Answering Machine Detection

Answering machine detection is the technology that tells a dialer whether a call was picked up by a person, an answering machine or voicemail system, or something it cannot classify. It sits between your dialer and the phone network, listening to the first moments of a connected call and deciding how the conversation should proceed, whether that means connecting an agent, dropping a voicemail, or moving on to the next call.

For outbound calling operations, this single decision point affects agent time, compliance exposure, and connection rates all at once. Getting it wrong in either direction has a cost: too conservative, and live answers get treated as machines; too aggressive, and voicemail greetings get treated as humans. This page covers what answering machine detection actually does, how the audio analysis works, what accuracy and speed mean in practice, and how it fits into power dialers, predictive dialers, and AI voice agents.

What Is Answering Machine Detection

Answering machine detection, usually shortened to AMD, is a feature built into dialing platforms and telephony APIs that analyzes the audio at the start of a connected call to determine whether a human or a machine answered. It is a classification step, not a conversation. The system listens for a short window after the call connects, scores what it hears, and returns a result the dialer logic can act on.

AMD exists because outbound calling at volume runs into a basic problem: a dialer cannot see who or what picked up. Without detection, every call looks the same to the system until a human listens to it. AMD automates that first listen so the dialer, or an AI agent sitting behind it, can route the call immediately instead of waiting on a person to triage it.

How AMD Works

AMD systems make their decision by analyzing a handful of audio characteristics in the opening seconds of a call:

• Greeting length. Voicemail greetings tend to run longer than a live "hello," since they often include a name, a business, or instructions.
• Silence patterns. A live person usually responds with some back and forth or pauses differently than a recorded greeting followed by a long stretch of silence before a beep.
• Cadence. Recorded greetings have a flatter, more uniform rhythm than live speech, which tends to have more variation in pace and tone.
• Beep detection. A tone at the end of a greeting is one of the strongest signals that the call reached a machine, since it is the cue a caller would use to start leaving a message.

These signals are combined into a classification, and the system returns one of three outcomes: human, machine, or unknown. Unknown results happen when the audio does not give the system enough to decide confidently, for example a very short greeting, background noise, or a caller who answers with silence.

AMD can run in two modes. Synchronous detection holds the call until a classification is reached before handing it off, which adds a short delay but guarantees the dialer has an answer before acting. Asynchronous detection connects the call immediately and delivers the classification a moment later through a callback or event, which avoids holding the line but means the first instant of the call happens before the system has classified it.

Misclassification happens for a few recurring reasons: a live person who answers with a scripted-sounding greeting ("Thank you for calling, this is Jamie"), background noise or a bad connection masking the audio signal, a voicemail greeting that is unusually short, or a human who stays silent long enough to resemble the gap before a beep. No detection method eliminates these edge cases entirely, which is why "unknown" exists as an honest third outcome rather than forcing every call into human or machine.

Accuracy and Speed

Speed and accuracy are both relevant to AMD, but they are not the same thing, and it is worth being clear about which one a vendor is actually telling you about.

That number describes processing speed, how fast the detection engine returns a classification once it has audio to analyze. It is not an accuracy claim, and no single number can honestly summarize accuracy across every calling list, audio quality, and carrier path. Accuracy depends on the specific data you are calling, so the only reliable way to know how AMD will perform for you is to test it against your own calls.

If you want to evaluate detection accuracy for your own use case, we've written a practical guide: How to test answering machine detection accuracy. It walks through building a labeled test set, running it against live classification, and scoring the results so you have a number you can trust, because it's yours.

AMD With Power and Predictive Dialers

AMD is only useful if the dialer logic around it does something sensible with the result. On a power dialer, a human classification connects the call to the next available agent immediately, while a machine classification can trigger an automated voicemail drop instead of tying up an agent's time.

On a predictive dialer, AMD results feed into the pacing algorithm itself. Predictive dialers place more calls than there are available agents, betting that a portion will go to voicemail, so detection accuracy directly affects how aggressively the system can dial without running into compliance limits. This connects to the TCPA's abandoned call rule, which the FTC enforces at no more than 3 percent of answered calls abandoned within a 30-day campaign, with a required greeting played within 2 seconds for any call that is abandoned. AMD and pacing decisions are part of how a predictive dialer stays inside that rule. We go into this in more detail in how the abandoned call rate rule affects predictive dialers.

AMD for AI Voice Agents

AI voice agents depend on the same classification, just with a different consumer sitting on the other end of the result. Instead of routing to a human agent, a machine classification on an AI dialer can trigger the agent to leave a structured voicemail message, log the outcome, and move to the next call, while a human classification hands the conversation to the live agent logic.

Because AI agents often operate through tool calls and structured outcomes rather than a hard-coded dialer script, the detection result needs to be available in a format the agent's reasoning layer can act on. We cover how voicemail outcomes integrate into agent tool-calling in AI agent voicemail outcomes and MCP.

Direct Carrier Routing

RizzDial calls ship directly through AT&T, Verizon, and T-Mobile with direct carrier connections, rather than being routed through an intermediate aggregation layer before reaching the destination network. Fewer hops between the dialer and the carrier means less variability introduced into the call path before AMD ever gets a chance to analyze the audio.

RizzDial vs Twilio AMD vs Typical Cloud Dialers

Twilio AMD is billed per call as an add-on and gives developers granular control, which suits teams building custom dialer logic from scratch. RizzDial bundles detection, dialing, and voicemail drop into one platform so there is less to assemble. Which fits better depends on whether you want to build the dialer layer yourself or use one that already includes it.

FAQ

What's the difference between synchronous and asynchronous AMD?

Synchronous AMD holds the call briefly until a classification is ready before connecting it anywhere. Asynchronous AMD connects the call right away and delivers the classification a moment later through a callback, which avoids holding the line but means the first instant of audio happens before classification completes.

Why does AMD sometimes return "unknown" instead of human or machine?

Unknown happens when the audio in the detection window doesn't give the system enough signal to classify confidently, for example a very short greeting, background noise, or a caller who stays silent. It's a deliberate third outcome rather than a forced guess.

Does AMD work the same on every call?

No. Audio quality, greeting length, background noise, and carrier path all affect the signal AMD has to work with, so results can vary across lists and call conditions. This is why testing against your own calls matters more than any single published number.

Can AMD be used with voicemail drop?

Yes. A machine classification is typically the trigger that starts an automated voicemail drop, so the message plays without an agent staying on the line to deliver it manually.

How does AMD relate to the TCPA abandoned call rule?

The FTC's abandoned call rule limits abandonment to 3 percent of answered calls within a 30-day campaign and requires a greeting within 2 seconds of answer for any abandoned call. AMD and dialer pacing work together to help predictive dialers stay within that limit.

Is Twilio's AMD a full dialer?

No. Twilio AMD is a detection feature inside Programmable Voice. It returns a classification, but the dialer logic, including what happens next, is built by the developer using the API.

How do I know if AMD accuracy is good enough for my use case?

Test it against a labeled sample of your own calls rather than relying on a vendor's general claim. Our guide on testing answering machine detection accuracy walks through building that test set and scoring the results.

See It on Your Own Calls

The only way to know how answering machine detection performs on your lists is to run it against them.