Answering Machine Detection and AI Voicemail Drop: How Modern Dialers Stop Talking to Machines

Most outbound dials never reach a person. Gong Labs' analysis of more than 300 million cold calls puts the average connect rate at 5.4% — 13.3% for top-quartile reps — which works out at roughly nineteen dials for every conversation. Most of the rest ring out or land on a voicemail greeting. Answering machine detection (AMD) is the piece of dialer software that decides, in the first seconds after pickup, whether a person or a recording answered, so that agents are not fed machines all day.

Classic AMD was built for predictive dialers, and it shows. It decides from silence lengths, speech cadence and word counts; it is wrong often enough that every vendor's documentation warns about it; and when it guesses 'machine' on a real person, it hangs up on them. This article explains how AMD actually works, why the trade-off between speed and accuracy is built in, what voicemail-drop tools fix and what they do not, and how an AI calling agent that listens first — classifying the answer from the words and the audio, detecting the beep and leaving a personalised voicemail — changes the economics of the unanswered call.

What answering machine detection is for

A dialer's job is to keep people talking to people. A predictive or power dialer places more calls than there are free agents because most will not be answered; when one connects, it has to decide within a second or two whether to bridge an agent. Bridging on every pickup wastes agents' time on greetings and beeps; waiting too long leaves a person saying 'hello?' into silence. AMD is the classifier that makes that call.

Depending on configuration, a machine result makes the dialer hang up, schedule a retry or play a recorded message — the voicemail drop. A human result bridges the call to the next free agent. Genesys Cloud's documentation states the rule its dialer applies plainly: a live person is 'less than 2,200 ms of speech followed by 700 ms or more of silence'; anything outside that pattern is treated as voicemail or left unresolved.

• 5.4%Average cold-call connect rate across 300M+ dials; 13.3% for top-quartile repsGong Labs, 300M cold calls (2024)
• 67%US adults who let an unknown number ring out and then check the voicemail (survey of 10,211)Pew Research Center, unknown-number calls (Dec 2020)
• 2,400 msDefault speech-length yardstick for 'machine' in one major carrier API's AMD (adjustable 1,000–6,000 ms)Twilio Docs, Answering Machine Detection
• 2.73% → 5.87%Email reply rate without vs with one or two voicemails; 2.2% after three or moreGong Labs, should you leave voicemails

How classic AMD works: silence, cadence and word counts

The open-source Asterisk PBX documents the knobs of its AMD() application, and they are a good map of the whole technique. The algorithm compares audio energy with a silenceThreshold and then times what it hears: initialSilence (too long before any greeting means a machine), greeting (a greeting longer than this means a machine), afterGreetingSilence (silence after the greeting means a person waiting for you to speak), maximumNumberOfWords and maximumWordLength (too many words, or one unbroken stretch of speech, means a recording), all bounded by a totalAnalysisTime. It returns MACHINE, HUMAN, NOTSURE or HANGUP, with a cause such as LONGGREETING or MAXWORDS.

Carrier APIs do the same thing with friendlier names. Twilio's AMD uses a speech-length yardstick of 2,400 ms by default (adjustable between 1,000 and 6,000 ms), treats speech as finished after 1,200 ms of silence, returns 'unknown' after 5,000 ms of initial silence, and times the whole analysis out at 30 seconds. Its two modes expose the central trade-off: Enable returns a verdict as soon as it can (human, machine_start, fax or unknown), while DetectMessageEnd waits through the whole greeting and returns machine_end_beep, machine_end_silence or machine_end_other — slower, but the only way to know when a message can start.

The built-in trade-off: decide fast or decide right

None of these heuristics understands words. 'Hello, this is John Smith' passes as a person because it is short; 'Hello, this is John Smith. Thank you for calling Widgets Inc, how can I assist you today?' is classed as voicemail because it runs past the speech limit — that is Genesys's own example. Twilio warns of the opposite error: 'a very short two-second voicemail greeting' can be read as a human picking up. Tighten the thresholds and you decide faster but misread more; loosen them and a real person waits longer in silence.

