01Introduce AI agents
Human Approval Checklist for AI Agents
A human-in-the-loop checklist for AI agents: which actions need approval, who approves, what they check, and a 60-second check for AI drafts.
- Topic
- Introduce AI agents
- Written by
- AI Ground Rules
- Last reviewed
- Sources
- 5 checked sources
A person should approve an AI agent’s action before it happens when the action reaches someone outside the business, changes something other people rely on, spends money, grants access, or would be hard to undo. Reading information and saving drafts can run without approval of each action, but only after the data decision owner has approved the exact tool, account, data types and task. That first approval covers what the agent may touch; approving an action covers one thing it is about to do. Sensitive-data and client restrictions still apply. This is what “human-in-the-loop” means in practice: for each risky action, name who approves it, decide what they check before they say yes, and agree what happens when they are away. The checklist below walks through those decisions with fictional examples.
An AI agent here means an AI tool that can take actions in connected systems, such as sending email, updating a client record or booking a meeting, rather than only suggesting text.
Before you approve: the 60-second check
Use this every time you approve AI-drafted work going to a client: an email, a letter, a report or a set of figures.
- Right client, all the way through. The name, company and details are this client’s, in every paragraph, the subject line and any attachment.
- Figures checked against the source. Every number, date and amount matches the original document, not just the draft.
- No other client’s data. Nothing from another client’s file, email or earlier draft has slipped in.
- Claims you can stand behind. You would say every statement yourself. Remove anything you cannot check.
- AI use declared where needed. If your policy or the client’s agreement says to mention AI use, it is mentioned.
If any point fails, do not approve. Fix it, or send it back.
No agents yet? The same check applies to any AI draft going to a client. A person checks AI-assisted work before it goes out, and staff own the output. Your AI Acceptable Use Policy sets the wider rules.
Step 1: Sort each action
List every action the agent can take. Your AI inventory template is the place to keep that list. Put each action into one of three groups.
| Group | Meaning | Typical actions |
|---|---|---|
| Allowed within a limit | The agent can do this on its own, inside a stated scope | Read a named folder, summarize, classify, save a draft |
| Approval every time | The agent prepares the action and waits for a person | Send an external email, update a client record, publish, book on a client’s calendar |
| Prohibited | The agent must not do this at all | Delete data, make payments, change permissions, connect a new system |
This is a suggested starting point. Some businesses will allow more once an agent has a track record. Move an action to a looser group only after a deliberate decision, and record it.
Step 2: Ask five questions about each action
If the answer to any question is “yes,” require approval or prohibit the action.
- Does it leave the business? Messages, documents or data going to a client, supplier or the public.
- Does it change or remove something other people rely on? Client records, shared files, prices or settings. A new draft in the agent’s own folder does not count.
- Does it commit money or obligations? Purchases, refunds, discounts, bookings or agreements.
- Does it change who can access what? Sharing a folder, adding a user, connecting another app.
- Would it be hard to undo, or would a mistake affect a person? Anything touching someone’s job, account, eligibility or personal information.
Make sure the approval is real
A line in the agent’s instructions saying “ask before sending” is not an approval gate. Agents can misread instructions. Content they read, such as an email or web page, can also try to tell them to do something else. The approval has to be a step in the tool itself that the agent cannot skip.
OWASP’s guidance on excessive agency in AI systems says to “require a human to approve high-impact actions before they are taken.” It also advises giving an agent only the tools and permissions it needs. (Source: OWASP LLM06:2025 Excessive Agency, checked 2026-11-14.)
Before going live, test with fictional data:
- An allowed action works.
- An action needing approval waits, and nothing happens if approval is refused.
- An action needing approval does not go ahead when nobody responds.
- A prohibited action is blocked.
- A document asking the agent to ignore its rules does not change what it can do.
- The pause procedure works.
If the tool cannot enforce a required approval, remove that permission or keep the agent switched off.
Where the approval switch is
These are the approval steps in two tools small firms often use, as described in each vendor’s documentation on 2026-11-14. Menus change, so check the current screen.
| Tool | Where the approval step is | What to set |
|---|---|---|
| Microsoft Copilot Studio agent flows | In the flow designer, select Insert a new action, then Request for information under Human review. The flow pauses and emails the request through Outlook until someone responds. | Add a Yes/No input for the decision, and make the next step act only on “Yes.” The first person to respond decides, so list only people allowed to approve. |
| Zapier | Add a Human in the Loop step with the Request Approval action. The Zap pauses until a reviewer acts. Available on Professional, Team and Enterprise plans, not Free. | Set Action if reviewer declines to Stop run. Set On timeout to End run. The other option, “Skip and continue,” lets the Zap carry on with no approval. |
(Sources: Microsoft Learn, Copilot Studio request for information; Zapier Help, Human in the Loop. Both checked 2026-11-14.)
