Human handoff: when an agent should stop

The most important setting in an AI agent is not what it can do. It is when it stops. A handoff is the moment the agent hands a conversation to a person on your team. Done well, the caller barely notices. Done badly, the caller repeats their story to a second voice and hangs up. This post covers when an agent should stop and what a good handoff carries.

Four reasons to stop

The first is a reserved topic. Some subjects you keep for people from the start: pricing negotiations, legal questions, anything medical, anything about a specific case. The agent recognises the topic and hands off before answering. The second is a request. A caller who asks for a person gets one. Making them argue with the agent first costs more than it saves. The third is emotion. A complaint, a raised voice, a caller who says this is the third time they have called. An agent can recognise these signals, and the right response is a person. The fourth is confidence. When the agent cannot find the answer in its knowledge, or the caller's words do not match anything it knows how to handle, it should say so and hand off rather than guess.

What a good handoff carries

The person taking over should not have to ask "how can I help you". A good handoff carries five things: a short written summary of what the caller wants; the full transcript so far; the caller's details and the record the agent found, such as the order or the appointment; the actions the agent already took, so nothing is done twice; and the reason for the handoff, so the person knows whether they are dealing with a reserved topic, a request, a complaint or a gap in the knowledge.

Three shapes of handoff

On a live call, the agent tells the caller it is transferring them, then connects a person who already has the summary on screen. When no one is free, the agent offers a callback, records the preferred time, and creates the task for your team. In chat, on WhatsApp or the web, the thread moves to a person's inbox with the history attached, and the person replies in the same window, so the caller sees one conversation, not two.

Tell the caller what is happening

In every shape, the caller should hear what comes next: that they are being transferred and to whom, or when to expect the callback, or that a colleague is now reading the chat. Silence during a transfer is what makes people hang up. A single plain sentence keeps them on the line.

After the handoff

The conversation stays in one record. The person's replies, the agent's earlier turns and the actions taken sit together, so whoever reads it later sees the whole thing. The agent can stay in the thread to draft replies for the person to approve, or step out entirely. Either way, the record is complete.

Writing the rules

Start conservative. Reserve more topics than you think you need, and let the agent handle only the conversations you are sure about. Then read the transcripts. Every week you will find handoffs that did not need a person, and you widen the agent's scope one topic at a time. You will also find the opposite: conversations the agent handled that a person should have seen. Those go back into the reserved list. The rules are never finished, and that is fine. They are a setting you own.

What to ask a vendor

Ask to see a handoff, not hear about one. Ask what the person on your side receives, and whether the caller has to repeat anything. Ask what happens when no one is free. Ask whether the agent can be told to stop on a topic, and how quickly that takes effect. The answers show whether the product was built around your team or around the agent.