Vendor accuracy claims deserve the same scepticism. One provider advertises 97% precision and a decision in under three seconds, while its own FAQ describes 85–90% as the industry norm. None of these figures comes with a published method or test set, and all are measured on the vendor's own traffic. Treat them as marketing until you have measured them on your own lists — and remember that 'accuracy' hides the direction of the errors, which is what actually costs you.

Dead air and false positives: what they cost

Every misclassification has a price, and the prices are not symmetrical. A false 'machine' means the dialer hangs up on a live prospect who took the trouble to answer — the most expensive outcome in outbound, and one that teaches that person to ignore your number next time. A false 'human' bridges an agent into a voicemail greeting, which is merely wasted seconds and a confused disposition.

The subtler cost is the delay itself. While AMD listens for its silence window, the person who answered hears nothing; when an agent is finally bridged, the opening line arrives two or three seconds late, after a second 'hello?' — the signature of an automated call that most people recognise and resent. Dropped calls, where no agent is free when the verdict comes in, are worse: the person answered and was hung up on. Both are experienced as disrespect, and both show up later as lower answer rates on your numbers and harder openings for the reps who do get through.

A false 'human' also costs more than it looks. The agent's wrap-up time, the wrong disposition ('no answer' instead of 'voicemail') and the retry schedule that follows all inherit the error. On a list of ten thousand numbers, a few percentage points of misreads is hundreds of leads routed wrongly — a quiet drain that rarely appears on a dashboard, because the dashboard trusts the classifier.

Voicemail drop and ringless voicemail: what they fix, what they don't

Voicemail drop is the dialer's answer to the roughly 94% of dials that do not connect: when AMD says 'machine', play a pre-recorded message, free the agent, move on. Ringless voicemail goes a step further and delivers the recording to the mailbox without ringing the phone at all. Both are cheap per message, and both share the same weaknesses: one recording for every lead, no awareness of what the greeting said or whether the beep has sounded, and nobody listening if the person picks up mid-message.

Is leaving a message worth it at all? Gong Labs says yes, in moderation. Across its 300M+ cold-call dataset, leaving one or two voicemails lifted the reply rate to the accompanying email from 2.73% to 5.87%; leaving three or more dropped it to 2.2%, below the no-voicemail baseline. Gong also notes that voicemails rarely earn a callback, that the first should run about 15 seconds and none past 30, and that the time reps spend leaving them would fund about 25% more dials — which is exactly the labour an AI agent removes.

Pew Research Center's 2020 survey of 10,211 US adults explains why the message still matters: 80% said they generally do not answer their mobile when an unknown number calls, but 67% said they let it ring and then check the voicemail. For most lists, the mailbox is the first touch you will actually get — so what is in it, and whether it starts after the beep, is not a detail.

How an AI agent classifies the answer and finds the beep

An AI calling agent inverts the classic design. Instead of a timer that runs before anyone speaks, the agent listens first and classifies the answer from two signals at once: the words — a live transcript of the first seconds — and the audio — energy, cadence, the three-tone intercept signal, the mailbox beep. The output is not a binary human/machine flag but a small set of situations, each with its own behaviour:

• A person. A short greeting, often rising into a question: 'Hello?', 'Yes?', 'Who's this?'. The agent speaks immediately, keeps its turns short and stops the moment it is interrupted.
• A voicemail. 'Hi, you've reached…', 'The person you are calling is not available…', 'leave a message after the tone'. The agent waits through the greeting, detects the beep and only then speaks.
• A silent line. A pickup with no speech at all. A brief prompt, then a disposition and a retry rule rather than a message into nothing.
• A carrier intercept. 'The number you have dialled is not in service.' The number is flagged and taken out of the rotation, so the list gets cleaner with every pass.

Beep detection matters because the message has to start after the tone, not during the greeting. A message that begins early is cut off at the start; one that begins late opens with silence the recipient will skip. An agent that monitors the audio for the tone and keeps listening while it speaks can also stop the instant a person picks up mid-message — a common event on home and office lines — and carry on as a conversation, which no recording can do.

