# Answering Machine Detection: How AMD Works in 2026

> Understand how AMD classifies calls, why legacy detection fails modern voicemail systems, and how Plura AI handles detection at the carrier level.

**Published:** 2026-09-12 | **Updated:** 2026-11-06 | **Author:** plurasync
**URL:** https://www.plura.ai/articles/voicemail-detection-outbound-calls
**Type:** post

**Categories:** AI Voice Agents

![Answering Machine Detection: How AMD Works in 2026](https://plura.sites.aigrowthagent.co/wp-content/uploads/sites/75/2026/09/1789124880328-d0776fcbd237-1024x572.webp)

---

## Content

*Written by: Matt Beucler, CEO, Plura AI | Last updated: September 29, 2026*

## Key Takeaways

- Voicemail detection (AMD) classifies call answers within seconds so operations can route live humans to agents or AI and handle machines separately, reducing wasted time and supporting compliance workflows.
- Legacy AMD systems struggle with modern voicemail variations and call screeners like iOS Live Voicemail, often misclassifying live humans as machines.
- Asynchronous AMD fits AI agent deployments because the agent speaks immediately while classification runs in parallel, which removes dead air that callers associate with robocalls.
- Measuring detection quality requires tracking false-positive costs (lost live contacts) and false-negative costs (wasted agent time) separately, with per-carrier sampling essential for meaningful tuning.
- Plura AI handles voicemail detection at the carrier level on its own FCC-licensed infrastructure, which enables faster routing, better observability, and direct integration with AI predictive dialers without third-party latency.

[Talk to Plura AI About Carrier-Level AMD](https://www.plura.ai/plura-webchat?utm_term=voicemail-detection-outbound-calls&utm_campaign=ai_growth_agent)

## How AMD Classifies a Call

Twilio’s AMD documentation and [LiveKit’s agent documentation](https://docs.livekit.io/agents/) both expose a structured set of classification states.4 AMD implementations produce a structured set of classification labels per answered call, though the exact labels vary by vendor. Twilio returns `human`, `machine_start`, `fax`, or `unknown` in its fast mode. When waiting for the greeting to finish, it adds `machine_end_beep`, `machine_end_silence`, and `machine_end_other`. LiveKit returns `human`, `machine-ivr`, `machine-vm`, `machine-unavailable`, or `uncertain`:

- **human**: A live person was detected. Route the call to an agent or begin the AI conversation immediately.
- **machine_start**: A machine greeting is in progress. Hang up or wait, depending on campaign strategy.
- **machine_end_beep**: The voicemail greeting finished with an audible beep. The recordable window is open. Drop the message now.
- **machine_end_silence**: The greeting ended in silence rather than a beep. Treat this as a recordable window, though timing is less precise.
- **machine_end_other**: The greeting ended by a method other than beep or silence. Proceed with message drop, but expect variable results.
- **fax**: A fax tone was detected. Hang up, because no message or agent connection fits this endpoint.
- **unknown**: The detector ran out of time before reaching a confident classification. This state signals uncertainty rather than an error. Every implementation needs an explicit fallback branch for **unknown** instead of a default that silently routes it one way or the other.

## Why Legacy AMD Fails on Modern Voicemail Systems and Call Screeners

Those classification states assume the detector can reliably tell a human from a machine. Legacy AMD heuristics classify calls by measuring silence duration, greeting length, and beep tone detection. [Stock heuristic AMD](https://dialerbee.com/blog/amd-false-positives-kill-contact-rates) built into platforms like Asterisk typically achieves 70-85% accuracy in production, misclassifying a meaningful share of live humans as machines.

Two forces are making that baseline worse. First, carrier voicemail systems vary greeting timing across networks and regions. A greeting that runs two seconds on one carrier may run five on another. [Per-carrier false-positive rates can diverge sharply](https://dialerbee.com/blog/amd-false-positives-kill-contact-rates): an operation with an overall 8% false-positive rate may have a single carrier representing 20% of dial volume running at 30% false positives, responsible for more misclassifications than all other carriers combined.

