When an AI receptionist should hand off to a person

An AI receptionist should hand off to a person on five triggers: an emergency, a question that needs professional judgement, a complaint, a request outside what it was given tools for, and a customer who asks for a human. The handoff should name who is taking over and when, keep the conversation history, and happen before the AI has tried to help. Trying first is how handoffs go wrong.

Trigger 1: an emergency

Medical symptoms, a gas smell, a flood, a pet that collapsed, a client in custody. These are matched by fixed rules, not by the model's judgement, and they skip generation entirely. The customer gets safety wording you approved, and the on-call person is alerted through the route you set up. The AI adds nothing of its own.

Trigger 2: professional judgement

Whether a treatment suits someone. Whether a case has merit. Whether a repair is covered. Whether an insurance claim will pay. These have a common shape: the honest answer depends on expertise and on facts the AI does not have. The handoff message says who will answer and roughly when, and the AI does not offer a guess "in the meantime".

Trigger 3: a complaint

Anger is not a question. When a message reads as a complaint, the right response is an apology in the business's voice and a person, not a cheerful offer to book again. The AI should recognise the tone, acknowledge it briefly, and route the conversation with the history attached so the person does not ask the customer to repeat themselves.

Trigger 4: outside its tools

The AI can only do what it was given the means to do. If it can book but not refund, a refund request is a handoff. If it can quote from the menu but not negotiate, a "can you do it for less" is a handoff. It should say so plainly: "I cannot change prices, but Sam can. I have passed this on."

Trigger 5: the customer asks

"Can I talk to a real person" is always honoured, immediately, without a further question. Some customers ask because they are frustrated; some because the matter is sensitive; some because they just prefer it. The reason does not matter.

What a good handoff looks like

It names a person or a team, gives an honest time ("this morning", "within the hour", "when we open at 8"), keeps the conversation in the same thread so the customer does not start over, and says what happens next. If nobody is available, it says that too, rather than promising a reply that will not come.

Why the rules run first

A model asked to be helpful will try to help. If the escalation check runs after the model has drafted a reply, the draft has already tried to answer the emergency or soothe the complaint. Escalation must be a gate in front of generation: match the trigger, skip the model, route. Ask any vendor where their checks sit.

Frequently asked questions

Will too many handoffs defeat the point?

In the first weeks, handoffs and the gaps log tell you what to add to knowledge. Most businesses find the share of conversations handed off falls quickly as the approved answers grow. The five triggers above should stay handoffs forever.

What if the person does not respond?

The conversation sits in an assigned queue with the customer's expectation recorded. A queue nobody watches is a management problem, not an AI one, and the inbox should make it visible.

Should the AI tell the customer it is handing off?

Always. "I am passing this to Dana" is honest and sets expectations. Silent handoffs feel like being ignored.

What Frontiva does here

Frontiva's AI agents check for prompt injection and sensitive topics before anything else runs, escalation keywords you configure hand a conversation to a person instantly, and a request for a human is one of the readiness scenarios every agent is scored on. Handed-off conversations wait in the unified inbox, where you assign them to a person, tag them and set priority.

Escalation keywords and the urgent-message rule are both settings on the AI agents page. For the wider picture, see what is an AI receptionist.