When an AI phone agent should hand the call to a human

Tuesday, 16:42. A woman calls a dental clinic in Odense. The phone host is an AI. It says so in its first sentence, asks what she is calling about, and she answers that a tooth has been knocked loose and it is bleeding. What happens in the next four seconds decides whether the clinic has a good phone host or a bad one.

The good one does one thing: it stops being clever. It says it will pass her to a person right away, it dials the clinic's emergency line, and while she waits the dentist gets a line on screen: "Woman, loose tooth after impact, bleeding, calling from 2x xx xx xx, sounds frightened." She repeats nothing. The human starts where the machine left off.

The bad one tries to book her an appointment on Thursday.

Handover is the product, not the exception

When we set up an AI voice agent for a company, we spend more time on the rules for when it should let go than on anything else. That surprises most customers. They bought a machine to answer the phone, and then we spend the first workshop talking about all the times it should not.

But that is where the difference lies. An AI phone host is not good because it handles many calls on its own. It is good because the customer on the other end never regrets that a machine picked up. And that feeling is almost always created in the transition.

A handover is the moment a call moves from AI to human, with everything the machine has understood sent along with it. If that last part is missing, it is not a handover. It is a transfer, and customers have hated those for thirty years.

The four rules that trigger a handover

The rules vary from company to company, but there are four we never deviate from. They are in the configuration before the first customer calls.

• The customer asks for it. A person who says "I want to talk to a human" gets a human. No "let me just try to help first". The first time. Every time.
• The topic is sensitive. Health, complaints, missing money, anything legal, a child, a death. The list is written together with the company, and it is always longer than they thought when we started.
• The machine has misunderstood twice. After the second time the customer corrects it, it gives up and passes the call on. A third attempt costs more trust than it ever saves.
• The company's own rules. Calls from a particular number, a customer with an open case, a word that must never be handled by a machine. At a property management firm the word was "water damage". At a car dealer it was "leasing".

Outside opening hours the rules change. There is no human to pass the call to, so the rule becomes: book a call from a human in the first free quarter hour tomorrow, and say exactly when. "We will call between 8:00 and 8:15" is a promise that can be kept. "We will get back to you as soon as possible" is not.

What the human must receive

The decisive thing in a handover is not that the call moves. It is what moves with it.

We send four things. Who is calling, with the name and history the number gives access to. What it is about, in one sentence. What the machine has already tried, so the human does not ask the same question again. And how the customer sounds: calm, irritated, frightened, in a hurry. That last item seems soft, but it is the one the staff use most. It sets the tone of the first sentence, and the first sentence sets the rest.

All of it is on screen before the phone rings at the human's desk. And afterwards it is stored as a summary, not as an audio file, unless the company has actively chosen otherwise. That is how we always do it, and it is written into our principles for responsible AI.

There is also a legal side that cannot be opted out of. Article 50 of the EU AI Act applies from 2 August 2026 and requires that people are told when they are talking to a system and not to a human. Our phone host says so in its first sentence, and it did so before the rule took effect. Not because the law requires it, but because a customer who discovers it halfway through the conversation never trusts the company again.

The number everyone measures, and the ones that matter

The industry loves one number: the share of calls the AI handles on its own. The containment rate. The higher, the better, say the sales decks.

It is the wrong number, and it produces bad phone hosts. A machine tuned to maximise that share holds on to calls it should release. It tries to book Thursday for a woman who is bleeding.

We measure three other things instead. How many seconds pass from the decision to hand over until a human is speaking. How often the customer has to repeat something after the handover; that number should be zero, and when it is not, something is missing from the summary. And how many of the calls the machine handled on its own lead to a new call about the same thing within two days. That last number exposes the calls the machine thought it had solved.

The containment rate takes care of itself when those three numbers are good. The other way round does not work.

An AI phone host is not measured by how many calls it handles alone, but by how few customers regret that it picked up.

How to write your own rules

If you are considering a phone host, start with the rules, not with the voice. Put the three people who take the most calls today in a room for an hour. Ask them to write down the ten calls they would never leave to a machine. That is your sensitive list. Then ask them for the ten calls they would happily be rid of. That is the machine's work. Everything in between is decided by the rule of two misunderstandings.

We wrote earlier about the technical foundation: latency, language and the three sets of rules a European call has to comply with. That is still the prerequisite. But a phone host with perfect latency and the wrong handover rules is still a bad phone host.

If you want to hear what it sounds like when it is set up properly, call ours on +45 52 51 74 95. It says what it is in the first sentence. Ask it for a human, and notice what happens.