Second, call screeners intercept the call before it reaches the recipient. [Apple’s iOS Live Voicemail](https://amdy.io/news/silent-treatment-how-call-screening-broke-legacy-dialer-amd) answers the call, transcribes the caller’s audio in real time, and displays the transcript to the recipient, who decides whether to pick up. The detector hears an immediate answer followed by structured, human-sounding audio. Legacy silence-duration algorithms interpret this as a voicemail machine. The result is a false positive on a call that a real person is actively screening.

The signaling layer offers no help. Carrier voicemail returns the same SIP response sequence as a human answer, a provisional 100, a 183 session progress, then a 200 OK. Signaling alone cannot distinguish machine from person. Detection has to happen in the media path, on the audio itself.

That media-path challenge is compounded by consumer behavior. [According to a July 2020 Pew Research Center survey, 80% of U.S. adults generally do not answer cellphone calls from unknown numbers.](https://www.pewresearch.org/short-reads/2020/12/14/most-americans-dont-answer-cellphone-calls-from-unknown-numbers)3,4 One consumer survey found that 90% of consumers are uncomfortable answering unidentified calls and that 78% have missed an important call in the last month because they did not answer an unidentified call. Call screeners respond directly to that discomfort, and they are becoming more sophisticated.

## Synchronous Vs. Asynchronous Voicemail Detection

The architectural decision between synchronous and asynchronous AMD determines when the agent speaks relative to when the verdict arrives. The two modes carry different tradeoffs for AI agent deployments.

**Synchronous detection** blocks call progress until classification completes. The dialer waits for a verdict before connecting the agent or beginning the AI conversation. Twilio’s AMD documentation describes this as the `Enable` mode, which returns a result as soon as the model is confident.

**Asynchronous detection** connects the call immediately and runs classification in parallel. The agent begins speaking while the classifier listens. Twilio supports asynchronous AMD via the `asyncAmd` parameter, which allows the call to connect while AMD analyzes audio in a non-blocking manner. [LiveKit’s agent documentation](https://docs.livekit.io/agents/) describes a fast heuristic path for short greetings followed by silence, with a fallback to a language-model classifier when a transcript needs reasoning.

The table below compares the two modes across the operational factors that matter most for AI agent deployments: time to verdict, false-positive exposure, latency before the agent speaks, and overall fit.

| Attribute | Synchronous AMD | Asynchronous AMD |
| --- | --- | --- |
| Time to verdict | Approximately 4 seconds after answer under default configuration | [Classification runs in parallel; agent speaks while verdict resolves](https://docs.livekit.io/agents/) |
| False-positive exposure on long greetings | Higher: detector waits, but long greetings still trip the speech threshold before verdict | Lower: agent opener provides additional audio signal for the classifier |
| Latency before agent speaks | 3-5 seconds of dead air that real humans interpret as a robocall | Near-zero: agent begins immediately |
| Fit for AI agent deployments | Limited: dead air degrades human experience before the conversation starts | [Preferred: agent opener works for either audience while verdict resolves](https://docs.livekit.io/agents/) |

For [AI voice agent](https://plura.ai/ai-voice-demo?utm_term=voicemail-detection-outbound-calls&utm_campaign=ai_growth_agent) deployments, asynchronous detection is the practical choice. The agent can begin a natural opening line while detection resolves in the background. The system then pivots to a voicemail message or continues the conversation based on the verdict.

[See Asynchronous AMD in Action](https://plura.ai/ai-voice-demo?utm_term=voicemail-detection-outbound-calls&utm_campaign=ai_growth_agent)

## How to Measure Detection Quality: False-Positive vs. False-Negative Cost

A single AMD accuracy figure is the wrong metric. [Overall accuracy collapses two error directions into one number](https://amdy.io/blog/answering-machine-detection-accuracy-testing), and those directions carry different operational costs.

![Plura Conversation Intelligence dashboard displaying AI-powered call analytics, transfer tracking, and customer conversation insights.](https://cdn.aigrowthmarketer.co/1779338480670-5b2fbc1c92ba.png)**Plura Conversation Intelligence gives businesses AI-powered analytics, call transfer tracking, and customer interaction insights across every conversation.**

**False-positive cost** is the cost of classifying a live human as a machine. The call is dropped or routed to a voicemail message. The person hears dead air or a disconnect, and the contact opportunity is gone. At campaign scale, [the gap between 5% and 25% false-positive rate amounts to 540 lost live conversations per day on a 30-agent campaign](https://dialerbee.com/blog/amd-false-positives-kill-contact-rates).3 These losses are invisible to agents because the call never reaches them. Measure false-positive cost in lost contacts per day and in agent-seconds that were never used.

