Best Answering Machine Detection Software for Call Centers (2026)

What Is Answering Machine Detection Software?

Answering machine detection (AMD) is the software that decides, in the first moments after an outbound call is answered, whether a live human or a machine picked up, so the dialer can route the call accordingly. If a human answered, the call goes to an agent before the prospect hangs up. If a voicemail, IVR, or recording answered, the dialer can drop a pre-recorded message, hang up, and move the agent to the next call instead.

Predictive dialers only work because of this decision. The dialer calls ahead of agent availability, which means some answered calls have no agent ready, and every answered call needs a fast verdict about what is on the other end. The quality of that verdict determines how much of your agents’ connected time is spent talking to people, and how much is spent listening to voicemail greetings and dead air.

Why Call Centers Need Better AMD in 2026

The raw mix of what answers an outbound call is brutal. Across roughly 2.3 billion answered calls per month on the AMDY network, only 12.5% were a live human. Around three quarters were machines: voicemail, IVRs, carrier recordings. Another 14% were carrier false answers, where the network signals “connected” even though nobody picked up. See the full network data →

That mix is exactly what the old detection technology was not built for. The heuristic AMD inside Asterisk was tuned for landline audio in the early 2000s: long human greetings, predictable silence patterns, analog answering machine beeps. Today’s traffic is compressed mobile VoIP, short clipped “hello?” answers, custom carrier greetings, iOS call screening, and carriers that fake the answer signal outright. Detection that counts syllables and measures silence misreads all of it, which is why stock AMD tops out around 70 to 85% accuracy in production and returns NOTSURE on the hardest calls.

The Three Approaches to AMD: Dialer-Native, Transcription-Based, and Acoustic AI

Every AMD product on the market falls into one of three technology families.

1. Dialer-native heuristic AMD

The AMD built into Asterisk (app_amd, used by ViciDial and GoAutoDial) and FreeSWITCH (mod_avmd) works on timing rules: how long the first audio lasts, how much silence follows, whether a beep tone appears, whether speech keeps going past greeting length. It is free and already installed. The ceiling is well documented: roughly 70 to 85% accuracy, and a NOTSURE gray zone on fast human greetings that the dialer has to route by guess. How NOTSURE drops your fastest answerers →

2. Transcription-based AMD

These services transcribe the answer audio and classify the words: greeting phrases mean voicemail, a question means a human. They read greetings well, but they need speech to work with. They are slow by construction (the system waits for words), bound to the languages they were trained on, and blind to a person who answers in silence or with background noise. On a live predictive dial that split-second matters. The case against transcribing every call →

3. Acoustic AI AMD

The newest generation classifies the sound itself: an ML model trained on the acoustic signature of what answered the phone. No transcription, no language dependency, and the decision arrives in milliseconds instead of seconds. Because it analyzes the signal rather than the words, it can also classify things word-based systems cannot see: carrier false answers, honeypot traps, fax tones, dead air.

A deeper technical breakdown, including FreeSWITCH specifics: FreeSWITCH vs Asterisk vs AI AMD →

What to Look for in Answering Machine Detection Software

Use the same criteria list on every vendor you evaluate:

• Accuracy with the errors split out. Ask for the false positive rate on humans and the false negative rate on machines separately. One headline number hides which mistakes the system makes.
• Decision latency, measured. Median detection time at production volume, not the best-case demo number. Accuracy at that latency, not accuracy given unlimited time.
• More than two buckets. Can it see carrier false answers, honeypots, disconnected numbers, silence? Two-class AMD is blind to roughly one in seven of your “answers.”
• Integration without migration. Does it work with your current dialer and carrier, or does it require moving your traffic?
• A queryable detection log. You cannot improve what you cannot audit. Every classification should be logged and reviewable.
• Transparent pricing. Included volume, overage rate, and what happens when you exceed it. Watch for uncapped per-detection billing.
• A free way to test on your own traffic. Any vendor confident in their accuracy will let you pilot it.

Best Answering Machine Detection Software for Call Centers: 2026 Shortlist

Full disclosure: this is our blog, and AMDY is our product. The section after this one tells you where we fit and where we do not. Here is the honest landscape, organized by category, because the right choice depends on what you already run.

