Your dialer is hanging up on real people — and you don’t even know it
The system that’s supposed to keep your agents off voicemail is quietly throwing away live prospects. Here’s why heuristic AMD fails in 2026 — and the one question that exposes “AI AMD” that only reads words.
An agent gets a connect. They hear a half-second of nothing and hang up. On the other end, a real prospect heard the same dead air and hung up too. Nobody logs it as a problem. It shows up as a slightly worse connect rate — a number everyone has quietly accepted. This happens thousands of times a day in outbound call centers, and the system responsible is the one that was supposed to help: your answering machine detection.
What AMD is actually doing to your campaigns
Answering Machine Detection (AMD) listens to the first few seconds of an answered call and decides: human, or machine? Between 40–70% of answered outbound calls reach a voicemail, an IVR, a carrier recording, or a disconnected-number intercept. Good AMD keeps your agents off those. Bad AMD does something worse than nothing — it throws away live people while you pay for the privilege.
Traditional AMD hasn’t changed since 2006
The AMD baked into most dialers — including stock Asterisk AMD — is heuristic: count syllables, measure silence, time the speech, listen for a beep. That logic is nearly two decades old. It worked on clean landline audio. In 2026, the average call list is cell phones over compressed VoIP, behind carrier filters and call-screening assistants. Heuristic AMD averages 70–80% accuracy in that environment, and it fails in expensive, invisible ways.
Five ways your AMD is costing you live conversations
- 1. It hangs up on real humans. Heuristic AMD misclassifies live people 5–15% of the time. On 10,000 daily dials, that’s hundreds of prospects a day your “helpful” AMD dropped before an agent ever spoke.
- 2. It doesn’t understand carrier-injected audio. Call-screening assistants and carrier filters answer with synthetic voices. Old AMD hears “machine” and drops the call — while a real person sits there, ready to talk.
- 3. Codec compression breaks the rules. VoIP audio runs through lossy codecs and carrier networks that add artifacts heuristics were never designed for. Accuracy swings wildly (65–90%) by carrier mix and call type.
- 4. It can’t read the greeting. Heuristic AMD doesn’t know the difference between “hello?” and “you’ve reached my voicemail.” It measures sound; it doesn’t understand it.
- 5. Disconnected numbers wreck your caller ID. Carriers return disconnected numbers as recorded intercepts, not error codes. Old AMD keeps dialing them — exactly the pattern carrier analytics flag as an aggressive autodialer. Your numbers get labeled “Spam Likely,” and your answer rate collapses.
The silent-pickup test
Here’s a question worth asking any AMD vendor: what happens when a real person answers and doesn’t say anything yet?
A lot of “AI AMD” is really speech-to-text with a classifier on top — it reads the words the caller says. That works great when someone answers with “hello, who’s this?” But a huge share of real answers start with silence: the person is waiting for you to speak, they’re in a loud room, or they just picked up and haven’t said a word. A word-reading model has nothing to read — so it falls back to “dead air,” and your live prospect gets dropped.
AMDY doesn’t read words. It classifies the acoustic signature of the audio — the sound of a live human picking up a real phone line, distinct from the produced opening of a voicemail or the flatness of true dead air. A person who answers and simply breathes is, acoustically, obviously a person. AMDY connects them. A word-first detector hangs them up.
In 2026, AMD has to classify more than two things
Two buckets — “human” or “machine” — is not enough. AMDY classifies the call the way it actually behaves:
- Live human
- Answering machine / voicemail — including provider-specific patterns (Google Voice, MagicJack)
- Carrier false-answer (FAS) — five distinct carrier patterns that fake a connect
- Honeypot / spam-trap — the trap numbers that get your CID blacklisted
- Disconnected / carrier intercept
- Fax
- Silence / dead air — a distinct class, not a guess
Plus 27 greeting subclasses underneath, because “voicemail” isn’t one sound.
How AMDY does it: an acoustic model that exits early
AMDY doesn’t measure audio — it classifies it. A multi-stage acoustic model exits the instant it’s confident and only listens longer when the call is genuinely ambiguous. Most calls — the easy humans who pick up and say “hello” — are classified in the first quarter-second.
- Sub-200ms decisions at the 95th percentile
- 99% production accuracy
- A
{ classification, human_conf, amd_conf, greeting_class, model_used }response over a plain WebSocket
The honeypot difference
Most “modern AI AMD” pitches skip this: detecting voicemail faster doesn’t protect your caller-ID reputation — and CID reputation is what decides whether anyone answers you at all. AMDY flags honeypot and spam-trap numbers before your dialer burns a CID on them, with reporting and alerting so you can scrub them out. That’s not a nicer voicemail filter. That’s the thing keeping your numbers off the “Spam Likely” list.
What better AMD is actually worth
On 10,000 daily dials, moving from 75% to 99% classification accuracy means thousands of correctly-handled calls a day — agents off voicemail, and live humans that used to get dropped now reaching an agent. Over a year that’s tens of thousands of conversations your team should have had and didn’t. You don’t have to take the math on faith: AMDY logs every detection to a queryable store and ships 23 analytics reports — human rate by IP, area code, state, carrier, and model stage, plus honeypot exposure. See exactly how many real people your dialer was hanging up on, on your own traffic. Run the numbers in the ROI calculator.
It drops into the dialer you already run
AMDY streams answer-audio over a WebSocket and returns the classification in real time. It works with VICIdial, Asterisk, and FreeSWITCH, keeps your existing carrier (no telephony migration), and installs on a Vicidial server with one command in about five minutes.
Try it on your own calls — free
The Sandbox plan is 50,000 detections a month, no credit card, no gateway swap. Run a real campaign for a few days and watch the dashboard fill with your actual human-vs-machine numbers. The first thing it’ll tell you is how many real people your dialer hung up on last month.