Building AI Chatbots That Know When to Hand Off to a Human

Every support chatbot we've shipped eventually hits a conversation it shouldn't try to finish alone, and the projects that succeeded were the ones where we designed the handoff path with as much care as the happy path, rather than bolting it on after launch as an afterthought once a frustrated customer complained loudly enough.

We track a small set of explicit signals rather than trying to infer frustration from tone alone: repeated rephrasing of the same question, explicit requests for a human, sentiment dropping across consecutive turns, and low retrieval confidence on a question the bot is about to answer anyway. Any one of these on its own triggers a soft nudge toward a human option; two together trigger an automatic handoff.

For a subscription billing client, the highest-value handoff trigger turned out to be topic-based rather than sentiment-based — the moment a conversation touched cancellation or a billing dispute, we routed straight to a human regardless of how well the bot was doing, because the cost of a bot mishandling a cancellation request was far higher than the cost of an unnecessary handoff on an easy question.

The handoff itself needs to carry context, not just the customer. Early versions of one client's system dropped the full conversation on handoff, and customers had to repeat everything to the human agent, which generated more complaints than the original problem. We now pass a structured summary — what was asked, what the bot tried, why it escalated — directly into the agent's queue view.

We measure handoff quality with a metric most teams skip: the rate of "unnecessary" handoffs, where a human agent resolves the ticket in under a minute using information the bot already had. High unnecessary-handoff rates usually mean the bot is under-confident, not that handoff logic is broken, and tuning that balance is an ongoing calibration, not a setting you get right once.