Your dialer’s built-in AMD (the free baseline)

Every ViciDial, GoAutoDial, and FreeSWITCH install ships with heuristic AMD. It costs nothing and takes zero integration work, which makes it the default for low-volume operations. The tradeoff is the 70 to 85% ceiling, the NOTSURE gray zone, and no visibility into carrier false answers or honeypots. If your connect rates are acceptable and your lists are clean, tuning it further (see every AMD setting explained) may be all you need.

Voice API providers with AMD add-ons

Twilio, Telnyx, Plivo, Bandwidth, and Vonage all offer machine detection as an add-on to their voice APIs, billed per detection on top of per-minute charges. Published rates we have verified: Twilio’s Enhanced AMD runs about $0.0075 per answered call, Telnyx Premium AMD about $0.0065 per invocation, and Plivo machine detection around $0.004 per request; some providers include a basic detector at no extra charge. The catch: detection runs in their network, so these make the most sense when your calls already travel on that provider’s platform.

Contact center platforms with bundled AMD

Large contact center suites like Genesys bundle machine detection into the platform. You configure it through admin settings (detection speed, when to disconnect) and pay for it inside a per-seat subscription. If you are running your entire operation on one of these platforms, that bundled AMD is the path of least resistance. It is not something you bolt onto an existing ViciDial deployment, and published accuracy benchmarks are generally not part of the deal.

Standalone AI AMD services

A growing category of acoustic AI classifiers that attach to your existing stack, usually by forking the answer audio over a WebSocket or SIP and returning a classification in real time. AMDY is one; the field also includes newer entrants chasing the same ViciDial upgrade story. Vendor accuracy claims in this space commonly run 95 to 99%. Treat every claim, including ours, as a hypothesis and verify it on your own traffic with the pilot checklist at the end of this guide.

Where AMDY Fits, and Where It Doesn’t

AMDY is an acoustic AI classifier built for outbound dialing operations that already run Asterisk-based or SIP-based dialers and want better detection without migrating. The verified facts:

• 99% accuracy in production.
• Classification in under 200 ms, with detection starting in 1/8 of a second.
• Acoustic signature model: it classifies sound, not words. Language independent, and it handles silent pickups.
• Seven detection classes: live human, voicemail, carrier false answer (five distinct patterns), honeypot/spam trap, disconnected/intercept, fax, and silence.
• ~5 minute install at the Asterisk layer on a ViciDial server, no dialer UI changes. Also supports GoAutoDial, FreePBX, and FreeSWITCH.

Where AMDY does not fit, honestly:

• It is not a dialer, not a carrier, and not a voice API. It classifies the answer; your existing stack does the dialing.
• If your whole operation already runs on a platform suite and its bundled detection is good enough for you, adding a standalone classifier buys you little.
• If your volume is tiny and your lists are clean, the free heuristic AMD may serve you fine. We wrote about that honestly too: default AMD vs AMDY, head to head →

Accuracy and Speed Benchmarks: False Positives, False Negatives, and False Answers

Headline accuracy is a marketing number until you split it into its two error types, because they cost very differently:

• False positive (human flagged as machine). The expensive one. A live prospect gets dropped or hears dead air. You lose the lead, and under TCPA rules the drop can count against your abandoned-call rate.
• False negative (machine passed as human). An agent’s time burned on a voicemail greeting. Costly in wages and morale, rarely a compliance event.

The third error dimension most comparisons ignore is the false answer. When the carrier signals “connected” on a call that never connected, a two-bucket detector has no category for it and guesses. On the AMDY network, about 14% of answered calls fall in this bucket. What FAS costs you, twice →

On speed: most operations want the verdict within one to three seconds of answer, and the TCPA connect window makes anything slower than that actively risky (more below). But there is a hard floor on how fast you can go with weak signals, and humans announce themselves quickly anyway. Across the AMDY network, 17.9% of human answers were identifiable in under half a second, and the large majority within about a second and a half. A detector that needs a full sentence of speech is, by construction, last to arrive on the fastest, most valuable humans.

