4) Guardrails
- Enforce least privilege: read-only by default; narrow write scopes; time-bound tokens where possible.
- Add thresholds and approvals for sensitive actions.
- Require dry-run or staging for production actions; use feature flags and rollback plans.
- Add human approval gates for irreversible actions (customer emails, merges to main, pricing changes).

5) Logging and auditability
- Log every agent action with: timestamp, agent/version, policy version, inputs/outputs, systems touched, trace ID, approval status.
- Store logs centrally (SIEM or log pipeline) with a clear retention policy.
- Make logs searchable by customer ID, order ID, pull request, or incident.

6) Metrics
- Track outcome metrics per workflow (CSAT, MTTR, change failure rate, forecast error).
- Track reliability metrics: override rate, exceptions per unit of work, rollback frequency.
- Set alert thresholds based on trend changes and incident risk.

7) Evaluation and change control (ongoing)
- Maintain a small eval set per agent (representative cases that match real work).
- Require an eval pass before increasing autonomy or shipping new policy versions.
- Version policies like code; publish changelogs; run postmortems on agent-caused incidents.

8) Culture and training (ongoing)
- Train teams on “draft vs decide”: AI can propose; humans own outcomes.
- Treat review as a skill with standards and examples.
- Maintain a safe channel for reporting risky automations; reward early escalation.

Definition of done
Every agent that can change state has: (1) a human owner, (2) risk classification, (3) least-privilege permissions, (4) action logging, and (5) at least one outcome KPI reviewed on a regular cadence.