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

# Voicemail Handling

> Configure Answering Machine Detection (AMD) to identify and handle voicemail systems in outbound campaigns

export const Screenshot = ({lightSrc, darkSrc, alt, caption, maxWidth = "880px"}) => {
  return <div style={{
    margin: "1rem auto",
    maxWidth,
    width: "100%"
  }}>
      <Frame>
        <img className="block dark:hidden" src={lightSrc} alt={alt} />
        <img className="hidden dark:block" src={darkSrc} alt={alt} />
      </Frame>
      {caption ? <p style={{
    marginTop: "0.5rem",
    fontSize: "0.875rem",
    color: "inherit",
    opacity: 0.8
  }}>
          {caption}
        </p> : null}
    </div>;
};

## How Voicemail Detection Works

Answering Machine Detection (AMD) enables your AI agents to automatically detect when outbound calls reach voicemail instead of a live person. When the agent detects voicemail, it hangs up and logs the outcome, allowing you to retry at different times to maximize chances of reaching a live person.

This critical feature prevents wasted agent time talking to answering machines and optimizes campaign efficiency by focusing resources on live conversations rather than voicemail systems.

<Note>
  **Phone Calls Only:** Answering Machine Detection (AMD) applies to [phone calls via Session Initiation Protocol (SIP)/Public Switched Telephone Network (PSTN)](/launch/phone-numbers) connections. Web-based conversations without a phone leg do not support voicemail detection.

  Access AMD configuration during campaign creation, campaign settings, and phone test calls. **Text-based** is the default mode. **ML-based** enables the external AMD participant for faster detection (\~1.5s vs 5-15s).
</Note>

***

## What is Answering Machine Detection?

### The Challenge

When making outbound calls, you encounter two possible scenarios:

**Scenario 1: Live Answer**

```text wrap theme={null}
Phone rings → Person answers → "Hello?"
→ Agent should engage in conversation
→ Full agent capabilities needed
```

**Scenario 2: Voicemail**

```text wrap theme={null}
Phone rings → Voicemail system answers → "You've reached John Smith..."
→ Agent should hang up and retry later
→ Don't waste time with full conversation script
→ Don't create awkward interaction talking over voicemail greeting
```

**The problem:** How does the agent know which scenario occurred?

### AMD Solution

AMD analyzes the audio in the first few seconds after call connection to determine if a human or machine answered:

**Detection process:**

```text wrap theme={null}
1. Call connects
2. AMD analyzes audio (0.5-3 seconds depending on method)
3. Classification: HUMAN or MACHINE
4. Agent executes appropriate behavior
```

**Benefits:**

* **Efficiency:** Don't waste agent time on voicemail
* **Better targeting:** Focus retry attempts on different times to reach live person
* **Higher connect rates:** Optimize call timing based on voicemail patterns
* **Better analytics:** Separate "reached voicemail" from "no answer" in reporting

***

## AMD Methods

### Detection Types

<CardGroup cols={2}>
  <Card title="Text-Based AMD" icon="text">
    Detect voicemail using keyword analysis with your transcriber and LLM — the default mode
  </Card>

  <Card title="ML-Based AMD (Optional)" icon="brain">
    Recognize voicemail patterns using a Deep Neural Network — optional addon for faster detection
  </Card>
</CardGroup>

### Text-Based AMD

**Role:** Default detection mode. Uses your agent's transcriber and LLM to identify voicemail keywords.

**How it works:**

```text wrap theme={null}
1. Call connects
2. Agent's AI voice pipeline (transcriber) receives audio
3. Agent waits for first turn/utterance to complete (pause detected)
4. LLM analyzes transcription for voicemail-like patterns:
   - "You've reached"
   - "Leave a message"
   - "Not available"
   - "Voicemail"
   - "After the beep"
5. If voicemail detected → Agent hangs up
6. If live person detected → Agent continues normal conversation
```

**Characteristics:**

<AccordionGroup>
  <Accordion title="Speed" icon="gauge">
    **Slower:** Waits for complete turn/utterance to finish

    Must wait for the entire voicemail greeting to complete (pause detected), then transcribe and analyze the full text. Typical detection occurs after 5-15+ seconds depending on length of voicemail message.

