An AI agent that only drafts a reply is one kind of risk. An agent that can issue a refund, cancel an order, send an email or move money is another. The second kind can create a result you cannot take back, or cannot take back easily, before a person has even seen it.
The practical move is simple, and it should happen before the agent goes live: list every action it can take, mark the ones you cannot undo and require a human sign-off for those.
Start with a complete list, not a guess
People often describe an agent by its job title: "it handles support," "it books appointments," "it follows up with leads." That is not a list of actions.
Sit down with whoever configured the agent and write down each thing it is allowed to do. Be concrete. Say your support agent can answer a question, look up an order, start a return, issue a refund, cancel a subscription, change an address, send an email or create a credit. A scheduling agent might book, cancel, reschedule or message a customer. A sales agent might quote a price, apply a discount, send a proposal or update a record.
If the agent can use tools, include those too. A tool is an action with a shorter name.
Then, for each action, ask one question: if this goes wrong, can we fully reverse it, quickly, and with little harm to the customer or the business?
What "can't be undone" usually looks like
Some actions are easy to reverse. A wrong draft that a person still has to send is not live yet. A note added to an internal record can be edited.
Other actions travel farther:
- A refund that has already been issued
- A cancellation that has already taken effect
- An email, text or public reply a customer has already received
- A payment, transfer or charge that has already been submitted
- A promise, price or date that now lives in a customer's inbox
You do not need a perfect legal definition. You need an honest line: if we would be uncomfortable seeing this go out with no person in the loop, it belongs on the "sign-off required" list.
When in doubt, put it on that list. You can always loosen a rule later. You cannot unsend a message or reverse a payment with the same ease.
Require a human sign-off for irreversible actions
Once you have the list, change the agent's permissions to match it.
- Let the agent propose, not execute, anything you cannot undo. It can draft the refund, the cancellation, the email or the payment. A person approves it before it goes out.
- Name who can approve those actions. Not "someone on the team." A role, a backup and a way to reach them during the hours the agent runs.
- Set a time limit for review, and a default that is safe. If nobody reviews a request, the safe default is to wait, not to proceed.
- Log every irreversible action, including who approved it. When a customer asks what happened, you should be able to see the trail.
- Test the gate. Before customers meet the agent, try to make it issue a refund, send a message or cancel an order on its own. If it can, the permission is still too broad.
The point is not to slow every answer. Most questions do not need a sign-off. The point is to keep a person in the loop for the few actions that can create a mess.
Questions to ask whoever runs your agent
- Do we have a written list of every action this agent can take?
- Which of those actions can we not fully reverse, or not reverse quickly?
- For each irreversible action, does a person have to approve it before it happens?
- Who can approve those actions, and what happens if they are unavailable?
- When was the last time we tested whether the agent can bypass that sign-off?
If you cannot answer the first two questions, you do not yet know what the agent can do in your name. That is the risk. The list comes first; the sign-off comes next.
I work with businesses on putting AI agents to work safely, through training, pilots and safety checks. If your agent can already take actions for you, I'd like to hear which ones require a person to sign off, and how you drew that line. Reply and tell me how you handle it.
