When an AI agent should stop trying and escalate to a person

There is a lot of talk about what an AI agent can automate, and much less about what happens when it should stop trying. That second part is what separates a well-designed system from one that frustrates customers until they stop using it.

Escalation is a function, not an error

It is easy to think of handing off to a person as “when the agent fails.” It is more useful to think of it as just another function of the system, with its own rules, as deliberately designed as any automated response. An agent that escalates at the right moment, with the right context, is as valuable as one that resolves things on its own.

Signals that should trigger escalation

• Repetition. If the customer asks the same thing twice in different words, the first answer did not work. Pushing a third automated variant only increases frustration.
• Explicit request. “I want to talk to someone” should always be respected, without trying to hold them back with one more reply.
• Emotional load. A tone of complaint, urgency or frustration detectable in the message is a sign the case needs empathy, not an automated flow.
• Out of scope. When the question falls outside what the agent reliably knows, escalating is better than risking an invented answer.

What should never happen: making the customer repeat themselves

Watch out: making the customer explain everything again from scratch to the person handling it undoes all the value of having qualified them with the agent first.

Technically correct escalation is not just “handing off the conversation.” It means giving the person a summary of what was already discussed: what the customer asked, what they were told, what data they already provided (order, account, specific issue). Without that, the customer has to repeat themselves, and the experience ends up worse than if they had never talked to the agent at all.

How to design the right “wait time”

An agent should not try to resolve something indefinitely before escalating. A useful rule is a maximum number of attempts on the same topic (usually two or three) before actively offering to hand off to a person, instead of waiting for the customer to ask out of frustration.

How to measure whether escalation is working well

Counting how many conversations the agent resolved alone is not enough. You also need to measure how long it took to escalate the cases that needed it, and whether the human team received enough context to avoid asking everything again. An “efficient” agent that holds onto cases it should have escalated sooner is not a good agent, it is a bottleneck in disguise.

How we do it at Bitora

We design escalation rules as a central part of the project, not a last-minute addition, and we build the context handoff so the person receiving the conversation has everything they need from the first message.

If your current agent (or the one you are evaluating) has not solved this, see how we approach it on AI agents and automation.