When the AI Should Stop and Fetch a Human

Every business that switches on a customer-facing AI eventually has the same argument: should it be handling this at all? The answer is easier if you decide the boundaries before the first live conversation instead of after the first complaint. Our guide to teaching an AI to sound like your business treats the handoff as a designed feature rather than a fallback for when things break.

Triggers that should always escalate

Some categories are not judgement calls. Write them down and route them regardless of how well the AI seems to be coping.

• Anger, distress or the word complaint. Nobody's frustration has ever been improved by a well-worded automated reply.
• Anything about money moving backwards — refunds, chargebacks, disputed invoices, a charge the customer does not recognise.
• Health, legal or safety questions, including anything where a wrong answer has consequences beyond the sale.
• Requests for exceptions: a discount, a late cancellation waived, a service outside your normal terms.
• The customer asking for a person. This one is absolute, and stalling on it does more damage than any other failure.

Beyond the fixed list, escalate on uncertainty — an AI that is unsure is one message away from the confident invention described in stopping an AI from inventing answers. If the AI has failed to answer the same question twice, or has said "I will check" once, the conversation has already stopped being useful — pushing for a third attempt only deepens the frustration.

Let the AI hand over on its own

The most reliable escalation is the one the AI initiates when it recognises it is out of depth. That requires an explicit instruction to do so and a mechanism for it — turning off AI replies for that thread and flagging it for a person, rather than continuing to answer while a notification sits somewhere unread.

It also requires the human side to be equally simple. A switch in the conversation header that stops the AI, and another that resumes it once the awkward part is over, keeps takeover from being a decision anyone hesitates over. The team mechanics around that — who picks it up, how it gets assigned — are covered in assigning conversations without chaos.

“

Two failed attempts is not persistence, it is a customer being processed. Escalate on the second, not the fourth.

— be digital ai team

Hand over context, not just the thread

The worst version of a handoff is the customer being asked to explain everything again to a person who has just arrived. Whoever takes over should see what the customer wanted, what the AI already said, what it looked up, and why it escalated — all of it in the thread rather than in a separate system.

A short internal note at the point of handoff does most of the work: what the customer is asking, what has already been promised, what remains open. This is the same practice that keeps human-to-human handovers from losing information, as described in keeping context between colleagues.

Say something honest to the customer

The transition should be visible and plainly worded. "I'll bring in a colleague who can sort this out — they'll reply here shortly" is enough. Two things to avoid: pretending the AI was a person all along, and vanishing silently so the customer sits watching an empty thread.

If it will be a while, say how long. "Someone will be with you within the hour" sets an expectation you can meet; silence sets one you cannot.

Plan for nobody being there

Escalations at two in the morning are the scenario most teams skip. Decide what happens: the AI acknowledges, states when a person will respond, and the conversation is queued for the morning — or, for genuinely urgent categories, an on-call notification goes out. What must not happen is the AI carrying on as though nothing was wrong because the alternative was unstaffed. Your response-time commitments should account for this explicitly, which is what setting SLAs and response time targets for chat support is for.

Use handoffs as a measurement

Count them, and read a sample every week as part of reading the AI audit log. A high handoff rate on one topic means a missing document. A high rate at one time of day means a staffing gap. A rate that climbs after a model or prompt change means the change made things worse. Handoffs are the most honest quality signal an AI deployment produces, precisely because they record the moments it could not cope — and they should trend downward for reasons you can name, not because you quietly raised the threshold.