Escalation is a feature, not a failure

Every AI employee will hit cases it should not handle. Whether that is a good experience or the moment a customer gives up is determined entirely by design decisions you make in advance. Teams that treat escalation as an exception build it last and badly.

Customers are remarkably tolerant of “a person will take this” and remarkably intolerant of having to start again.

Never make them repeat themselves

The handoff must carry the full conversation, what the AI attempted, what it retrieved, and why it stopped. An agent who opens the ticket cold and asks the customer to explain again destroys the goodwill the automation just earned.

This is a plumbing problem, not a modelling one, and it is where a surprising share of these projects actually fail.

Escalate before the customer asks

By the time someone types “talk to a human”, you have already lost the interaction. Trigger on the signals that precede it: repeated rephrasing, rising frustration, a second failed answer, or any topic on your stakes list. Pre-emptive escalation reads as competence; reactive escalation reads as an obstacle.

Every “speak to a human” request is an escalation you should have triggered one turn earlier.

Be honest about what they’re talking to

Disclose that it is an AI teammate, plainly, at the start. The alternative — a customer discovering it mid-conversation after being misled — converts a routine interaction into a complaint about your integrity rather than your service.

Disclosure also lowers the bar in your favour: people forgive an AI for not knowing something and don’t forgive a person for pretending to be one.

Hand back deliberately, if at all

If work returns to the AI after a human resolves the hard part, the customer should be told, and the AI needs the human’s resolution in its context. Silent hand-backs produce contradictions that undo the recovery you just paid an agent to perform.

Watch the escalation rate in both directions

A rate that is too high means the scope is wrong; a rate near zero usually means it is escalating too little, not performing too well. The cases it should have escalated and didn’t are the expensive ones, and they only show up if you sample resolved conversations rather than only reviewing escalated ones.