Agents rarely get switched off because the AI was bad. They get switched off because a customer went round in circles for four minutes, could not reach a person, and complained to someone senior. The escalation path is the product. The chat is the interface.
What a good handoff does
-
Triggers before the customer asks
Two failed attempts at the same question, a detected complaint, or a confidence score below threshold. Waiting for the customer to demand a human means you already lost them.
-
Carries the whole conversation
The person picking up sees everything already said, what the agent looked up, and what it could not resolve. Making a customer repeat themselves is the specific thing that makes people hate chatbots.
-
Knows when nobody is there
Out of hours it should say so, take the details, and commit to a time, rather than silently queueing into a void until Monday.
-
Never pretends
If it cannot do something, it says so plainly. Agents that bluff generate the complaints that get them turned off.
The threshold nobody sets
Most agents are configured to always produce an answer, because that is the default behaviour of the underlying model. That is the wrong default for a business. An agent should have a confidence level below which it stops answering and fetches a person, and that number should be set deliberately and reviewed against real conversations.
Set it too high and everything escalates and you have an expensive switchboard. Too low and it confidently tells a customer something untrue. This is a tuning exercise on real traffic, not a setting you get right on day one.
Design the failure first
Before writing a single prompt, answer these: who picks up an escalation, within what time, during which hours, and what happens outside them. If those four answers are vague, the agent will fail in exactly that gap.
Measure the right thing
Containment rate — the share of conversations the agent resolved alone — is the metric vendors quote, and it is easy to game by making escalation hard. Watch resolution and repeat-contact rate instead. An agent that contains eighty percent while generating callbacks the next day has not saved anyone anything.
We build the escalation before the conversation. See AI agents and chatbots.
Questions we get asked
Should the agent say it is an AI?
Yes. It becomes obvious eventually, and discovering it late damages trust more than knowing up front ever does. Most customers are entirely comfortable with an assistant that is honest about being one.
How fast should escalation be?
Within the conversation if someone is available, otherwise with a stated time and a commitment that is actually kept. Silence is the failure mode that generates complaints.
What if we have nobody to escalate to?
Then the agent should collect details and set expectations rather than pretending it can resolve things. An agent without a human behind it needs much tighter limits on what it attempts.
Can the agent hand to a specific person?
Yes, routing by topic or account is straightforward once it can read your systems. It is one of the more valuable things a connected agent does.