Answering Machine Detection: How It Works and How to Tune It

Answering machine detection (AMD) is software on a dialer that listens to the first seconds of a connected call and decides: live person or machine. Humans go to an agent. Voicemail greetings get a disposition, a drop, or a redial later. Done well, it gives agents back hours of dead air. Done badly, it hangs up on real prospects and quietly distorts your compliance numbers.

This guide covers how AMD works, how accurate it can be, why it is a regulatory filter and not just a productivity feature, and how to audit it in your own campaigns.

TL;DR

• AMD classifies the first 1 to 3 seconds of audio as human or machine. It is a statistical guess, never a certainty.
• The costly error is the false positive: a live person treated as a machine. It wastes a lead and can hide abandoned calls.
• Under FCC and FTC rules, calls correctly identified as machine-answered are excluded from the 3% abandonment cap, so misclassification can flatter your numbers.
• Factory settings rarely fit your lists. Sample recordings, track a one-click voicemail disposition, and retune per campaign.
• AI-voiced outreach and healthcare voicemails carry their own consent and HIPAA limits that AMD does not solve.

Table of Contents

• What answering machine detection does
• How AMD actually decides
• Why AMD is a compliance filter, not just a speed tool
• The two errors, and which one hurts
• How to audit AMD in your own campaigns
• AMD tuning by industry
• Leaving messages and AI voices: consent still applies
• Our perspective: treat AMD as a measured control
• Where RevRing fits
• A short checklist before you switch AMD on
• FAQ
• Sources

What answering machine detection does

When an outbound call connects, the dialer has no idea who picked up. Someone saying “hello” and a recorded greeting both look like audio on the line. AMD analyzes that audio and returns a verdict, usually within a couple of seconds, so the call can be routed.

The outcomes are simple:

• Human: the call is bridged to an available agent immediately.
• Machine: the dialer disconnects, logs a disposition, schedules a callback, or plays a prerecorded message after the beep (where the law and your consent records allow it).
• Unknown or silence: the system falls back to a rule you set, often routing to an agent.

This matters most on predictive and multi-line dialers. When a dialer places several calls per agent, every voicemail it filters out is a call an agent never has to sit through. That is why AMD and predictive dialing are tied together: the pacing algorithm assumes a certain share of answers will be machines. For the pacing side, see our predictive dialing guide for contact center managers.

How AMD actually decides

Most detectors use some mix of the following signals.

Cadence and silence

A person typically answers with a short word or two, then stops and waits. A voicemail greeting runs on for several seconds of continuous speech. Classic detectors measure the length of the initial greeting and the silence that follows. A long unbroken stretch of speech reads as a machine. A short utterance followed by a pause reads as a human.

The weakness is obvious. A person who answers with a long business greeting (“Thank you for calling Smith Dental, this is Dana, how can I help you”) looks like a machine. A short voicemail greeting looks like a person.

Beep and tone detection

Many systems also listen for the beep that ends a greeting. This is more reliable for deciding when to leave a message than for deciding who answered, because it only arrives after the greeting has played.

Machine learning on audio or transcripts

Newer approaches classify the audio with neural networks or transcribe the first seconds and classify the text. Published research shows the gap. One recurrent neural network study notes that traditional methods often reach about 70 to 80% accuracy, while deep learning combined with silence detection can exceed 98%, with one comparison moving from 96.7% to 98.1% when silence detection was added.

Treat those numbers as research results on specific datasets. They are not a promise for your lead lists, your carriers or your regional voicemail greetings.

Why AMD is a compliance filter, not just a speed tool

This is the part most AMD explainers skip.

Under the FCC and FTC rules, an abandoned call is one where a person answers and the caller does not connect them to a representative within two seconds of the person’s completed greeting. Telemarketers must keep abandoned calls to no more than 3% of live-answered calls, measured per campaign over a 30-day period. Calls correctly identified as machine-answered fall outside that count.

Now follow the math. The denominator of your abandonment rate is calls answered by a live person. If AMD wrongly labels a human as a machine and hangs up:

1. The prospect hears silence or a dead line.
2. The call is logged as a machine, not an abandon.
3. The denominator shrinks and the numerator misses the call.

