by NodicalASDIC is a 2027-2028 research project. Published measurements are those of the notebook; targets are targets.
NotebookSeptember 2026: first measurement
Before building anything, we measured what our agents do today: on 60,287 calls answered by a human in September 2026, half last ten seconds or less, and a voicemail is only recognised after 8 seconds (median).
The question
Where do our voice agents stand before the ASDIC programme? Not “are they good”, but: how long does an answered call last, after how long is a voicemail recognised, how long does the agent take to answer, and when does a handover to a human happen. Without these numbers, no target means anything.
The call set
All outbound calls in Nodical's production database for September 2026: 60,287 calls answered by a human, across 14 campaigns, in French. Unanswered calls and those that reached a voicemail from the start are not counted in durations. No conversation was listened to or read: the measurement only uses the timestamps written by the platform during the call.
The method
For each call, the platform records the moment of pick-up, the moment the agent concluded it was talking to a voicemail, the moment the other party stops speaking and the moment the agent starts answering, and the moment of any handover. We compute durations between these moments, then the median and the value under which 9 calls out of 10 fall (p90). We publish the median rather than the mean because a few very long calls would distort it.
The results
| Measurement | Median | 9 calls out of 10 |
|---|---|---|
| Duration of an answered call | 10 s | under 20 s for 8 calls out of 10 |
| Voicemail recognised after | 8 s | 14 s |
| Agent response delay, excluding end-of-speech detection | 0.14 s | 1.32 s |
| Handover to a human | 1.19% of answered calls, at 135 s | — |
Taken from Nodical's production database, September 2026, 14 campaigns, without reading the content of conversations. For calls under twenty seconds: 8 out of 10, not a p90.
What it says
Voicemail detection comes late: at 8 seconds (median), it falls after the end of half of the human calls. The agent's response delay, on the other hand, is already short — but it is measured excluding end-of-speech detection, that is, from the moment the system decided the person had finished; it is that decision, not the generation of the answer, that makes people wait. Handover is rare and late: 135 seconds (median), on calls most of which last under twenty.
What does not work yet
- We do not know the error rate of voicemail detection: that would require a hand-annotated corpus (human or voicemail, and at what moment it could be known). It is the programme's first task.
- The published response delay excludes end-of-speech detection; the full measurement, from the person's last word to the agent's first, needs that same annotation.
- One platform, one language, one month: these numbers describe Nodical in September 2026, not “voice agents”.
Next
Assemble a pilot set of about ten hours of pseudonymised calls, have them annotated (human or voicemail with the moment of recognition, ends of turns), and publish here the error rate of the current detection and the full response delay. Those numbers will be the programme's real starting line.
And in practice? how Nodical lets its agents make calls
Corrections
No correction so far.