AI agents · THE LAB AI Work Audit™
AI agent approval checklist: where must a human say yes?
AI agents do not just draft; they act — sending, booking, filing, replying, updating records. The question for any business running one is simple to ask and hard to answer: at which steps must a person approve before the agent proceeds?
Use the checklist below on any AI agent or automation currently acting on your behalf. It is deliberately short. If you cannot answer a question, that is the finding. If you can answer it but the answer is uncomfortable, that is also the finding.
THE LAB AI Work Audit™ applies this checklist — and the rest of the 11-Gate Assessment — to one real agent workflow in a 60–90 minute session, and returns named approval gates, a verification checklist, and three corrections in priority order.
The checklist
1. Can you list every action the agent can take?
Not the actions it usually takes — every action it is permitted to take, including ones it has never taken yet.
2. Which actions reach outside the business?
Any step that emails a customer, updates a shared system, books a commitment, or spends money should be marked. These are the candidates for approval gates.
3. What does the agent do when information is missing?
Does it stop and ask, proceed with a default, or infer? Each is a design decision. Do you know which one was made?
4. Where is a person required to approve, and do they actually do so?
An approval step that is always clicked through is not an approval step. Identify which gates are real.
5. What does the approver see?
Only the proposed action, or also the inputs and reasoning behind it? An approver who cannot see the basis cannot really approve.
6. Can you reconstruct a specific action later?
If a customer asks why the agent did something, can you show what it saw, what it decided, and who approved?
7. Who is told when it fails, and how fast?
Name the person, the signal, and the expected time. If any of the three is missing, the correction path does not exist yet.
8. Can you stop it?
Who has the ability to pause the agent today, and does that person know they have it?
What the audit looks at
One workflow, examined end to end in a 60–90 minute session.
- One agent or automation that acts on your behalf: sends, books, updates, files, replies.
- The full action list, including permitted-but-rare actions.
- Where approval gates exist on paper and whether they function in practice.
- What the approver sees, and whether that is enough to approve responsibly.
- Stop, notify, and correct: the three controls every acting agent needs.
You receive written findings within 48 hours: AI Workflow Map, Human Approval Gates, Verification Checklist, Three Priority Corrections, and a 30-Day Action Target. See the full method.
Questions
- Does every agent action need human approval?
- No. Requiring approval everywhere is how approval gates become click-through. The goal is to place real approval where consequence is high and lighter verification where it is low.
- We use an off-the-shelf agent product. Is the checklist still relevant?
- Yes. The checklist is about how the agent operates in your workflow, not about the vendor. Where behavior is outside your control, the corrections focus on your approval and monitoring steps.
- Is this a compliance certification or legal advice?
- No. The audit is an operational review of one workflow. It is not legal advice, a compliance certification, a cybersecurity penetration test, or a guarantee that risk has been eliminated.
- Do you need access to our systems or client data?
- No. The audit is a structured walkthrough with the people who run the work. We do not ask for logins, and we ask you not to share credentials or sensitive customer records.
Related audits
Founding Pilot · $500
Find out what you're looking at.
One AI-assisted workflow, a 60–90 minute walkthrough, and concise written findings within 48 hours: workflow map, approval gates, verification checklist, three priority corrections, one 30-day target.