Chatbot to human handover: the measure of a good assistant is how cleanly it gives up

Handover, or escalation, is the moment an assistant recognises it should not answer and passes the conversation to a person instead of continuing. Done properly it carries the context across so nobody repeats themselves, and it happens on a deliberate trigger rather than when the visitor gives up. It is the feature that decides whether people trust the rest of the assistant, because everyone eventually asks it something it cannot handle.

On three signals. First, an explicit request — someone asking for a person should get one. Second, low confidence: if the assistant has nothing relevant in your content, guessing is worse than handing over. Third, category — refunds outside policy, complaints, anything contractual, anything where a customer is upset. Those three cover almost every case, and configuring them deliberately is more reliable than hoping the model reads the room.

Almost always, yes — refusing to fetch a person when asked is the fastest way to lose someone. The one reasonable exception is doing something useful on the way: collecting the order number or email the agent will need anyway, and telling the visitor what happens next. That is a second of work that saves a round trip. Making them argue with a bot to reach a human is not.

The full transcript rather than a summary, anything the visitor already supplied, the page they were on, which trigger fired, and the time. The transcript matters most and summaries are the common mistake — an agent needs to see the wrong answer the assistant gave, because that is often what the conversation is now about.

Yes, and you can write the message it uses. The default connects the visitor to a human rather than leaving them stranded; most people rewrite it to name a real contact route for their business.

Handover collects the visitor’s details and routes the conversation to your team rather than opening a real-time agent seat inside the widget. If a live-agent console is a hard requirement for you, that is worth knowing before you evaluate further.

Two levers. The system prompt is yours, so you can name the categories it must not attempt — refunds, complaints, anything contractual. And leaving gaps in the knowledge base unfilled is itself a way of keeping certain topics off the assistant.

Usually the fix is content, not configuration. Read the transcripts where it gave up: nearly always there is a page that does not exist or a page that is ambiguous. Writing one paragraph often removes a whole category of handovers.

Modern models pick up on tone reasonably well, and you can instruct it explicitly in the system prompt to offer a human when someone seems frustrated or has asked the same thing twice. It is worth writing that instruction rather than assuming it.

They are stored with the conversation so your team can follow up, subject to the same retention as other sessions — 7 days on the free plan, 30 days on paid plans. Sessions can be deleted individually or in bulk.