voicemail_detected.
There are two layers of detection. They work independently and complement each other.
Layer 1 — AI detection (always on)
The first thing the other party says is checked by the AI: voicemail greeting, or real person?AI evaluates the opening
If it sounds like a voicemail (mentions leaving a message, a beep or tone, “you have reached”, “not available”, and similar — in any language), the AI ends the call silently.
- Always on — no setup, runs on every outbound call.
- Free — no extra charge.
- Any language.
- Speed: around 8–9 seconds — it waits for the full greeting before deciding.
- A transcript of the greeting is saved to the call record.
Layer 2 — Carrier AMD (optional)
AMD (Answering Machine Detection) is a faster, carrier-level check that runs on your own telephony account. It listens to the audio directly and reports human or machine — often within ~3 seconds, before the greeting finishes.- Off by default. Enable it per agent with the Voicemail detection toggle in Call Settings.
- Speed: around 3 seconds.
- Cost: roughly $0.0075 per answered call, billed on your telephony account.
- Best for: US and Canada.
- When AMD ends the call, it happens before the AI is involved, so the call has no transcript, sentiment, or summary. This is expected.
Comparison
| AI detection | Carrier AMD | |
|---|---|---|
| Default | Always on | Off (toggle) |
| Speed | ~8–9 sec | ~3 sec |
| Cost | Free | ~$0.0075 / answered call |
| Languages | Any | Best in US / CA |
| Transcript saved | Yes | No |
| Ended reason | voicemail_detected | voicemail_detected |
What happens on detection
Either layer ends the call the same way:- The call ends immediately — the agent does not leave a message.
- The call record gets
ended_reason: voicemail_detected. - Your webhook fires so you can react in your own system.