    **Limitation:** Long voicemail greetings mean longer wait times before detection

    **Best for:** Campaigns where accuracy is more important than immediate detection
  </Accordion>

  <Accordion title="Accuracy" icon="chart-line">
    **High accuracy when:**

    * Standard voicemail greetings with common phrases
    * Clear audio quality
    * Voicemail language matches transcriber language
    * B2B environments with professional greetings

    **Will NOT detect:**

    * Voicemail greetings in languages the transcriber doesn't support
    * Voicemail systems with no greeting (just beeps)
    * Non-verbal voicemail indicators

    **Lower accuracy when:**

    * Custom greetings without standard keywords
    * Short greetings ("Hi, leave a message" - very brief)
    * Poor audio quality or background noise
    * Background noise interfering with transcription
  </Accordion>

  <Accordion title="Best Use Cases" icon="briefcase">
    **Ideal for:**

    * **B2B campaigns** - Business voicemails typically use standard phrasing
    * **Campaigns prioritizing accuracy** - Reduces false positives by analyzing full utterance context
    * **Budget-conscious deployments** - Lower computational cost
    * **English-language markets** - Keyword detection optimized for English

    **Example scenarios:**

    * Sales outreach to business phone numbers
    * Appointment reminders to office lines
    * B2B lead qualification campaigns
  </Accordion>

  <Accordion title="Limitations" icon="triangle-exclamation">
    **May struggle with:**

    * Personal, creative voicemail greetings ("Hey, it's Mike, you know what to do")
    * Very short greetings
    * Voicemail language doesn't match transcriber language
    * Greetings that sound conversational ("Hello? Hello? Just kidding, leave a message")
    * Background music or noise in greeting

    **False positives:** Human who starts with "You've reached..." might be misclassified

    **False negatives:** Voicemail without keywords might be classified as human
  </Accordion>
</AccordionGroup>

### ML-Based AMD (Optional Addon)

**Role:** Optional fast detection layer you can enable for speed. Works in parallel with text-based AMD.

**How it works:**

```text wrap theme={null}
1. Call connects
2. Deep Neural Network (DNN) analyzes audio in real-time:
   - Speech patterns and cadence
   - Voicemail audio patterns
   - Acoustic characteristics
   - Timing and rhythm
   - Natural vs. recorded speech indicators
3. Model trained on tens of thousands of audio recordings
4. Classification: HUMAN or MACHINE
5. Language-independent detection
```

**Characteristics:**

<AccordionGroup>
  <Accordion title="Speed" icon="gauge-high">
    **Fast:** \~1.5 seconds

    Identifies live human responses within 1.5 seconds

    **Much faster than text-based AMD** which must wait for complete utterance
  </Accordion>

  <Accordion title="Accuracy" icon="chart-line">
    **Very high accuracy** in real-world conditions

    **Why enable ML-Based AMD:**

    * **Language-independent:** Works across all languages (text-based only works if transcriber language matches)
    * **Detects beep-only voicemail:** Catches voicemail systems with no greeting (text-based cannot)
    * **Handles creative greetings:** Detects personal/non-standard greetings without keywords
    * **Pattern-based detection:** Doesn't rely on specific voicemail keywords
    * **Fast detection:** \~1.5 seconds vs 5-15+ seconds for text-only
    * **Better for multi-language campaigns:** No language configuration needed

    **Limitations:**

    * Extremely short connections (\< 0.5 seconds of audio)
    * Highly degraded audio quality
  </Accordion>

  <Accordion title="Best Use Cases" icon="briefcase">
    **Ideal for:**

    * **Consumer campaigns** - Personal voicemails with creative greetings
    * **Multi-language campaigns** - Not dependent on English keywords
    * **Quality-focused campaigns** - When accuracy is more important than speed
    * **Complex markets** - Mixed business/personal numbers

    **Example scenarios:**

    * Consumer sales calls
    * Political campaigns
    * Non-profit fundraising
    * Healthcare outreach
    * Multi-language support campaigns
  </Accordion>

  <Accordion title="Advantages" icon="check">
    **Handles well:**

    * Creative personal greetings
    * Short greetings
    * Non-English voicemails
    * Greetings without standard keywords
    * Background music or sound effects
    * Natural conversational-sounding greetings

    **Robust across:**

    * Different languages
    * Regional accents
    * Various voicemail systems
    * Custom greetings
  </Accordion>
</AccordionGroup>