For other tools, ask the vendor or your IT provider: “Where is the approval step, what happens if nobody responds, and where is each approval recorded?” If there is no clear answer, treat the action as prohibited.
Step 3: Decide who can approve
The approver needs both the authority and the knowledge to judge the action.
- Name the approver and one backup. In a small firm, that is often you and one backup.
- Match the approver to the risk. A team member can approve a routine reply. A partner approves a fee change.
- Do not let the person who asked the agent to do the work be the only check on high-impact actions.
- Absent approver. If the approver is away, requests go to the backup. If both are away, the action stays on hold. No response is not approval.
Step 4: Decide what the approver sees
An approval is only as good as the information behind it. Before deciding, the approver should see:
| What to show | Why |
|---|---|
| The exact action | “Send this email” or “change this field,” not a summary |
| The recipient or record | Who or what is affected |
| The full content | The actual message, values or file |
| The source information | What the agent relied on, so the approver can check it |
| Any cost or commitment | Amount, currency and limit |
| What the agent was asked to do | The original request, to spot drift |
If the approval screen does not show these, the approver cannot judge the action. Change the setup so the request includes them, or keep that action prohibited.
Step 5: Keep approvals meaningful
Approvals fail quietly when people stop reading them.
- Watch for approval fatigue. If someone approves dozens of requests a day without changes, they may be rubber-stamping. Reduce the volume. Move safe, repeated actions to “allowed within a limit,” or prohibit risky ones.
- Approvals expire. An approval covers one action, as shown. If the content, recipient or amount changes afterwards, approve it again. Set a time limit on waiting requests, and make sure a request that runs out ends with no action.
- Log each approval. Record who approved, when, and what they saw. Check that the tool keeps this record, or keep it yourself.
Step 6: Plan how to pause and escalate
Agree these before the agent goes live:
- Who can pause it. You and one backup, each with the access to do it.
- How to pause it. The exact steps, written down and tested.
- Stop conditions. For example, an unexpected recipient, access to data outside its scope, repeated errors, or a limit reached.
- What to check after pausing. Any queued or half-finished actions.
- Who to tell. The data decision owner or backup, following your incident steps.
- Who approves restarting. The agent’s owner fixes the cause and tests the fix. The data decision owner then decides whether the agent comes back on. It stays off until the cause is fixed.
Being able to see what an agent did helps here. Microsoft’s guidance on agents says every action “should be traceable to a specific agent, in a specific context, at a specific moment in time.” (Source: Microsoft Learn, governing agents guidance, checked 2026-11-14.) That guidance is written for large organisations, but the point applies to any firm. Ask your vendor or IT provider for the agent’s activity log and how long it is kept.
Fictional examples
| Scenario | Action | Group | Approver and what they see |
|---|---|---|---|
| At Lantern Talent, an agent drafts follow-up emails to candidates after interviews | Send email to candidate | Approval every time | The recruiter sees the full email, recipient and interview notes used. The agent sends only after the recruiter approves |
| At Harbor Accounts, an agent updates client contact details from incoming emails | Change a client record | Approval every time | The office lead sees the old value, new value and source email |
| At Northgate Advisory, an agent summarizes meeting notes into a project folder | Save a summary | Allowed within a limit | No approval. Limited to that project’s folder. A person checks it before sharing |
| At Fernway Plumbing, an agent offers to issue a refund to a customer who complained | Spend money | Prohibited | Not allowed. The agent drafts a note for the account manager instead |
| At Calder & Rowe Solicitors, an agent wants to share a folder with a new client contact | Grant access | Prohibited | Not allowed. A person shares the folder through the normal process |
Record the decision
Record the result for each agent in your AI inventory: the action groups, approver and backup, the approval step, the stop procedure and the test results. Review it every quarter and whenever the agent gains a new connection, data source or action.
Questions
Is “ask me before sending” in the agent’s instructions enough? No. Instructions can be misread or overridden by content the agent reads. Use an approval step built into the tool, and test that it holds.
02Evidence
Sources
- OWASP GenAI Security Project, LLM06:2025 Excessive AgencyChecked
- Microsoft Learn, How do we start governing agents without slowing innovation?Checked
- Microsoft Learn, Request information from human review in agent flows (Copilot Studio)Checked
- Zapier Help, Use Human in the Loop to pause Zap workflows pending human reviewChecked
- Zapier Help, Request approval to keep your workflow running with Human in the LoopChecked