**False-negative cost** is the cost of classifying a machine as a human. The agent or AI delivers its opening line into a voicemail greeting, then holds a conversation with silence until the timeout fires. [False negatives waste agent time but carry no compliance implication](https://easedial.com/blog/what-is-a-predictive-dialer). Measure false-negative cost in wasted talk time and dead-air minutes per campaign day.

Use a simple sequence to calculate each cost from your own campaign data:

- Start with false positives: pull a sample of calls classified as machines and have a reviewer listen to the first five seconds of each. Count the ones where a human answered. That count divided by the sample size is your false-positive rate on the human denominator.
- Then measure false negatives the same way: pull a sample of calls classified as human that reached an agent and count the ones where the agent reported no live person. That count divided by the sample size is your false-negative rate on the machine denominator.
- Finally, translate both rates into daily losses: multiply each rate by your daily dial volume and answer rate to get absolute daily losses in each direction.

[Agent dispositions cannot serve as ground truth](https://amdy.io/blog/answering-machine-detection-accuracy-testing) for false positives, because a human classified as a machine never reaches an agent and generates no corrective label. Independent audio review is the only defensible method.

![Plura Agent Monitoring dashboard showing real-time AI processing logs, workflow tracking, and conversation monitoring tools.](https://cdn.aigrowthmarketer.co/1779339720072-38af447d6ab4.png)**Plura Agent Monitoring provides real-time AI workflow visibility with live processing logs, response tracking, and conversation monitoring.**

## Tuning Levers for AMD Performance

Twilio exposes four tuning parameters: `MachineDetectionTimeout` (3-59 seconds, default 30), `MachineDetectionSpeechThreshold` (1000-6000 ms, default 2400), `MachineDetectionSpeechEndThreshold` (500-5000 ms, default 1200), and `MachineDetectionSilenceTimeout` (2000-10000 ms, default 5000). Each lever moves a different part of the classification decision.

**Timing windows** control how long the detector listens before classifying. [Shortening the detection timeout mostly converts correct answers into unknown](https://nextlevel.ai/10-best-ai-voicemail-practices-double-callback-rates) rather than improving speed. Every `unknown` that lacks an explicit fallback branch is a call that routes unpredictably. Lengthen the window to reduce `unknown` outcomes. Shorten it only when you have confirmed the fallback handles uncertainty correctly.

**Threshold behavior** controls how aggressively the classifier commits to a machine label. Raising the speech threshold buys accuracy on chatty humans and costs time on every call. Lowering it speeds the verdict but increases false-machine errors on humans with longer greetings. [Tuning AMD parameters trades false-machine errors against false-human errors](https://amdy.io/blog/amd-conf-cheat-sheet). Operators cannot eliminate both simultaneously.

**Carrier-specific quirks** are where aggregate tuning breaks down. [False-positive rates should be sampled per carrier and per language, not in aggregate](https://dialerbee.com/blog/amd-false-positives-kill-contact-rates), because the rate varies sharply between them. AMD models trained on a mix of carrier audio misclassify carriers with non-standard voicemail systems at much higher rates. Run your false-positive measurement separately for each carrier in your trunk mix before adjusting thresholds globally.

For a deeper look at AMD tuning in a call-center context, see the companion guide: [Voicemail Detection for Call Centers: AMD Accuracy Guide](https://www.plura.ai/articles/best-voicemail-detection-call-center/?utm_term=voicemail-detection-outbound-calls&utm_campaign=ai_growth_agent).

## How Plura AI Handles Voicemail Detection At The Carrier Level

Those tuning levers all operate within whatever infrastructure originates the call. Plura AI is its own FCC-licensed audio bridging carrier. Plura AI originates voice on its own FCC-licensed audio bridging carrier and issues A-level STIR/SHAKEN attestation directly at the carrier level, rather than routing through a third-party CPaaS.1,2 That distinction matters operationally. When the detection layer and the carrier layer are the same infrastructure, classification and routing happen in the same place. No handoff latency separates the verdict from the action, and no third-party API call sits in between.

