The handoff problem: moving a chat from AI to human without losing context
Most AI support tools get the answer right and the handoff wrong. The moment a person is needed is the moment the experience usually falls apart — here's how we designed around it.
Every AI support demo ends at the good answer. Real support ends somewhere else: the customer has an edge case, the AI can't ground an answer, and a person has to take over. That transition is where most tools quietly fall apart, and it's the part customers actually remember.
Three ways handoff goes wrong
- 1The bot keeps talking. A human replies, and thirty seconds later the AI answers the same message too — because nothing told it to stop.
- 2The customer repeats themselves. The agent arrives cold, opens with “can you describe the problem?”, and undoes every minute the AI just saved.
- 3Nobody says who's who. The tone changes mid-conversation and the customer is left guessing whether they're still talking to software.
Rule one: one clear owner at a time
We tied the AI to conversation state rather than to a heuristic. The agent answers while a conversation is new. The moment a teammate replies, or the conversation moves out of that state for any reason, the AI is switched off for that conversation — permanently, not until the next message.
It's a blunt rule and that's the point. Anything cleverer — cooldowns, confidence thresholds, “only interject if the agent hasn't replied in five minutes” — produces a system nobody can predict, and support teams do not want a colleague whose behaviour they can't predict.
Rule two: the context travels
When an agent opens an escalated conversation, the sidebar already shows who the contact is, which company they belong to, and every previous conversation they've had. The transcript is right there, including everything they told the AI. There is no version of “can you describe the problem again” that is acceptable when the customer already described it twice.
We also put a summary at the top — one line on what's actually being asked, plus intent and sentiment. On a forty-message thread, that's the difference between picking it up in ten seconds and picking it up in three minutes.
Rule three: say who is talking
When a person joins, the customer sees a notice with that person's real name and photo, and the reply that follows carries the same identity. AI replies are labelled as AI. Nobody has to infer it from tone.
Customers are remarkably tolerant of talking to a bot. They are not tolerant of being misled about it.
This one is worth insisting on even when it looks like a downgrade. There's a temptation to give the AI a human name and let it pass — it tests well right up until someone realises, and then you've spent trust you can't earn back cheaply.
The unglamorous conclusion
Handoff quality is mostly product decisions, not model quality. Who owns the conversation, what the agent sees when they arrive, and what the customer is told — none of that gets better with a bigger model, and all of it is what makes the difference between AI support that helps and AI support that people learn to route around.