Answering-machine detection (AMD) listens to the first seconds of an answered call and decides whether a person or a recorded greeting picked up, so a predictive dialer only passes live conversations to agents. It saves agents a lot of wasted time, but it has costs: the billed seconds spent detecting, the short pause a real person hears while the system decides, and the live calls it gets wrong.
Depending on the list and the time of day, many answered calls reach a voicemail rather than a person: phones switched off, out of coverage, or set to send unknown numbers straight to the mailbox. Without detection, a predictive dialer passes those calls to agents, who listen to a greeting, hang up, select a disposition and wait for the next call. Over a shift, that adds up to a lot of time spent listening to recordings.
AMD moves that work from the agent to the dialer. The question is how well it does it, and what the mistakes cost.
How answering-machine detection works
When the call is answered, the detector analyzes the audio before deciding where to send it. The signals are simple:
- The first words. People tend to answer with a short greeting, then pause and wait for a reply. Recorded greetings tend to be longer and run on without leaving room for the caller.
- Silence. A long silence at the start, or a particular rhythm of speech and silence, shifts the decision.
- The beep. A tone at the end of a greeting is a strong sign of a machine, though by then the decision comes late.
The system needs a moment to hear enough, usually a few seconds at most. Each call ends up classified as a person, a machine or uncertain. A person goes to an agent and a machine is handled by the campaign's rules. Every system has to decide what to do with the uncertain calls, and that choice pushes errors one way or the other.
The two ways it gets things wrong
No detector is perfect, and its errors run in two directions:
| Error | What happens | Who pays |
|---|---|---|
| False positive: a person classified as a machine | The person answers, hears nothing, then the line drops | The person, your number's reputation and the lost conversation |
| False negative: a machine classified as a person | An agent receives a voicemail greeting | Agent time |
Settings that catch more machines usually hang up on more people, and the reverse. There is no setting without errors, only a balance you choose.
False positives deserve the most attention. From the person's side they are silent calls, the experience that drives spam reports and complaints. Some regulators treat them as abandoned calls: in the UK, for example, Ofcom expects call centers to include an estimate of AMD false positives in their abandoned call rate. Check the rules that apply in the countries you call.
Then there is the pause. Even when detection is right, the person who said hello hears a short silence before an agent comes on, and some hang up. Keeping that gap short, and training agents to open quickly and clearly, matters as much as the detection itself.
What AMD costs you
Four costs, some visible on the invoice and some not:
- Billed seconds. A machine-answered call is answered from the network's point of view, so it is billed from pickup until hang-up. With per-second billing, that is a few seconds per machine; with 60/60 billing, every one costs a full minute. See per-second vs 60/60 billing.
- Lost conversations. Every false positive is a person who answered and never reached you.
- Reputation. Silent calls lead to reports that can get your numbers flagged. More in why your caller ID gets flagged as spam.
- Agent time on false negatives. Much smaller than without AMD, but not zero.
Against those costs sits the time AMD gives back to agents.
A worked example for one agent
Assumptions, per dialing hour for one agent: 20 calls answered by machines and 20 by people. Without AMD, an agent needs about 15 seconds to recognize a machine, hang up and set a disposition. With AMD, detection and hang-up take about 5 seconds. For illustration, we assume AMD gets 1 call in 20 wrong in each direction.
| Per dialing hour | AMD off | AMD on |
|---|---|---|
| Machine calls reaching an agent | 20 | 1 |
| Agent time spent on machines | 5 minutes | 15 seconds |
| Billed time on machine calls, per second | 5 minutes | about 2 minutes |
| People hung up on by mistake | 0 | 1 |
Over a month of 126 dialing hours (6 hours a day for 21 days), AMD gives this agent back roughly 10 hours of talk capacity and saves about 6.5 hours of billed time on machines. In exchange, about 126 people who answered are hung up on.
Whether that is a good trade depends on the list. On a large consumer list where each contact is one of thousands, the agent hours usually win. On a small B2B list where each contact is a named prospect, losing 126 of them a month may cost more than the time saved.
When to switch AMD off
- B2B lists. Switchboards, receptionists and automated attendants often answer with a scripted greeting that looks like a recording to a detector.
- Small, high-value lists. When each contact matters, let agents hear every answer.
- Agreed callbacks. A person expecting your call should not be dropped by mistake.
- Calls where a message matters. If agents need to leave a voicemail, use preview mode so they can.
- Testing a new list. Run a sample with AMD off to see how many machines and people the list produces before you tune the campaign.
How to measure AMD on your campaigns
Track a few figures per campaign and per list:
- Machine rate = calls classified as machines ÷ answered calls. A sudden change on the same list suggests a detection or network issue rather than a change in your contacts.
- Machines reaching agents. Give agents a dedicated disposition for voicemail greetings. That count estimates your false negatives.
- Callbacks after a machine result. People who were hung up on sometimes call the number back within minutes. Calls from numbers that were just classified as machines are a rough signal of false positives.
- Outbound abandon rate = live answers not connected to an agent ÷ live answers. Watch it alongside the machine rate, since both shape what people experience.
- Hello, hello. If agents keep connecting to people already repeating their greeting, the gap before connection is too long.
Hang up or leave a message?
Some dialers can play a recorded message to a machine, often called voicemail drop. OVOKY does not: when it detects a machine on a predictive campaign, it hangs up and logs the attempt against the contact so you can plan the next try. If a message matters for a campaign, run it in preview mode so agents leave it themselves, or consider an SMS step, which OVOKY campaigns can send immediately, on a schedule or as a drip.
How OVOKY handles answering machines
- Answering-machine detection is on by default for predictive campaigns and can be switched off per campaign.
- Machines are hung up and the attempt is logged, so agents only receive calls a person answered.
- Voice is billed per second with no minimum, so each machine call costs only the seconds it lasted.
- The Predictive Efficiency report shows dial ratio, machine rate and abandon rate, and supervisors can change the dial ratio mid-campaign.
- The Best Time To Call report shows connect rates by hour and weekday, so you can dial when people tend to answer.
More on the predictive dialer and on OVOKY reports.