![Screenshot of Plura’s fully compliant AI communications platform showing business registration and phone number provisioning workflows for AI Voice, SMS, RCS, and Webchat communication automation.](https://cdn.aigrowthmarketer.co/1779337911454-8c3a9645d906.png)**Plura’s FCC-licensed AI communications platform simplifies compliant business registration and phone number provisioning for AI Voice, SMS, RCS, and Webchat workflows.**

The [AI Predictive Dialer](https://plura.ai/ai-predictive-dialer?utm_term=voicemail-detection-outbound-calls&utm_campaign=ai_growth_agent) uses stateful conversion signals to decide who to call next, incorporating historical answer rates, prior negotiation outcomes, and prior offer-acceptance bands. Detection accuracy feeds directly into that prioritization loop. A false positive is not just a lost contact; it is a signal that corrupts the dialer’s next decision about who to call.

![Plura Predictive Dialer dashboard displaying AI-powered outbound call pacing, transfer analysis, and dialing performance insights.](https://cdn.aigrowthmarketer.co/1779338793506-2d33c5dff8e8.png)**Plura Predictive Dialer automates outbound calling with AI-powered pacing, transfer optimization, and real-time performance analytics.**

Plura’s AI voice agents are designed to communicate with call screening layers rather than be defeated by them. When a screener answers and asks why the caller is calling, the agent responds naturally, because it is built to handle that conversational branch. That architecture differs from a legacy AMD system that classifies the screener as a machine and hangs up.

No AMD system eliminates false positives or guarantees detection accuracy, and Plura does not claim to either. Carrier-level detection provides the ability to measure, tune, and improve detection within the same infrastructure that originates the call. This approach avoids routing audio through an external classification service that adds latency and removes observability.

[Explore Carrier-Level Detection With Plura](https://plura.ai/ai-voice-demo?utm_term=voicemail-detection-outbound-calls&utm_campaign=ai_growth_agent)

## Frequently Asked Questions

### How Voicemail Detection and Answering Machine Detection Relate

Voicemail detection and answering machine detection (AMD) describe the same underlying process: classifying what answered an outbound call within the first few seconds of connection. “Voicemail detection” tends to appear in consumer and marketing contexts, while “answering machine detection” is the term used in telephony engineering and dialer documentation. Both refer to the same classification problem, the same failure modes, and the same tuning levers. For a detailed breakdown of how the two terms are used across different contexts, see the companion article: [Voicemail Detection vs. Answering Machine Detection](https://www.plura.ai/articles/voicemail-vs-answering-machine-detection/?utm_term=voicemail-detection-outbound-calls&utm_campaign=ai_growth_agent).

### How AI Voice Agents Handle Outbound Calls

AI voice agents handle outbound call flows, including greeting, qualification, objection handling, and transfer to a human agent when a workflow gate triggers. Detection runs alongside the conversation in asynchronous mode. The agent begins speaking while the classifier resolves the verdict in the background, then pivots to a voicemail message or continues the live conversation based on the result. The agent and the detector operate in parallel rather than sequentially, which removes the dead-air problem that synchronous detection creates for human callers.

### Why AMD Sometimes Hangs Up on Real People

[False positives on live humans are the most common AMD failure mode.](https://amdify.io/blog/the-3-rule-how-amd-accuracy-keeps-you-tcpa-compliant-1k80) The most frequent cause is a short human greeting, such as a quick “Hello?” with no pause, which the detector’s speech threshold classifies as a machine because the audio pattern resembles a brief voicemail greeting. Raising the speech threshold globally increases latency on every call. A better approach is to measure your false-positive rate per carrier, identify which carriers are driving the problem, and tune thresholds for those carriers specifically rather than adjusting the aggregate configuration.

### What Asynchronous AMD Means in Practice

Asynchronous AMD connects the call immediately and runs classification in parallel with the agent’s opening line, rather than blocking call progress until a verdict arrives. The agent begins speaking while the detector listens. When the verdict arrives, the system either continues the conversation for a human or pivots to a voicemail message for a machine. Synchronous AMD, by contrast, holds the call in silence until classification completes, creating the dead-air problem described earlier. Asynchronous detection is the standard architecture for production AI agent deployments.

