Why VICIdial AMD Flags Real People as Machines and How to Fix It

A false positive (AMD false positive) is the most expensive mistake answering machine detection (AMD (answering machine detection)) can make. It is when AMD decides a real, live person is an answering machine and either hangs up on them or plays a recorded message at them. You paid to reach that person, they picked up, and the system threw them away. Here is why it happens and what to change.

Why a human gets flagged as a machine

AMD listens to the first moments of a call and judges whether the audio sounds like a person saying hello or a recorded greeting. People who give a long greeting fool it. Someone who answers with "hello, this is the Johnson residence, how can I help you" sounds a lot like a voicemail prompt. So do people in noisy rooms, people on speakerphone, and people who pause oddly before speaking.

Telecom conditions add to it. A live answer (Live answer) that arrives with extra carrier (Carrier) delay, a clipped first syllable, or background tones can read as a machine. None of this means AMD is broken. It means the thresholds are set tighter than your traffic can support.

flowchart TD
  A[Person answers with long greeting] --> B[AMD hears speech then silence]
  B --> C{Greeting too long}
  C -->|Yes| D[Flagged as MACHINE]
  C -->|No| E[Flagged as HUMAN]
  D --> F[Hang up or play message]
  E --> G[Send to agent]
  F --> H[Lost live customer]

Confirm the problem before you change anything

Pull the AMD Log Report and look at how many calls got a machine result. Then listen to a sample of recordings against those results. If calls logged as machines turn out to be people, you have a false-positive problem, and the disposition (Disposition) data on those leads is wrong too. Read the AMD Log Report guide for how to read it.

Loosen the detection to protect humans

The fix is to make AMD more forgiving so borderline calls fall on the human side. In the AMD Asterisk parameters, that generally means allowing a longer greeting before a call is judged a machine and giving the analysis more time and tolerance for silence between words. The trade-off is real: loosening detection lets more actual machines slip through to agents. You are choosing to protect live answers at the cost of a few wasted agent (Agent) seconds.

There is also a routing safeguard. AMD can return a NOTSURE result for calls it cannot confidently classify, and by default those go to agents along with humans. If you are losing real people, make sure NOTSURE calls are routed to an agent rather than hung up or sent to a message. A human ear sorts out the ambiguous cases far better than the algorithm.

Or sidestep detection entirely

If false positives keep costing you good answers, weigh whether you need AMD at all. Setting a Dial Timeout around 22 to 26 seconds dodges most voicemail pickups without ever risking a live person, since it simply drops the call before machines tend to answer. Compare the approaches in the AMD versus Dial Timeout comparison, and get the full background in the AMD and CPD complete guide.

Cutting false positives is a tuning loop, and it is easier on a dialer that is set up right. VICIfast manages the AMD application and your campaign (Campaign) settings so you can adjust thresholds and measure results without touching the server. See VICIfast pricing to stop losing live customers to bad detection.