Answering machine detection (AMD) is the software that listens to the first seconds of an answered call and decides whether a human or a machine picked up. A dialer uses that verdict to connect a rep, hang up, or wait for the beep and drop a recorded message. On a power dialer, AMD saves time. On a parallel dialer, it decides whether the product works at all.

No detector is perfect, and the two ways it fails cost very different amounts. Here is how AMD works, where it breaks, and how to test it on your own calls before you trust a vendor's accuracy figure.

How answering machine detection works

When a call is answered, the detector analyses the audio for a short window. It looks at patterns that separate a person from a recording:

• Greeting length. A person says "Hello?" or "Allô?" and stops. A recording keeps going for several seconds.
• Cadence and silence. People pause to listen. Recordings rarely do.
• Audio signature. Modern detectors use trained models on the waveform, not only fixed rules about timing.
• The beep. A second detector listens for the tone that marks the start of the recording, so a message drop starts after it and not over the greeting.

The result is a label: human, machine, or unknown. What the dialer does with the label is a separate setting.

Speed against accuracy

Every detector balances two things. Decide fast and you risk guessing wrong. Wait for more audio and you improve the guess but leave the prospect on the line in silence.

That silence is the hidden cost. A prospect who says "Allô?" and hears nothing for two seconds has already started to hang up. In a parallel setup, the delay can be the difference between a conversation and a dropped call that the prospect remembers as a spam call.

So ask for two numbers, not one: the accuracy, and the time between the prospect's first word and the moment the rep is connected.

Two ways AMD fails

False positive: a human classed as a machine

The detector hears a long greeting, a quiet line or an unusual voice and labels the person a machine. The dialer hangs up or drops a message on someone who is listening. This is the expensive error. You lose a real conversation, and you may leave a recorded message on a person who picked up.

The arithmetic is simple. Our 1,000,000-call benchmark in France shows an 11.3% connect rate: about 113 intended contacts per 1,000 dials. If a detector wrongly drops 5% of the humans who answer (an illustrative rate), that is a handful of lost conversations per 1,000 dials, and they are the best ones, since someone actually picked up.

False negative: a machine classed as a human

The detector bridges a voicemail to a rep. The rep hears a greeting, hangs up and loses a few seconds. This error is cheaper per event, but it adds up. In our experience, the ratio of true connects to bridged calls should be well above half. If it falls below, the detector is bridging machines, the list is full of switchboards, or reps are hanging up early.

Where detection struggles in Europe

Most published accuracy figures are measured on one language and one carrier mix. European prospect lines add variety:

• Network-generated messages that do not sound like a personal greeting.
• Switchboards and reception desks that answer with a long company name.
• Greetings in several languages on the same list.
• Call screening on phones, where an assistant answers and asks for your name before the person is on the line. Our piece on iOS call screening explains what it looks like from the dialer side. How a given detector labels it is something to test, not assume.

How to test AMD on your own calls

A vendor demo proves little. Run your own test, using this protocol:

1. Take a sample of about 200 answered calls, spread across your personas, countries and hours.
2. Listen to each recording and label it yourself: human, voicemail, switchboard, screening, other.
3. Compare your label with the detector's label and count both error types.
4. Measure the time from the first spoken word to the rep being connected on the human calls.
5. Split the results by segment: mobile versus landline, switchboards, each language.

Two hundred calls is not a scientific study, but it is enough to see a detector that drops one human in ten. Repeat it after any settings change, and again each quarter. Detectors and carriers both change.

AMD, voicemail drop and compliance

Detection and message drop work together: AMD says a machine answered, the beep detector says the recording has started, and the message plays. Done well, the voicemail costs the rep nothing. We cover what to say in our guide on whether to leave a voicemail and our voicemail templates.

Compliance sits on two edges. Silent or dropped calls to people who did pick up are the kind of thing regulators and prospects complain about, and pre-recorded messages are regulated differently depending on whether the recipient is a consumer or a business. Check the rules for each country you call, and see our map of auto dialer rules in France.

How Sonik approaches answering machine detection

We treat AMD as the feature that makes or breaks parallel dialing, which gives reps 5x more conversations. If the detector fails, the speed advantage turns into hang-ups.

So we measure the outcome and not the label. The numbers we watch are connects against bridged calls and the delay before the rep is on the line. Our connect-rate definitions separate "bridged" from "connected" for this reason. We suggest the same for any dialer you test.

Conclusion

AMD is a probability, not a fact. Judge it on the two errors and the delay, on your own calls. This week, record 200 answered calls, label them by ear and count how many humans the detector got wrong. That single number tells you more than any accuracy claim on a pricing page.