Whatever a vendor claims, benchmark it the same way: a manual audit of real answered calls, labeled by a human, compared against the software’s verdicts, plus an A/B split of live traffic. Step-by-step: how to test AMD accuracy before going live →

AMD Pricing: Per-Detection Fees, Bundled Plans, and Subscriptions

Three pricing models dominate, and they behave very differently as your volume grows.

1. Per-detection fees on voice APIs

Twilio charges about $0.0075 per answered call for Enhanced AMD, Telnyx Premium about $0.0065 per invocation, Plivo around $0.004 per request, always on top of per-minute voice charges. The math is linear and uncapped: double your volume, double the detection bill. At a million answered calls a month, that is $4,000 to $7,500 in detection fees alone.

2. Bundled into dialer or platform subscriptions

Heuristic AMD is “free” inside ViciDial and Asterisk; platform suites fold it into per-seat pricing. The cost does not disappear, it moves off the invoice: dropped live humans, idle agents, and FAS billing you never see. The real cost of free AMD, totaled up →

3. Monthly subscriptions with included detections

Standalone AI AMD services, AMDY included, sell a monthly base with a detection allowance and a flat overage rate. AMDY’s plans:

• Sandbox: free. 50,000 detections a month, hard cap, no card. Built for pilots.
• Starter: $79/month. 500,000 detections included, then $0.00025 per detection.
• Growth: $299/month. 5 million included, then $0.00015 per detection.
• Scale: $999/month. 25 million included, then $0.00010 per detection.

A buyer’s note: at call-center volumes, the detection fee is usually small next to the agent time a better verdict recovers. Compare total cost against recovered agent hours, not against the per-detection price in isolation. Current plan details live on the pricing page.

Integration With Vicidial, Asterisk, and FreeSWITCH

There are three integration paths, matching the three technology families above.

• Tune what you have. Heuristic AMD lives in amd.conf and dialer variables. Tuning squeezes out a few points of accuracy inside the same 70 to 85% ceiling. Every ViciDial AMD variable, explained →
• Move your traffic to a voice API. Detection rides along with the calls, which means migrating dialing onto that provider’s network. Fine if you were planning to move anyway; a large project otherwise.
• Attach an AI classifier to your existing stack. The answer audio is forked to the service, which returns a classification over WebSocket or API. AMDY installs at the Asterisk layer with a small integration script, about five minutes on a ViciDial server, no changes to the dialer UI, and your carrier stays put. The same pattern on GoAutoDial →

If your dialer speaks SIP, the attach path almost always exists. AMDY works with ViciDial, Asterisk, FreePBX, GoAutoDial, FreeSWITCH, and any SIP-based predictive dialer through its WebSocket API.

AMD and TCPA Compliance: Abandon Rates and the 2-Second Window

AMD is not mandated by the TCPA, but it directly drives the numbers the rules cap. The relevant FCC rules for predictive dialers:

• Abandoned calls are capped at 3% of calls answered by a person over any 30-day period.
• When a person answers, a live agent must be connected within 2 seconds of the greeting ending, or the call counts as abandoned.
• A call disconnected on reaching an answering machine is not an abandoned call.
• Violations carry statutory damages of $500 to $1,500 each.

Inaccurate AMD pushes you into violations from two directions at once. A false machine verdict on a real human produces dead air or a dropped call, which is exactly what the 3% cap measures. A slow verdict blows the 2-second connect window even when the classification is eventually right. Accurate, fast detection is therefore a practical compliance safeguard: it keeps the abandon numerator small without making you drop real people. The deeper mechanics: the 3% rule and how AMD accuracy keeps you compliant → and how most operators compute abandonment wrong →

One caution: AMD is one control in a compliance program, not the whole program. Consent management, DNC scrubbing, and calling-hour rules still sit on you.

