What voicemail detection does
An outbound call that reaches an answering machine sounds, for the first few seconds, like a call that reached a person. Something picks up, something speaks. If the Finn treats that as a live caller it will talk over the greeting, wait for a reply that never comes, and burn a full call's worth of telephony and AI minutes on a recording.
Voicemail detection is the check that decides whether the thing that answered is a person or a machine, and the setting that decides what happens next.
Where you configure it
Two places, and they are separate.
| Where | Field | Scope |
|---|---|---|
| Finn → Call Settings tab | voicemail_detection | Default for every call this Finn makes |
| Deployment launcher → Advanced | Voicemail behavior | Overrides the Finn default for that campaign only |
Both are dashboard settings. There is no separate voicemail endpoint to call.
The Finn-level default is End call if voicemail detected. See Call settings for the neighbouring fields (max_idle_duration, limit_call_duration, handoff_number) and Deployments for where the campaign-level override lives.
The two behaviours
End call
The Finn hangs up as soon as it concludes a machine answered. The call is recorded with a voicemail outcome and becomes eligible for retry under your retry policy.
This is the default and the right choice when the point of the call is a conversation. A qualification call, a support callback, an intake interview: none of them are worth anything as a monologue into a mailbox.
Leave a message
The Finn plays a message via TTS instead of hanging up, then ends the call.
The important consequence, and the one that surprises people: if voicemail behaviour is set to leave a message, the platform does not retry that contact. The call is treated as delivered. A campaign configured to leave messages will reach each contact once, not up to three times.
That is correct for reminders and notifications. It is wrong for anything where you actually needed to talk to the person, because you have spent your only attempt on a recording.
Interaction with retries
Retry settings live on the deployment.
| Setting | Default | Notes |
|---|---|---|
| Max retries | 2 | Applies to no-answer, busy, and voicemail-end |
| Retry interval | 4 hours | Minimum wait between attempts |
Retries are not dialled at the exact interval. They are dialled at the next valid calling-hour window for that contact's local time. A voicemail hit at 5:50 PM with a 4-hour interval and a 9 AM to 6 PM window does not redial at 9:50 PM. It redials the following morning.
So the practical retry cadence for voicemail-heavy audiences is roughly once per day per contact, not once per interval. Plan campaign end times accordingly, or the deployment will stop before its retries have run.
What goes wrong
The greeting gets clipped. Detection needs a moment of audio to decide. If you set behaviour to leave a message, the first fraction of a second of your TTS message can land while the mailbox is still finishing its beep-and-record setup. Write the message with a throwaway opening word rather than leading with the critical detail.
Live callers get classified as machines. A person who answers with a long, flat, scripted greeting looks structurally like an outgoing voicemail message. It is uncommon, but it is not zero, and the failure is silent: the caller hears a hang-up or an unexpected recorded message. If a deployment shows a voicemail rate far above what the audience should produce, listen to a sample of the calls flagged as voicemail before assuming the number is real. Use the per-call table in Live deployments and open the call to hear the audio.
Machines get classified as live callers. The opposite failure. The Finn talks to the mailbox, waits through max_idle_duration, re-prompts, waits again, and eventually hits limit_call_duration. You pay for the full call. These show up in analytics as long-duration calls with near-empty transcripts, which is a useful thing to filter for.
Compliance openings collide with mailboxes. If your welcome message is a defined script carrying a required disclosure, and voicemail behaviour is set to leave a message, decide explicitly whether that disclosure belongs on the recording. See Welcome message and Recording consent.
Measuring it
Voicemail shows up as a call outcome alongside success, failed, and no-answer. See Call outcomes for the full set and Webhooks for the events that carry them.
Post-call analysis still runs on voicemail calls, but the transcript contains the mailbox greeting and your message, not a conversation. Fields like appointment_booked will come back false or as failed extractions. When you compute conversion rates, exclude voicemail-outcome calls from the denominator or your numbers will look worse than the campaign actually performed. See Post-call analysis.
Recommended starting point
- Conversational campaigns: End call, max retries 2, and check the voicemail rate after the first hundred calls.
- Reminders and notifications: Leave a message, and accept that each contact gets one attempt.
- Never set Leave a message on a first-run campaign you have not tested on yourself. Test on Call from the Finn page reaches your own phone, and letting it go to your own voicemail is the only way to hear what the message actually sounds like end to end.