Your report can show a clean 1.5% while your true rate, counting those mishandled humans, is well above the cap. Default dialer logs will not show you this. The regulator does not care what the log says. It cares what happened on the call.

For the full rule set, see our 3% ceiling checklist for abandoned calls.

The two errors, and which one hurts

The first error is annoying and measurable. The second is expensive and invisible unless you go looking. Most teams should tune to avoid false positives, even if that lets more voicemails reach agents.

A practical rule: when the cost of a lost live contact is high, as with a Medicare lead you paid for or a legal intake inquiry, accept a few more voicemails reaching agents. When leads are cheap and volume is huge, you can run more aggressively, but still audit.

How to audit AMD in your own campaigns

No detector is perfect. Telecom practitioners consistently say AMD is a heuristic guess, and that the best strategy is to measure it on your own traffic, because stock settings rarely match how your lists behave. Here is a workable audit.

Step 1: Give agents a one-click voicemail disposition

If an agent hears a greeting, one click should mark it as voicemail. Specialists recommend tracking the percentage of these dispositions per list per week. A sudden jump means AMD is letting more machines through. A sudden drop can mean it is misclassifying in the other direction or that a list has changed.

Step 2: Sample recordings, not summaries

Summary reports tell you what the system decided. Recordings tell you whether it was right. Each week:

• Pull a random sample of calls the dialer labeled machine, across each campaign and each carrier route.
• Listen to the first five seconds of each.
• Count how many were actually people.
• Do the same for a sample labeled human, to count voicemails that reached agents.

If you handle recordings, check the consent rules first. Our call recording compliance guide covers the state-by-state issues.

Step 3: Recalculate your abandonment rate

Add the live humans you found in the machine sample back into the denominator and numerator, scaled to the full volume. If the adjusted rate moves toward 3%, loosen the detector.

Step 4: Tune one variable at a time

Typical settings include the maximum greeting length, the silence threshold, and the analysis window. Change one, run it for a week, and compare. Changing three at once tells you nothing.

Step 5: Segment by list source and time of day

Aging lists, mobile-heavy lists and business lists behave differently. Mobile carriers’ call screening and visual voicemail add odd audio patterns. A setting that works for a fresh internet lead list may fail on a recycled one. Run separate thresholds where your dialer allows.

Teams that already run structured monitoring will recognize this loop. A marketing automation checklist uses the same logic of measuring, reviewing and adjusting on a schedule, applied to a different channel.

AMD tuning by industry

Insurance

Insurance outreach lives under TCPA rules, with consent records for every number. Final expense and Medicare agents often call older consumers who answer slowly, with a hesitant “hello?” and a pause. That pattern can look like silence, so test your detector on exactly those calls. Our TCPA walkthrough for agents is here.

Healthcare

A missed human is a missed appointment confirmation. And a message left on the wrong machine is a privacy risk. HHS says the HIPAA Privacy Rule permits providers to leave messages on answering machines, but they must reasonably safeguard patient privacy. In practice, that means limiting the message to minimal information such as the practice name, a callback number and an appointment reminder, with no diagnosis, test names or medication details.

Any cloud platform that handles ePHI on your behalf needs a Business Associate Agreement. Details are in our guides to HIPAA compliant outreach and the 5 step BAA checklist.

Real estate and lead generation

Speed to lead is everything. A false positive on a fresh inquiry can cost the deal. Real estate teams on multi-line dialers should favor conservative detection and watch talk time per agent hour, not just dials. Lead generation operators who route live calls to buyers should check their detection before paying for transfers that turn out to be voicemails.

Legal intake

Intake calls are often returns of an inquiry the prospect just made. They are expecting the call, so they answer fast and say little. Audit those greetings, and keep a clear callback path for voicemails so a prospect is not lost to a single missed attempt.

Leaving messages and AI voices: consent still applies

AMD tells you a machine answered. It does not tell you that you may leave a recording. A prerecorded voicemail drop is a prerecorded message under the TCPA, and it needs the right prior consent for the type of call. Our express written consent playbook explains what that consent must contain.

