A customer asks your AI agent to waive a fee that the company normally does not refund.
The agent understands the request. It can read the customer’s history, see the charge, and write a perfectly polite response. It may even be fairly sure that making an exception would calm things down.
But the fee is not its decision.
This is the point of an AI agent escalation policy. It tells the agent what to do when a routine task turns into something a person needs to judge.
The phrase sounds formal. The useful version is not. It is a clear answer to a very ordinary question: when the situation gets weird, who needs to see it and what do they need to know?
A good agent knows when the job has changed
Most agent demos stay on the happy path. The customer asks a familiar question, the information is complete, and the next step is obvious.
Real work is messier.
The customer asks for the owner. A supplier sends new bank details an hour before payment. Two records disagree about where an order is going. The request starts as a booking change and ends with someone explaining that the service failed at the worst possible time.
At some point, the job changes. It is no longer “process this routine request.” Now it is “make an exception,” “verify something suspicious,” or “repair a relationship.”
That is when the agent needs to stop trying to be helpful on its own.
Think about a good employee in their first few weeks. You do not expect them to bring you every question. You also do not want them improvising a refund policy because they feel bad for a customer. You want them to recognize the edge of their authority and arrive with the facts already organized.
An agent deserves the same clarity.
Begin with the moments that make your team pause
You probably already know where the awkward moments are.
Ask the people doing the work when they lean back from the screen and say, “I need to check this.” Do not ask for every theoretical risk. Ask for last month’s real examples.
The fee waiver is one. The employee can explain the standard terms, but only the owner can make the exception.
The supplier email is another. The message may look legitimate, but changing bank details before a payment carries enough downside that somebody needs to verify it through a trusted channel.
Then there is the muddier case. A customer says the delivery address is wrong. One system shows the old location, another shows the new one, and the email came from an address nobody recognizes. The agent does not need a better guess. It needs to stop making guesses.
These examples usually reveal the boundary faster than a long policy discussion. The agent has reached the edge when it lacks reliable facts, when the decision belongs to someone else, or when being wrong would be hard to undo.
You do not need five pages of categories. You need examples concrete enough that the agent can recognize the family resemblance.
Let it ask once, then stop digging
Sometimes one question clears things up.
“Can you send the order number?” may be all the agent needs to find the right record. “Are you asking to change the appointment time or cancel it?” can turn a vague message into a routine request.
The trouble starts when the agent keeps questioning, rephrasing, and nudging because it has been told to complete the task at all costs. Customers feel trapped. The agent gathers more words without getting closer to the truth.
A useful house rule is this: one sensible clarification is fine. If the answer still leaves the agent guessing, bring in a person.
The same goes for failed actions. If a payment update or account change does not work, repeated attempts can make the situation worse. Preserve what happened and stop. Persistence is valuable when following up on a missing document. It is dangerous when the action affects money, access, or a customer promise.
The handoff should save the human time
“Manual review required” is not a handoff. It is a shrug.
Imagine the owner receiving the fee-waiver case. They should not have to read a long chat, open three systems, and reconstruct why the agent gave up. The note can be brief:
Taylor is asking us to waive a $75 late-cancellation fee. The appointment was cancelled six hours before the scheduled time. Our standard terms apply the fee inside 24 hours. Taylor says a family emergency caused the cancellation and has asked for the owner. I acknowledged the request and said someone would reply by 3 p.m. I did not promise a refund.
That is a useful handoff. The person knows what is being requested, what the normal rule says, what the customer said, what has already been communicated, and what remains undecided.
The supplier case needs a different detail set, but the principle is the same. Show the requested change, the records checked, anything that does not match, and whether a payment is waiting. Do not bury the reason for concern in a transcript.
When the right person can understand the situation in thirty seconds, the agent is still doing valuable work even though it did not finish the task.
Send it to a person, not a department
“Escalate to finance” sounds clear until the message lands in a shared inbox at 4:55 on Friday.
Use names. The supplier-bank-change request goes to Priya, with Marco as backup. The customer exception goes to the service manager, then the owner if nobody responds within the promised time.
The response time matters because the agent has to tell the waiting person something honest. “I have passed this to our service manager, and you will hear from us by 3 p.m.” is useful. “I am connecting you now” is not, especially if nobody is joining the conversation.
Some cases can wait until tomorrow. Some need an immediate hold. Decide that while everyone is calm, not while money is moving or a customer is getting angrier.
Also tell the agent what it can do in the meantime. It can collect a missing invoice or confirm a callback number. It cannot imply that the exception is likely, retry the bank change, or make nearby changes that complicate the decision.
Keep the written policy short enough to use
Here is a practical starting version:
Handle the request when it fits the assigned job, the important facts agree, and the next action falls within your stated authority.
Stop when key information is missing or conflicting, identity is uncertain, someone asks for an exception, the action crosses a money or authority limit, the result would be difficult to reverse, or the customer asks for a person.
You may ask one clear question to resolve uncertainty. If you still need to guess, stop.
Send the named person a short note explaining what is being requested, which facts you verified, what remains unclear, what you already did, and any response time given. Preserve the original messages.
While waiting, acknowledge the request and gather clearly relevant information. Do not promise an outcome, repeat a high-impact action that failed, or change anything that could make the case harder to resolve.
If the named person does not respond on time, send the case to the listed backup and update the person waiting.
Now replace the general language with the details of the actual job. What dollar amount can the agent handle? Who decides exceptions? How quickly can that person realistically respond? What counts as confirmed identity in your business?
If those answers are vague, the agent is not the problem. The business has been relying on unwritten judgment, and the new tool has simply made that visible.
Try the uncomfortable examples before going live
Do not test only the request the agent was built to handle. Give it the Friday-afternoon version.
Ask for the ordinary refund, then ask for the exception. Submit a normal supplier update, then change the bank details with a payment due. Give it two customer records that agree, then two that do not. Have the customer ask for a person directly.
You are listening for the moment the agent stops. Too early, and every slightly unusual case becomes extra work. Too late, and the agent starts inventing certainty it does not have.
Then look at what arrives on the human side. Can the right person understand the issue quickly? Do they know what has already happened? Can they act without asking the customer to repeat the whole story?
That is the real test of the policy.
The goal is not endless autonomy
The worst agent mistakes are often small, confident choices. A promised exception. A changed record. A repeated charge. A customer kept in a loop because the agent was determined to finish.
A good agent has a better move. It notices that the situation got weird, stops cleanly, and brings the right person a useful summary.
Write down the moments your experienced people already treat carefully. Give each one a named destination and an honest response time. Keep the handoff short enough to read at a glance.
If you are working out those boundaries for a real business process, Brainiac can help you turn the awkward cases into an agent your team and customers can trust.