The message itself is written from context rather than recorded once: who the lead is, where the lead came from, what the campaign is about, and a callback number read in digit groups so it can be written down. That is what turns a drop into a voicemail someone might act on, and it is why Gong's finding that reply rates roughly double after one or two voicemails is worth more to an AI campaign than to a drop campaign.

Classic AMD vs voicemail drop vs AI agents

The point-tools column is deliberately mixed. Some AI dialers classify well but hand a connected person to a separate CRM or dialer, and few can bring a human into the same call. The question to ask of any tool is whether the voicemail, the callback it asks for and the hand-off all land in one record — because the follow-up is where the value of a good voicemail is collected.

How LeadPedlar handles machines, beeps and voicemails

LeadPedlar combines an escrow-protected lead marketplace, a CRM, a cloud dialer and AI calling agents in one platform, so machine handling is a campaign setting rather than an integration project. Buy or import a list, write the campaign context and review it with the AI — business name, opening line, goals, the voicemail — set calling hours by the lead's time zone, and choose how calls are handled:

• AI start to finish. The agent dials, listens first, talks to people, leaves voicemails on the beep, books callbacks and respects 'not interested' and do-not-call requests, with the least human intervention. Our walkthrough of unattended campaigns covers the set-up step by step.
• People only. Manual dialing, or autodial for teams with answering-machine handling options, call recording and live monitoring — the dialer feeds the next lead and your reps do the talking.
• Blended. AI first, people when the caller asks hands the call to a person or a SIP extension on request, with a summary of who the lead is and what was said. AI steps in covers voicemails and the moments when nobody on the team is free.

On every answered call the agent classifies the answer — a person, an answering machine or voicemail, a silent line, a carrier intercept — and acts on it. On a mailbox it leaves a personalised message written from the campaign context, or the campaign's own voicemail if you prefer a fixed script, speaking after the beep and stopping if a person picks up mid-message. Phone numbers are read digit-group by digit-group. Every outcome lands in the CRM and in the campaign stats — connected, voicemails left, callbacks booked, transfers, sales — and each stat opens the leads behind it with 'Call with AI' or 'Call myself'. You can listen live to AI calls, read transcripts and recordings, and star calls with notes.

The agent's brain can be the platform's AI or your own connected model; AI calling hours are paid in crypto (see pricing), and the full list is on the features page. For the wider workflow — first-touch calls within minutes and qualification before a person is involved — see agentic AI lead qualification and speed to lead; for how the same agent answers inbound calls, see AI receptionist for inbound calls; and for the dialer categories themselves, AI dialer vs predictive dialer.

Setting up voicemail handling for an outbound campaign

1. 1

   Step 1: Decide what a machine should get

   Nothing (retry later), the campaign's fixed voicemail, or a personalised message. For a cold list, one personalised message on the first mailbox and one more on a later attempt is the pattern Gong's data supports.

2. 2

   Step 2: Write the voicemail into the campaign context

   Name the business, say why you are calling in the lead's terms, give one callback number to be read in digit groups, and keep it under 30 seconds. Leave the pitch for the follow-up email or the callback.

3. 3

   Step 3: Cap attempts per lead

   Two voicemails, then stop: Gong's data shows a third does worse than none. Spread attempts across days and hours within the lead's calling window.

4. 4

   Step 4: Set calling hours and the stop rules

   Calling hours by the lead's time zone, and suppression of leads who said 'not interested' or asked not to be called — both are campaign settings, not agent judgement calls.

5. 5

   Step 5: Choose the mode and the hand-off

   Fully autonomous, people only, or blended; hand-off rules decide who is rung when a person asks for one, and the result of every hand-off is recorded.

6. 6

   Step 6: Listen, then tune

   Review a sample of voicemail recordings and transcripts in the first days: was the message cut at the start? Did the agent stop when someone picked up? Adjust the context, not the thresholds.

Frequently asked questions