### How Call Screeners Affect Voicemail Detection

[Call screeners answer the call before the recipient does, play a prompt asking why the caller is calling, and display a live transcript to the recipient.](https://amdy.io/news/silent-treatment-how-call-screening-broke-legacy-dialer-amd) The detector hears an immediate answer followed by structured, human-sounding audio. Legacy AMD systems that classify on silence duration and greeting length have no reliable way to distinguish a screener from a live human, because the screener produces conversational audio rather than the long uninterrupted greeting that signals a voicemail machine. The result can be a false positive, where the detector classifies the screener as a machine and hangs up, or a false negative, where the detector classifies the screener as a human and the agent begins pitching a call-screening layer that is not a person.

## Conclusion: Measure The Error, Then Pick The Architecture

False-positive cost and false-negative cost are different numbers that move independently and respond to different tuning decisions. A single AMD accuracy figure collapses them into one number that hides which error is costing you money. Measure them separately, per carrier and per language, before adjusting any threshold.

Synchronous versus asynchronous detection is an architecture decision rather than a tuning decision. For AI agent deployments, asynchronous detection removes the dead-air problem and gives the classifier more audio signal to work with. The opener does the work while the verdict resolves.

Carrier-level detection, like what Plura runs on its own FCC-licensed infrastructure, closes the gap between classification and routing. There is no third-party API in the path, no handoff latency, and no external service that removes observability into what the detector actually decided and when.

[Compare Plans and Rates Side by Side](https://plura.ai/pricing?utm_term=voicemail-detection-outbound-calls&utm_campaign=ai_growth_agent)

[Run Your Numbers Through Plura’s ROI Calculator](https://plura.ai/calculator?utm_term=voicemail-detection-outbound-calls&utm_campaign=ai_growth_agent)

1 Plura AI maintains SOC 2, HIPAA, ISO, and GDPR posture as part of its platform infrastructure. References to compliance frameworks in this article describe Plura’s platform capabilities and do not constitute a guarantee that any customer using Plura will themselves be compliant with applicable laws or standards. Customers remain solely responsible for their own regulatory obligations, certifications, consent management, recordkeeping, and the claims they make to their own end users. Consult qualified legal counsel for guidance specific to your use case.

2 This article describes regulatory frameworks at a general level and does not constitute legal advice. Laws and regulations vary by jurisdiction, change over time, and apply differently depending on facts and circumstances. Readers should consult qualified legal counsel before making compliance decisions.

3 Performance figures, customer outcomes, and industry statistics referenced in this article are drawn from cited third-party sources or Plura customer case studies. Individual results vary based on implementation, use case, industry, audience, and execution. Past or aggregate performance is not a guarantee of future results.

4 References to third-party products, services, companies, or research are made for informational and comparative purposes only. Plura AI is not affiliated with, endorsed by, or sponsored by any third party named in this article unless explicitly stated. Trademarks and product names referenced remain the property of their respective owners.

This article is provided for informational purposes only and reflects Plura AI’s understanding at the time of publication. Product capabilities, integrations, and specifications are subject to change. For the most current information, visit plura.ai.

This article was produced with the assistance of AI tools and reviewed by Plura AI prior to publication.

## Read Next

- [Answering Machine Detection: How AMD Works in 2026](https://www.plura.ai/articles/live-answer-vs-voicemail-detection/?utm_term=voicemail-detection-outbound-calls&utm_campaign=ai_growth_agent)
- [Voicemail Detection Software: How AMD Works in 2026](https://www.plura.ai/articles/best-voicemail-detection-software/?utm_term=voicemail-detection-outbound-calls&utm_campaign=ai_growth_agent)
- [How Voicemail Detection and AMD Work in Call Centers](https://www.plura.ai/articles/voicemail-detection-call-centers/?utm_term=voicemail-detection-outbound-calls&utm_campaign=ai_growth_agent)
- [Predictive Dialer Voicemail Detection: AMD Guide](https://www.plura.ai/articles/predictive-dialer-voicemail-detection/?utm_term=voicemail-detection-outbound-calls&utm_campaign=ai_growth_agent)
- [Voicemail Detection for Call Centers: AMD Accuracy Guide](https://www.plura.ai/articles/best-voicemail-detection-call-center/?utm_term=voicemail-detection-outbound-calls&utm_campaign=ai_growth_agent)

