pq
Free Checklist · One Page, Zero Excuses

The AI governance checklist

Perpetual Quest · 2026

The Operator’s
AI Governance
Checklist

Before your agents touch production. Outputs are easy — receipts are rare.

Audit logs, human approval gates, permission scoping, the five governance pillars — the baseline every agent deployment should ship with, before compliance asks what your agents did last month.

Instant access · Also sent to your email · No spam

Dear Operator,

Here’s a question worth sitting with: “What did your AI agents actually do last month?”If your compliance team — or your board, or a regulator, or a major customer’s security review — asked that today, could you answer with documentation, or with a shrug?

Most AI platforms give you outputs. Grown-up deployments give you outputs with a record. The checklist below is the baseline we build into every system we deploy — and run in our own businesses. Print it, tape it to the wall, and don’t let any agent touch production until the boxes are checked.

The checklist

1 · Audit logs — every action, reconstructible

  • Every agent action is logged: what it did, when, what triggered it, what data it touched.
  • Logs are queryable after the fact — "show me everything agent X did on Tuesday" is a 30-second answer.
  • Log retention matches your industry's record-keeping requirements.

2 · Human approval gates on high-stakes operations

  • You have a written list of operations that always pause for human sign-off (payments, customer-facing sends, and other high-stakes actions).
  • Every hold-for-approval surfaces an approver packet: proposed action, supporting evidence, agent reasoning summary, confidence or uncertainty, policy triggered, potential downside, reversibility, recommended decision, and similar prior outcomes.
  • Approvals are recorded — who approved, when, and the packet they reviewed. An Approve button without that context is theater.
  • The gate list is reviewed when scope expands, not just at launch.
  • If reviewers routinely approve large batches of nearly identical actions without using the packet, the gate is a human bottleneck — redesign autonomy zones, don't add more rubber stamps.

How we draw those zones: Human-in-the-Loop — when to let agents act vs require approval →

3 · Permission scoping — least privilege by default

  • Each agent has exactly the access it needs for its workflow. No more.
  • Read and write permissions are separated — an agent can read a system without write access to it.
  • Every write tool is individually approved — no blanket "can write" grant.
  • Destructive and irreversible actions have separate controls beyond ordinary write access.
  • Permission increases require evidence and approval — scope expands only with a recorded rationale and sign-off.
  • No agent runs on a human's personal credentials or a shared admin account.
  • Someone can produce the access list per agent on demand — and revoke it in minutes.

4 · Oversight loop — trust earned with evidence

  • A defined human review loop covers the first 30 days of any new agent in production.
  • Edge cases route to a named person, not a queue nobody owns.
  • Oversight health is tracked: approval volume, average review time, override rate, rubber-stamp rate (approvals with negligible review time or no packet engagement), reviewer disagreement, and actions automatically approved after repeated success.
  • Sampling rates and progressive autonomy step up only when these metrics show real review — not on a calendar alone.

5 · Transparency & escalation — answers on demand

  • You can produce a plain-English summary of what your agents do, for a customer or auditor, without a scramble.
  • There's a documented kill switch: who can pause an agent, and how fast.
  • Incidents (wrong output reached a human or a system) are logged with cause and correction.

6 · Agent change control — no silent self-modification

  • Action logs record what agents do. Change control governs how the agent system itself may change — they are not the same.
  • Every prompt, tool, policy, model, and permission set is versioned.
  • Each change records who proposed it and who approved it before it reaches production.
  • Regression evaluations run on a fixed eval set before any material change is deployed.
  • Changes use staged rollouts — not a single flip to 100% traffic.
  • New and previous versions are compared before and after promote.
  • Threshold breaches trigger automatic rollback to the last-known-good version.
  • Production agents cannot change their own permissions or governing instructions.
  • Feedback collection is separated from deployment authority — suggestions never auto-deploy.

The five pillars behind the checklist

The sections above roll up into five pillars we score every deployment against — data, access, oversight, transparency, compliance. Agent change control is how that estate stays trustworthy after go-live: the system itself cannot silently rewrite what it is allowed to do. In Forge, the live governance-health score ships with the deployment as the baseline dashboard and go-live score. In Run, that same score becomes an operating discipline: drift monitoring, change review, and quarterly re-scoring so the dashboard keeps meaning something. Whether you work with us or not, score yourself against the five pillars quarterly. Drift is silent; the score isn’t.

Why this is an operator issue, not an IT issue

Governance sounds like overhead until the day it’s the only thing standing between you and a very bad week: a customer security review you can’t pass, an auditor question you can’t answer, an agent mistake you can’t reconstruct. The companies deploying AI durably aren’t the ones moving fastest — they’re the ones who can show their work. Receipts are rare. Be the operator who has them.

And the honest counterpoint: governance without deployment is just paperwork. Don’t let this checklist become the reason nothing ships. Build the baseline into the first deployment — it’s a week of work when designed in, and a quarter of pain when bolted on.

Who we are

Perpetual Quest deploys AI agent systems for mid-market operators. Every system we build ships with this governance layer as the foundation, not an add-on tier — audit logs, approval gates, permission scoping, agent change control, and live governance health scoring. The 28-item checklist is the shared standard; Forge includes the live score at go-live, and Run keeps that score current over time. Validated in our own operating businesses first.

Book the 30-minute call — we’ll review your governance posture, no pitch →

— Perpetual Quest
perpetualquest.com · [email protected]