Escalation: When an AI Agent Hands Off to a Person
The most important thing an AI agent does is know when to stop. An agent that answers everything is worse than one that answers most things and hands the rest to a person cleanly. Escalation is the rule set that decides which is which, and it is the part worth spending an afternoon on before you go live.
The four triggers worth setting
Almost every good escalation policy is built from four kinds of trigger.
- Named topics. A written list of subjects that always reach a person: complaints, refunds above a threshold, legal questions, anything medical, press enquiries, cancellations of a large contract.
- Sentiment. A caller who is angry, distressed, or repeating themselves gets a person, regardless of topic. Nothing burns goodwill faster than an unbothered voice talking to someone upset.
- Repeated failure. If the agent has failed twice to answer the same question, it stops trying. Two attempts is a fumble; five is an insult.
- Explicit request. "Can I speak to someone?" is honoured immediately, every time, with no retention loop and no "I can help with that".
Warm transfer beats cold transfer
There are two ways to move a call. A cold transfer dumps the caller on a colleague who knows nothing. A warm transfer passes the conversation so far — who is calling, what they want, what has already been tried, what the agent promised.
The second one is the only one worth having. The measure of a good hand-off is that the person picking up does not have to ask a single question the customer has already answered. If your team is re-interviewing every transferred caller, the escalation is configured badly, not the agent.
What happens when nobody picks up
Escalation policies fall apart at 7pm. Decide in advance what the agent does when the transfer goes unanswered:
- Try a second person or a rota, if you have one.
- Take a message with a callback number and a written summary of the issue.
- Set the expectation out loud: "someone will call you back before ten tomorrow" is fine; a vague "someone will be in touch" is not.
- Flag it. Escalations that ended in voicemail should be the first thing your team sees in the morning, not buried in a list of all calls.
Some topics deserve to wake someone up. Most do not. Splitting the list into 'now' and 'morning' is usually a five-minute decision that saves a lot of nights.
Reviewing escalations is how the agent improves
Every escalation is a small piece of feedback. Over a week you get two useful piles.
The first pile is correct escalations: the agent hit a topic it should never handle, and the rule worked. Leave those alone.
The second pile is avoidable ones: the agent handed off because it did not know your parking situation, or your Saturday hours, or which insurers you accept. Every one of those is a missing line in the script, and adding it converts a hand-off into a handled call. Most businesses cut avoidable escalations by half in the first fortnight simply by reading them.
Frequently asked
Can I force certain topics to always reach a person?
Yes. A named-topic list is the first rule in every escalation policy, and it takes priority over everything the agent thinks it can handle.
Does the person who picks up see the conversation?
Yes. A warm transfer carries the summary and the transcript so far, so the caller does not repeat themselves.
What if my team is on another call?
You set the fallback: a second person, a rota, or a message with a committed callback window. Escalations that ended in a message are flagged for review.
Can the agent call a customer back itself?
Outbound follow-up is a separate role you can turn on, and it is gated by region because consent rules differ. Most businesses start with inbound only.
Will the agent argue with an upset caller?
No. Frustration is an escalation trigger in its own right. The agent acknowledges, stops selling, and gets a person.
How do I know if my rules are too strict?
Read a week of escalations. If most of them are questions the agent could have answered with one more line in the script, your rules are fine and your script is thin.