Voicemail Detection

Voicemail Detection classifies the very first turn of a call as either a live conversation or a voicemail greeting, and ends the call automatically if it's a voicemail — useful for outbound calling (e.g. campaigns) where a meaningful fraction of calls land on voicemail.

How it works

On the call's first turn, a classification LLM call runs alongside normal call setup: it's given the transcribed first-turn speech and asked to respond with either CONVERSATION or VOICEMAIL. If it responds VOICEMAIL, the call ends automatically rather than the agent proceeding to talk to a recording.

Two safeguards against a slow or ambiguous voicemail greeting:

• Speech Cutoff — if the first turn's speech runs longer than this many seconds, classification is triggered early rather than waiting for the caller (or recording) to finish talking. Default: 8 seconds, range 1–30s. Long unbroken speech on a first turn is itself a signal of a voicemail greeting.

Enabling it

Open an agent's Settings → Voicemail Detection and toggle it on. It's off by default for every agent.

Classification LLM

By default, voicemail classification uses the same LLM configured for the rest of the agent's conversation (Use Workflow LLM, on by default). Turn this off to use a separate provider/model/API key just for classification — useful if you want a cheaper or faster model for this narrow yes/no task without changing the agent's main conversational LLM.

System Prompt

The classification prompt is fully editable. It must instruct the model to respond with exactly CONVERSATION or VOICEMAIL — a sensible default prompt is provided, but you can customize it (e.g. to tune for a specific voicemail-greeting style you're seeing false negatives/positives on).

Related

• Agent Node — the conversation logic that runs once a call is classified as a live conversation.
• Campaigns — outbound calling, the primary use case for voicemail detection.