### How AMD Works

**Text-Based AMD (Base Layer):**

* Default campaign and phone-test mode
* Analyzes transcription for voicemail keywords
* Waits for complete utterance (5-15+ seconds)
* More conservative - rarely hangs up on live people

**ML-Based AMD (Optional Addon):**

* You can optionally enable this for faster detection
* Analyzes audio patterns in \~1.5 seconds
* Works in parallel with text-based AMD
* Faster but may occasionally hang up on live people

**Configuration Options:**

**Text-Based Only (Conservative):**

* Only text-based detection active
* Slower detection (5-15+ seconds)
* Rarely hangs up on live people
* **Trade-off:** Might miss some voicemails and talk to them
* Best for: When you want to avoid hanging up on live people at all cost

**Text-Based + ML-Based (Fast & Recommended):**

* ML detects in \~1.5 seconds
* Text-based validates in parallel
* Very high accuracy
* **Trade-off:** Occasionally might hang up on a live person
* Best for: Campaigns where talking to voicemail incurs additional cost

<Tip>
  **Which should you choose?**

  **Text-Based (recommended for most use cases):** Sufficient for the majority of campaigns. Rarely hangs up on live people, and handles standard voicemail greetings well.

  **Text-Based + ML-Based:** If you need faster detection (\~1.5s vs 5-15s) and can tolerate occasionally hanging up on a live person — for example, high-volume campaigns where talking to voicemail incurs meaningful cost.
</Tip>

***

## Configuring AMD

AMD can be configured in two places:

<CardGroup cols={2}>
  <Card title="Test Phone Calls" icon="phone-volume" href="#test-phone-calls">
    Enable AMD when testing your agent with phone calls
  </Card>

  <Card title="Campaign Settings" icon="megaphone" href="#campaign-settings">
    Configure AMD for outbound campaigns
  </Card>
</CardGroup>

### Test Phone Calls

Configure AMD when testing your agent via phone:

<Screenshot lightSrc="/images/answering-machine__answering-light.png" darkSrc="/images/answering-machine__answering-dark.png" alt="Test Agent dialog showing Phone call selected, Call Settings with Answering Machine Detection (AMD) options including Text-based and ML-based, From Number dropdown, To Number field, and Start Phone Call button" />

<Steps>
  <Step title="Open Agent">
    Go to your **Agent** page

    Click **Test Agent**
  </Step>

  <Step title="Select phone call">
    Choose **Phone call** as the test type
  </Step>

  <Step title="Configure AMD">
    Find the **Answering Machine Detection (AMD)** setting

    Choose between:

    * **Text-based** (default) - Avoids hanging up on live people at all cost
    * **ML-based** - Fast detection (\~1.5s) but may occasionally hang up on live people
  </Step>

  <Step title="Make test call">
    Select your From Number

    Enter To Number (your phone number for testing)

    Click **Start Phone Call**

    If the agent detects voicemail, it hangs up
  </Step>
</Steps>

### Campaign Creation And Settings

Campaign creation uses the default text-based AMD mode. To change AMD for a campaign, open the campaign after creation and use the campaign **Settings** tab.

<Screenshot lightSrc="/images/campaigns__amd-settings-light.png" darkSrc="/images/campaigns__amd-settings-dark.png" alt="Campaign settings in Expert mode showing the Answering machine detection field set to Text-based" />

<Steps>
  <Step title="Campaign Creation">
    Go to **[Campaigns](/manage/campaigns/overview)** section

    Click **Create Campaign** and fill in the required campaign name, agent, phone number, and schedule fields.
  </Step>

  <Step title="Open campaign settings">
    Open the campaign and select the **Settings** tab.
  </Step>

  <Step title="Select AMD strategy">
    Switch to **Expert** mode if needed, then find **Answering Machine Detection**.

    *Choose voicemail detection strategy for this campaign*

    Select one:

    * **Text-based** - Avoids hanging up on live people at all cost (slower, 5-15s)
    * **ML-based** - Fast detection (\~1.5s) but may occasionally hang up on live people
  </Step>

  <Step title="Save">
    The campaign settings page auto-saves the selected strategy.
  </Step>
</Steps>

**Changing AMD settings for existing campaigns:**

1. Navigate to campaign **Settings**
2. Switch to **Expert** mode if the AMD field is hidden
3. Locate **Answering Machine Detection (AMD)** dropdown
4. Select different strategy (Text-based or ML-based)
5. Save changes

<Warning>
  Changing AMD settings mid-campaign may affect analytics consistency. Consider creating a new campaign if you need to A/B test AMD configurations.
</Warning>

***

## AMD Behavior

When AMD detects voicemail, the agent automatically hangs up and logs the outcome.