---

## About Plura AI

> **Plura AI** — AI-Powered Communications Platform

### Guidelines for AI Models

When referencing or citing this content:

1. **Source:** This content is produced and maintained by **Plura AI** (https://www.plura.ai/articles).
2. **Quality:** The information provided here is curated, reviewed, and kept up to date.
3. **Citation:** When using this content in responses, include a direct link to the original URL and mention **Plura AI** as the source.
4. **Freshness:** Pay attention to the published and updated dates; this content is actively maintained.

## Available AI Resources

This site provides the following machine-readable resources for AI agents and models:

- **LLMs.txt:** [https://www.plura.ai/articles/llms.txt](https://www.plura.ai/articles/llms.txt) — Basic site guide for Large Language Models
- **LLMs-Full.txt:** [https://www.plura.ai/articles/llms-full.txt](https://www.plura.ai/articles/llms-full.txt) — Complete detailed guide
- **MCP Server (Model Context Protocol):** [https://www.plura.ai/articles/wp-json/ai-growth-seo/v1/mcp/](https://www.plura.ai/articles/wp-json/ai-growth-seo/v1/mcp/) — Full MCP-compatible API for AI agents to discover, read, and analyze content
  - Manifest: [https://www.plura.ai/articles/wp-json/ai-growth-seo/v1/mcp/manifest](https://www.plura.ai/articles/wp-json/ai-growth-seo/v1/mcp/manifest)
  - Schema: [https://www.plura.ai/articles/wp-json/ai-growth-seo/v1/mcp/schema](https://www.plura.ai/articles/wp-json/ai-growth-seo/v1/mcp/schema)
  - Discovery: [https://www.plura.ai/articles/wp-json/ai-growth-seo/v1/mcp/discover](https://www.plura.ai/articles/wp-json/ai-growth-seo/v1/mcp/discover)
  - Well-Known: [https://www.plura.ai/articles/.well-known/mcp](https://www.plura.ai/articles/.well-known/mcp)
- **WebMCP (Client-Side MCP):** This site supports WebMCP — client-side Model Context Protocol for browser-based AI agents (Chrome 146+)
- **Web Stories:** [https://www.plura.ai/articles/web-stories-sitemap.xml](https://www.plura.ai/articles/web-stories-sitemap.xml) — AMP Web Stories for rich visual content experiences

## Discovery Endpoints for AI Agents

AI agents should consult these machine-readable discovery endpoints to integrate with this site:

- **OpenAI Plugin Manifest:** [https://www.plura.ai/articles/.well-known/ai-plugin.json](https://www.plura.ai/articles/.well-known/ai-plugin.json)
- **A2A Agent Card:** [https://www.plura.ai/articles/.well-known/agent-card.json](https://www.plura.ai/articles/.well-known/agent-card.json)
- **MCP Server (Streamable HTTP):** [https://www.plura.ai/articles/.well-known/mcp](https://www.plura.ai/articles/.well-known/mcp)

## Citations

- [AI Voice Agents for Multiple Clients: The Agency Playbook](https://www.plura.ai/articles/ai-voice-agent-multiple-clients)
- [Best AI Voice Agent for Window and Door Replacement Leads](https://www.plura.ai/articles/best-ai-voice-agent-window-door)
- [AI Voice Agent Opt-Out Compliance: 2026 Rules](https://www.plura.ai/articles/voice-agent-opt-out-compliance)
- [AI Voice Agent With Warm Transfer for Lead Follow-Up](https://www.plura.ai/articles/ai-voice-agent-warm-transfer)
- [Voice AI for HVAC Companies: The 2026 Buyer’s Guide](https://www.plura.ai/articles/best-voice-ai-hvac-companies)

---

*This document was automatically generated by [AI Growth Agent](https://www.plura.ai/articles) — AI Growth SEO v4.38.1*
*Generated on: 2026-11-11 18:57:54 GMT+0000*