The FCC has also confirmed that the TCPA applies to AI technologies that generate human voices. If your outreach uses an AI voice agent that handles a call after AMD decides it is a human, it needs prior express consent as well. Better detection does not change that. For more on rules affecting AI in regulated calling, read AI compliance monitoring: 4 U.S. rules for regulated call teams.

Caller identity matters too. If the voicemail you leave or the call you place shows a spoofed or mismatched number, you add a separate risk. See U.S. neighbor spoofing laws.

Our perspective: treat AMD as a measured control

We build dialers for regulated sales and intake teams, so we see the same pattern in agencies of every size. They buy a dialer, leave AMD on its default, and look only at the headline numbers. Talk time goes up, the abandon rate looks fine, and nobody listens to the calls the system threw away.

Our view is simple.

• AMD is a control, not a feature. Like any control in a regulated process, it needs testing, sampling and a record that you did it.
• Compliance and productivity pull in opposite directions at the edges. Aggressive settings raise agent occupancy. Conservative settings protect live contacts and your abandon rate. Decide on purpose where you sit, per campaign.
• Detection belongs next to consent, scripts and reporting. An accurate verdict is worth little if the message you leave breaks a consent rule or a HIPAA limit. That is why we put TCPA, DNC and HIPAA BAA infrastructure in the same platform as the dialer, with workflows built for each industry rather than one generic setup.
• Your CRM should not have to change. We connect to the CRM you already use, so dispositions, callbacks and voicemail rates flow into the reports your managers read.

If you take one habit from this guide, make it the weekly recording sample. It takes under an hour and tells you more than any dashboard.

— Marc

Where RevRing fits

RevRing is a cloud platform that brings predictive dialers (single-, multi- and unlimited-line), blended inbound and outbound calling, AI call scoring, compliance infrastructure and CRM connectivity into one system, with workflows for insurance, real estate, healthcare, legal and customer service teams. Plans start at 39.99 USD per month per seat on Starter, 59.99 USD on Scale and 89.99 USD on Pro, with telephony and other usage-based items billed on top. See the pricing page for the current details, or see how it works before you talk to us.

A short checklist before you switch AMD on

• Confirm consent records exist for every number in the campaign.
• Set a conservative detection profile for the first week.
• Add a one-click voicemail disposition for agents.
• Schedule a weekly sample of machine-labeled recordings.
• Recalculate abandonment with misclassified humans included.
• Write voicemail scripts that follow your industry’s limits, such as HIPAA’s minimum necessary information.
• Keep a log of tuning changes, with dates and results.

FAQ

How does answering machine detection work?

AMD analyzes the first seconds of audio after a call connects. It looks at how long the greeting runs, how much silence follows, and in newer systems the sound and words themselves, using trained models. It then returns human, machine or unknown so the dialer can route the call.

How accurate is answering machine detection?

It varies widely. Research on traditional cadence-based methods puts accuracy around 70 to 80%, while deep learning approaches report above 98% on test data. No detector is perfect on real lists, so measure it on your own calls.

Can AMD cause TCPA problems?

Yes, indirectly. If it mislabels live people as machines, those calls drop out of the abandonment calculation, which can make a campaign look compliant when it is not. Leaving prerecorded messages on machines also requires proper consent.

Is a voicemail the same as an answering machine?

Functionally, for dialing, yes. Both are recorded greetings that take a message instead of a person. Dialers treat carrier voicemail and physical answering machines the same way, though greeting patterns differ.

Can I leave a healthcare voicemail after AMD detects a machine?

Often yes, with limits. HHS says HIPAA allows it if you reasonably safeguard privacy, which generally means a practice name, callback number and minimal appointment information. Check your own policies and your compliance officer before scripting messages.

Sources

• FCC order on TCPA rules, including the abandoned call definition
• A Recurrent Neural Network Approach to the Answering Machine Detection Problem (arXiv)
• HHS: may providers leave messages on answering machines
• FCC: TCPA applies to AI technologies that generate human voices
• AL Performance: answering machine detection tuning