> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tryopencomputer.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Voicemail and screening

> Configure outbound voicemail actions, screening messages, and opening delays.

The **Settings → Voicemail & Screening** panel applies to outbound calls using separate speech and language models. The panel's settings are not a promise of support on realtime speech-to-speech models. Native Athena's phone-menu/answer handling is a separate runtime capability; use [Phone menus and hold](/guides/phone-menus.md) for that setup.

## Choose what happens on voicemail

1. Open the agent's **Settings → Voicemail & Screening**.
2. Enable **Enable voicemail and screening handling**.
3. Under **When voicemail is detected**, choose **Disconnect the call** or **Leave a message**.
4. For **Leave a message**, fill in **Voicemail message** using Text or choose a library Recording. A nonempty message or recording is required; the agent plays it and disconnects.
5. Choose **Save Voicemail Settings**, test the draft on a mailbox you control, and publish when the behavior is correct.

Use a concise message with the name, purpose, and approved callback details. Confirm the message was actually left; a voicemail disposition alone does not prove that recording reached the mailbox.

## Configure screening

In **Screening message**, state the agent's name and reason for calling. Choose Text or a library Recording. After playing it, the agent waits silently for a person. Leave this field blank if screened calls should end instead.

Open **Timing** and set **Screening wait (seconds)** from 1 to 60; the default is 30 seconds. Check a controlled screening flow where a person accepts, one where no one accepts, and a case with recorded prompts. Recordings require a supported playback path; see [Recordings](/guides/recordings.md).

## Tune the opening listening window

Edit **Start Call** on the canvas. Enable **Delayed Start** and set **Delay Duration (seconds)** from 0.1 to 10 to replace the default 1.2-second initial listening window. A brief human greeting can end the wait sooner. Check both a silent answer and a person who says hello before choosing a value.

## Classification model and instructions

Under **Classification model**, leave **Use Workflow LLM** enabled to use the configured workflow model, or select a separate supported provider/model and its required account fields.

Keep the built-in **Classifier instructions** unless you need to describe greetings or carrier announcements in another language. A custom prompt must preserve the full reply-label contract:

```text theme={null}
CONVERSATION, VOICEMAIL, NO_MESSAGE, SCREENER, SCREENING_WAIT, IVR, UNKNOWN
```

The response must be a single label. An unsupported response is treated as `UNKNOWN` and lets the call proceed; a prompt that only returns CONVERSATION or VOICEMAIL can silently disable the other branches. Clearing custom instructions restores the built-in behavior.

## Check results

Review classification activity, the message playback, silent wait, and final outcome in the run. Check human pickup, voicemail, screening, and an IVR separately. A campaign can automatically retry a voicemail outcome only when that campaign's retry policy allows it; configure this separately from the voicemail message. See [Campaigns](/guides/campaigns.md#campaign-auto-retries).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.