How to Evaluate AMD: A Pilot Checklist

Run any candidate, including ours, through the same seven steps:

1. Capture your baseline. Human connect rate, abandonment rate, NOTSURE or unknown rate, agent idle time on machines.
2. Audit manually. Pull about 50 answered calls, have a human label what actually answered, and compare against your current AMD’s verdicts.
3. Split live traffic. Run the candidate on a slice of campaigns at the same hours as a control group, so the comparison is apples to apples.
4. Measure latency and accuracy together. Ask each vendor for median decision time and accuracy at that time, plus false positive and false negative rates separately.
5. Ask about false answers. How does the product detect carrier FAS? If the answer is “we don’t,” you now know roughly 14% of your answers are invisible to it.
6. Check the logs. You want every classification stored and queryable, so your own team can verify rather than trust.
7. Start free. Use a sandbox or pilot tier before any contract. AMDY’s is 50,000 detections a month with a hard cap and no card.

Judge the result on cost per recovered agent hour, not on the detection fee. That is the number that pays for the software.

Frequently Asked Questions

How accurate is answering machine detection software?

Accuracy varies widely by technology. Legacy rule-based AMD built into dialers like Asterisk typically reaches around 70 to 80 percent in practice, missing custom greetings and short human hellos. Modern AI-driven systems commonly claim 95 to 99 percent accuracy. When comparing vendors, ask for both the false positive rate (humans flagged as machines) and the false negative rate (voicemails passed to agents), since the two errors have very different costs.

How fast does AMD need to classify a call?

Most operations want a decision within one to three seconds of answer. Under TCPA and FCC rules, a predictive dialer must connect a live agent within two seconds of the called party finishing their greeting, or the call counts as abandoned. Slow detection wastes agent time and raises abandon-rate exposure, while overly fast detection on weak signals causes false positives. Ask vendors for both median detection latency and their accuracy at that latency.

How much does answering machine detection cost?

Pricing falls into three buckets. Voice APIs charge per detection, roughly $0.004 per request at Plivo up to about $0.0075 at Twilio, on top of minute charges. Dialer platforms bundle AMD into per-seat subscriptions. Standalone AI AMD services like AMDY use monthly subscriptions with included detection volumes, starting at $79 per month for 500,000 detections, plus small per-detection overage. For high-volume centers, recovered agent time matters more than detection fees.

Does AMD software work with Vicidial, Asterisk, and FreeSWITCH?

Yes, through three common paths. Dialer-native AMD uses Asterisk’s built-in call progress analysis, which is free but less accurate. Voice API providers sit in the SIP path and return detection results via webhooks. Standalone AI services like AMDY connect through a small integration script or WebSocket API and work with Vicidial, Asterisk, FreePBX, GoAutoDial, FreeSWITCH, and most SIP-based predictive dialers without replacing your existing stack.

Is answering machine detection required for TCPA compliance?

AMD itself is not mandated by the TCPA, but it directly affects compliance. FCC rules cap abandoned calls at 3 percent of calls answered by a person, and a call dropped on reaching an answering machine does not count as abandoned. Inaccurate AMD causes two problems: dead air after false machine verdicts, and misclassified humans that inflate your abandon rate. Accurate, fast AMD is a practical compliance safeguard, not a luxury.

What are false answers (FAS) and does AMD catch them?

False answer supervision is when a carrier or network signals that a call connected when it actually rang, hit a busy tone, or reached a disconnected number. You get billed for the fake connection and your dialer wastes cycles. Classic cadence-based AMD cannot reliably catch FAS. AI acoustic detection analyzes the actual audio, so it can flag these events. Ask vendors specifically how they detect false answers before you buy.

What is the difference between AMD and voicemail drop?

They solve different problems together. AMD decides whether a call reached a human or a machine and routes accordingly. Voicemail drop plays a pre-recorded message when AMD identifies a machine, then moves the agent to the next call instead of leaving a manual voicemail. Voicemail drop depends on accurate AMD: if the detector misclassifies a human as a machine, your recording plays to a live person, which can violate telemarketing rules.

Can AI voice agents use answering machine detection?

Yes, and it is one of the fastest-growing use cases. AI agents still burn per-minute fees when they talk into voicemail. AMD classifies the call in the first seconds and either hands live calls to the AI agent or triggers a voicemail flow. The same accuracy criteria apply: low latency, low false positives, and the ability to detect custom greetings and iOS call screening before the agent or bot starts speaking.