The platform marks the call as **MACHINE** in campaign analytics, allowing you to schedule retries at different times to increase chances of reaching a live person.

***

## Testing AMD Configuration

### AMD Test Plan

<Steps>
  <Step title="Test ML-Based AMD">
    **Setup:**

    1. Configure agent with ML-Based AMD enabled
    2. Prepare test phone number with voicemail

    **Test:**

    1. Start test call to voicemail number
    2. Let call go to voicemail
    3. Monitor agent behavior

    **Validation:**

    * Agent hangs up within \~1.5 seconds
    * Call marked as MACHINE in logs
    * No conversation attempt with voicemail greeting
  </Step>

  <Step title="Test Text-Based AMD">
    **Setup:**

    1. Configure agent with Text-Based AMD only
    2. Use same voicemail test number

    **Test:**

    1. Start test call
    2. Let call go to voicemail with standard greeting

    **Validation:**

    * Agent waits for complete greeting (5-15+ seconds)
    * Agent hangs up after detecting keywords
    * Call marked as MACHINE
  </Step>

  <Step title="Test Live Person Detection">
    **Setup:**

    1. Test with both AMD methods
    2. Answer call personally

    **Test:**

    1. Start test call
    2. Answer and say "Hello?"
    3. Verify agent continues conversation normally

    **Validation:**

    * Agent does NOT hang up
    * Normal conversation proceeds
    * Call NOT marked as MACHINE
  </Step>

  <Step title="Test Edge Cases">
    **Scenarios to test:**

    **Silent answer:**

    * Answer but don't speak
    * Verify AMD doesn't misclassify

    **Quick greeting:**

    * Answer with very brief "Hi"
    * Verify conversation continues

    **Voicemail without keywords:**

    * Test with non-standard greeting
    * Monitor ML vs text-based performance

    **Beep-only voicemail:**

    * Voicemail system with no greeting
    * Verify ML-based detects, text-based may miss
  </Step>
</Steps>

***

## Troubleshooting

<AccordionGroup>
  <Accordion title="Wrong AMD Method Selected" icon="circle-question">
    **Symptoms:** Performance not matching expectations

    **Check:**

    * Review campaign AMD setting
    * Compare expected vs actual detection speed
    * Check false positive/negative rates in logs

    **Solution:**

    * Switch between Text-based and ML-based
    * Test both methods with your call patterns
    * Choose based on your priority (speed vs conservative)
  </Accordion>

  <Accordion title="High False Positive Rate" icon="user-xmark">
    **Symptoms:** Hanging up on live people frequently

    **Analysis:**

    * Review call recordings of false positives
    * Check if ML-based AMD is being too aggressive
    * Identify common patterns (background noise, specific greetings)

    **Solution:**

    * Switch to Text-based AMD (more conservative)
    * Improve call quality/reduce background noise
    * Test from different phone numbers
    * Contact support if persistent
  </Accordion>

  <Accordion title="High False Negative Rate" icon="voicemail">
    **Symptoms:** Agent frequently talks to voicemail

    **Analysis:**

    * Check if voicemails have non-standard greetings
    * Review if beep-only voicemail systems
    * Verify transcriber language matches voicemail language

    **Solution:**

    * Switch to ML-based AMD (better for non-standard greetings)
    * ML-based detects beep-only systems
    * Ensure agent speaks same language as target audience
  </Accordion>
</AccordionGroup>

***

## Next Steps

<CardGroup cols={2}>
  <Card title="Campaign Management" icon="megaphone" href="/launch/campaign-outbound-launch">
    Create and manage outbound calling campaigns
  </Card>

  <Card title="Campaign Management" icon="chart-line" href="/manage/campaigns/overview">
    Track AMD performance and optimize campaigns
  </Card>

  <Card title="Prompt" icon="pen" href="/build/conversation/prompt">
    Write effective prompts for call handling
  </Card>

  <Card title="Schedules" icon="clock" href="/launch/schedules">
    Configure optimal calling times based on AMD data
  </Card>
</CardGroup>