Ten seconds: what happens when a call is picked up

Why tell the story of ten seconds

The other pages of this site deal with one question at a time: recognising a voicemail, knowing when the person has finished speaking, measuring an agent. This one does the opposite: it follows a call from start to finish, in the order things happen. The length is not chosen at random. In Nodical's production database, in September 2026, 60,287 calls were answered by a human, across 14 campaigns, and their median duration is 10 seconds: half the calls are shorter, half are longer. Ten seconds is the ordinary call, not an extreme case.

One rule holds for the whole page: we only write what was measured. The measurement neither read nor listened to any conversation; it only uses the timestamps the platform writes during the call. When a moment of the call has no timestamp, we describe it without giving it a duration. The method is detailed in the first notebook entry.

The pick-up

The line opens. For the agent, it is a precise instant, recorded by the platform: everything that follows is counted from there. For the person, it is a reflex. They take the call without knowing who is calling, often busy with something else. The calls in this measurement are French, so they say “Allô?” — the French “Hello?” — or “Oui?”, or their name, then they stop and wait.

At that instant, the agent does not know what it is dealing with. What it hears may be a person, the start of a voicemail greeting, or silence: the person has not yet put the phone to their ear, or is saying nothing. A voicemail sometimes starts, too, with a voice that sounds like a “Hello”. Nothing in the first sounds settles it for sure.

Human or voicemail: the question that drags on

It is the first decision of the call, and the slowest. When the call has reached a voicemail, our agent currently recognises it after 8 seconds (median), and after 14 seconds for 9 calls out of 10.

Put those figures next to the length of a human call: 8 seconds to recognise a voicemail, 10 seconds for the median human call. These durations are of the same order. In other words, the machine takes almost as long to know who it is talking to as an ordinary conversation takes to unfold. The measurement does not say how many human calls end before that point; it says that the simplest decision of the call comes late, too late to guide what happens at the start. Why it takes so long, and what exists to go faster, is explained on the page about detecting voicemail; the beep, which does not always come when expected, has its own page.

The agent speaks, the person answers

Meanwhile, the call goes on. The agent introduces itself and says why it is calling; the person listens, cuts in, asks a question or answers. Each time they stop talking, the agent must first decide they have finished, then produce its answer.

That last step, we measure. Once the decision is made, the agent starts answering in 0.14 seconds (median), and in 1.32 seconds for 9 answers out of 10. That is short. But the figure is measured excluding end-of-speech detection: it starts at the moment the system concluded the person was done. The wait before it, while the agent listens to the silence wondering whether the sentence is over, is something we cannot yet measure without listening to the calls: we would need to know, for each turn, the instant the person had really finished. Yet it is the wait the person hears. It is explained on the page about end of turn.

The hang-up

Then the call stops. For half of the answered calls, that happens after 10 seconds or less; for 8 calls out of 10, before 20 seconds. Everything above has to fit into that span: the first contact, the human-or-voicemail question, the agent's introduction, and the person's decision to carry on or not.

What the measurement does not say, and it matters: why the call stops. Nobody listened to these conversations. A short call may be a polite refusal, a “call me back later”, a wrong number, someone hanging up without a word, or an answer obtained quickly. The measurement gives the shape of the call, not its content. The reasons will come from a corpus of pseudonymised, hand-annotated calls, whose method will be published in the notebook.

The calls that last

A small share of calls takes another path: the conversation settles in, and the agent hands over to an adviser. That is the case for 1.19% of answered calls, and the handover happens at 135 seconds (median). These calls have nothing in common with the ordinary call: they run far beyond most of the others. Handover is therefore rare and late. We do not yet know what share of those handovers came at the right moment; it is one of the measurements waiting for the annotated corpus, and one of the reasons an agent should be judged on several numbers, not just one.

What these ten seconds say about ASDIC

Put end to end, these measurements describe an imbalance. The agent answers fast once it has decided; it is slow to decide almost everything else: who it is talking to, when the person has finished. And Nodical does not own, today, any comprehension model of its own: 0% is the starting point shown on this site.

The ASDIC programme targets exactly these early-call decisions. Its targets are expressed in numbers, and they are targets: none is reached, and each will be published with the measurement that proves it, not before. They are detailed on the page about measuring a voice agent. This page will follow the notebook, with the same rule: what is measured, and nothing else.

The other possible start: when nobody picks up

Everything above assumes that someone answered. There is also the call that rings out, on the side of a business receiving calls: it leaves no visible trace, which is exactly what makes it hard to count. Nodical explains on its site what you lose when nobody picks up, and how to measure it with your own data.

And in practice? what you lose when nobody picks up

Sources

accessed on 1 October 2026.

1. ASDIC, carnet : « Septembre 2026 : première mesure » — https://asdic.ai/carnet/2026-09-premiere-mesure

Corrections

No correction so far.