Picture a customer in a kitchen, phone buzzing on the counter. She picks up, says "Hello?" and hears nothing. One second, two seconds, three. She says "Hello?" again, then hangs up. On your side, the dialer was still deciding whether she was a person or a voicemail box. That pause is the price of answering machine detection, and most teams never measure it. They switch AMD on, see fewer wasted agent minutes, and move on. This article explains what happens inside those few seconds, what the silence costs you, and how to set it up differently for each queue.
What AMD actually listens to
When a call is answered, the dialer does not know who or what picked up. Answering machine detection analyzes the live audio and the signaling around it to put the call into one of a few buckets: a human, a voicemail, a number that is unavailable, or an invalid number.
The logic is simpler than it sounds. A person usually says something short: "Hello?", "Yes?", "Dental clinic, good morning." Then they stop and wait. A voicemail greeting is a long, continuous stretch of speech, often recorded in a calm, even tone, and it ends with a beep. Carrier messages ("the number you have dialed is not in service") and special tones sit in a different category again, and they point to unavailable or invalid numbers rather than to a person.
So the detector is looking at things like how long the first utterance lasts, how much silence follows it, whether the speech keeps going, and whether a tone appears at the end. It is a judgment from patterns, not a certainty. And like any judgment, it gets better the more time it has.
The trade-off: time versus accuracy
Here is the core tension. To tell "Hello?" from the start of a long greeting, the detector has to listen for a while. If you give it very little time, it will guess early and guess wrong more often. If you give it more time, it will be more accurate, but the human on the other end waits in silence for longer.
There are two kinds of mistakes, and they cost different things.
- A human classified as a machine. The call is dropped or sent to a voicemail routine. You just lost a real conversation with a real person who was ready to talk. In a sales queue, this is the expensive error.
- A machine classified as a human. The call is passed to an agent, or to the AI voice agent, which starts speaking into a voicemail. You lose a few seconds of capacity and possibly leave an awkward half message.
Most teams fear the second error because it is visible: an operator complains that they keep hearing beeps. The first error is silent. Nobody complains about the customer who never got a call. That is why it deserves more attention than it usually gets.
What the silence costs
The silence at the start of a call is not neutral. People who answer an unknown number and hear nothing tend to assume it is a robocall or a scam, and they hang up. Even those who stay are already a little irritated by the time you start speaking.
Imagine a queue that makes a few thousand calls a day. If the detector holds the line for an extra second or two on every answered call, that is a few seconds multiplied by every person who picked up. Some fraction of them will hang up in that gap. You will never see the exact number in a report, because those calls look like "answered, short, no outcome." But your conversion rate on that queue will quietly carry the loss.
There is a second, regulatory side. Many markets have rules about silent or abandoned calls, and the details differ by country and by campaign type. Check the rules that apply to you with whoever handles compliance, and treat the detection delay as something that belongs in that conversation.
Tune it per queue, not per account
One global AMD setting is convenient and almost always wrong. Different queues call different people, at different times, in different ways. The same time limit that works for a consumer reactivation campaign will misfire in a B2B queue where the first thing you hear is a receptionist reading a full company greeting.
Some examples of how the queue changes the right setting:
- B2B calls to company numbers. Business greetings are long: "Good morning, you have reached Kowalski Logistics, how can I help you?" That sounds a lot like a recorded message. A tighter limit will misclassify these as machines. Give the detector more room, and accept slightly more silence.
- B2C evening calls to mobiles. People answer with one word. Short greetings are easy to recognize, so a shorter limit works well and keeps the call feeling natural.
- Warm callbacks. Someone left a request or abandoned a cart an hour ago. They expect your call. Here the cost of losing a human is highest, so favor accuracy over speed.
- Cold lists with old numbers. Expect many voicemails and many dead numbers. Faster detection saves capacity, and the retry ladder will give each number another chance later.
On a platform like Alphacall, the time limit for the decision is set per queue, next to the calling windows, retry ladders and attempt limits. That is the right place for it, because the detection setting only makes sense together with what you do after each outcome.
Decide what happens after each verdict
Detection is only half of the setup. The other half is what the system does with the answer.
For a human, the call goes to an operator or to the AI agent immediately, with no extra greeting delay. For voicemail, you have a choice: hang up and put the contact back on the retry ladder for a different time of day, or send an SMS that says who called and why. For unavailable, retry later within your attempt limit. For an invalid number, remove it from the list so you stop paying for calls that cannot succeed.
Be careful with leaving voice messages. A scripted message left on every voicemail can feel intrusive, and in some queues it reduces callbacks rather than increasing them. Test it on one queue before applying it everywhere.
A practical tuning checklist
- Split your queues by audience: business numbers, consumer mobiles, warm callbacks, cold lists.
- Pull a sample of calls classified as voicemail and listen to them. Count how many were actually people.
- Pull a sample of calls passed to agents and count how many were machines.
- Compare the two errors. In a sales queue, lean towards fewer lost humans, even at the cost of a few more beeps.
- Change the time limit for one queue at a time, by a small step, and keep the rest as a control.
- Watch the short-call rate and the conversion rate, not only the "machine detected" rate.
- Set a clear action for each verdict: connect, retry, SMS, or remove.
- Review again after a few weeks, because greetings and carrier behavior change.
Where to start
You do not need a perfect detector. You need one that is tuned to the people each queue actually calls, with a known price in silence that you have chosen on purpose instead of inherited from a default. Start with one queue, listen to real recordings, and move the time limit in small steps.
If you would like to hear how this sounds in practice, with the detection decision, the handover to a person and the AI agent speaking on a real call, you can ask Alphacall for a live call demo and judge the first three seconds yourself.