What an AMD false positive is
Answering machine detection makes two kinds of mistakes:
- A live person judged to be a machine. The person says hello and hears nothing, or hears a recorded voicemail message meant for an answering machine. In this post, that is the false positive.
- A machine judged to be a person. An agent gets a voicemail greeting and has to drop it. That wastes agent time, but the agent can recover in a few seconds.
Vendors don't all use the same names for these errors, so when someone quotes an accuracy figure, ask which error it describes. A single "accuracy" number can hide a poor result on the error that matters most to you.
The two errors are not equal. A machine sent to an agent costs a few seconds of agent time. A live person sent away costs the whole reason you dialed. That is why false positives deserve their own line in your budget.
What a false positive costs you
A lost conversation
Every answered call is the end of a paid process: list cost, carrier minutes, dialer capacity and the attempts it took to get an answer. When AMD hangs up on the person who finally picks up, all of that is spent and nothing comes back. On consumer lists, live people are often a minority of answered calls, so each one you lose matters more.
A burned lead
A person who answers and hears silence, then a click, has a reason not to answer that number again. Some will block it or report it. The lead is still in your list, but it is now harder to reach than it was before you called.
Dead air and caller ID reputation
Detection takes time. While AMD listens, the person on the line hears nothing. Long silences and hang-ups are what people complain about, and they can hurt how your numbers are treated. Our dead air guide covers this in detail.
Compliance exposure
Under the US Telemarketing Sales Rule and FCC rules, abandoned calls are limited to 3% of answered calls per campaign per 30 days. A call counts as abandoned if the person isn't connected to a live agent within 2 seconds of their completed greeting. A live person that AMD drops, or holds in silence too long, can count against that limit. The UK's Ofcom has similar rules on silent and abandoned calls. This is not legal advice; check with counsel for your campaigns.
Why false positives happen
Most VICIdial and Asterisk centers use Asterisk's built-in AMD(), which VICIdial runs through extension 8369. AMD() doesn't understand speech. It measures timing: how long the first silence lasts, how long the greeting is, how many words it hears and how long each word runs. Shipped defaults include initial_silence 2500 ms, greeting 1500 ms and maximum_number_of_words of 2 or 3, depending on version.
Real people break those rules all the time:
| What the person does | What AMD() may decide | AMDCAUSE |
|---|---|---|
| Answers with a full sentence, such as a business greeting | Greeting too long, so a machine | LONGGREETING |
| Says several words before pausing | Too many words, so a machine | MAXWORDS |
| Picks up and waits for the caller to speak | Too much initial silence, so a machine | INITIALSILENCE |
| Talks for a long stretch without a pause | Word too long, so a machine | MAXWORDLENGTH |
Noisy lines, speakerphones and background conversations make it worse. You can loosen the rules, but then more voicemail greetings slip through to agents. Timing rules force you to pick which error you would rather have. Our page on AI answering machine detection explains why tightening one rule tends to create another mistake, and the AMD FAQ answers common questions.
A worked example (hypothetical)
The numbers below are made up to show the method. Replace every one of them with your own.
Imagine a center that answers 40,000 calls a day and dials 22 days a month:
- Answered calls a month: 40,000 × 22 = 880,000.
- Live people: assume 30% of answered calls are live people, so 264,000.
- Live people reaching agents today: assume the current AMD sends 85% of them to agents, so 224,400 reach an agent and 39,600 are sent away every month.
- With a 94.9% rate (nomoVM's measured share of live people sent to agents): 250,536 reach agents and 13,464 are sent away.
- Difference: 26,136 more live conversations a month.
- Value: if 3% of conversations convert and each conversion is worth $40, that is about 784 extra conversions, or about $31,360 a month.
For comparison, 880,000 answered calls fit in nomoVM's Starter plan at $99 a month. Again, these inputs are hypothetical. The share of live people on your lists, what your AMD keeps today and your conversion value are what decide the real answer.
Run it with your own numbers
The AMD ROI calculator does this math in your browser. It asks for answered calls a day, dialing days, the share of live people among answered calls, the share your AMD sends to agents today, your conversion rate and the value of one conversion. It then shows how many more live conversations reach agents each month, what they are worth, the nomoVM plan cost and the net value.
How to measure your false positives
You can't fix a number you haven't measured. This takes an afternoon:
Sample machine verdicts
Pull a random sample of a few hundred calls your AMD marked as machines over a normal week. Random matters: don't pick only the short calls or only one list.
Listen to the first seconds
Have a person listen to the start of each call and label it: live person, voicemail greeting, screening assistant, silence, fax or tone. You need audio from answer onward; if your setup doesn't keep it for calls AMD hangs up on, see the next section.
Sample human verdicts too
Do the same for calls sent to agents. This tells you how many machines reached agents, the other error.
Scale to your volume
Multiply each sample's share by the total calls in that group. Live people sent away divided by all live people is your false positive share. Live people at agents divided by all live people is the share you keep.
Repeat after any change
Every settings change or new list shifts the numbers. Re-measure before you trust a new setup.
Report both numbers together. nomoVM's accuracy page does the same: 94.9% of live people sent to agents, and 88% of calls sent to agents were live people, measured on 400 real outbound calls checked by a person, none used to build nomoVM.
How to reduce false positives
- Tune AMD() with data, not guesses. If most of your lost live people show
LONGGREETING, a longergreetingsetting helps, at the cost of more machines reaching agents. Change one setting at a time and re-measure. - Keep listening time short. Every extra second of analysis is dead air for a live person.
- Treat call screening as its own case. Google Call Screen and iPhone screening are neither a person nor a voicemail greeting. Knowing how many you get helps you plan for them.
- Use detection that understands the audio. nomoVM is hosted AI AMD that decides from the sound of the answer itself, not from timing rules. Most calls are decided about 2 seconds after answer, and unclear calls listen up to 3.2 seconds. It returns
AMDSTATUSasHUMAN,MACHINEorFALLBACK, andAMDCAUSEsays why, includingSCREENING,SILENCE,TONEandBUSY.
It is honest to say where any AMD still slips. nomoVM's remaining errors are mostly very quiet answers, people talking to someone else, noisy lines, short personal greetings and brand-new recordings built to sound human.
There is no additional infrastructure required: no new servers, hardware or carrier change, just one script on the dialer you already run, and about 10 minutes to set up. The dashboard keeps the seconds the service heard, with playback and CSV export, so measuring is built in. If the service can't be used, Asterisk's own AMD() runs for that call, and fallback calls are never billed. See the VICIdial AMD page and pricing.
Next steps
Measure your current false positives with a sample this week, then put the numbers into the ROI calculator. To compare directly, create a free nomoVM account (30,000 calls a month, no card), follow the VICIdial AMD setup guide, and run one campaign side by side for a week. For the fee side of the picture, read our answering machine detection cost comparison.
Frequently asked questions
What is an AMD false positive?
In this post, an AMD false positive is a live person that answering machine detection judges to be a machine. The call is then hung up, sent to a voicemail drop, or otherwise kept from an agent. Some vendors use the term the other way round, so always check which error a number describes.
Why does Asterisk AMD hang up on live people?
Asterisk AMD() decides with timing rules such as greeting length, initial silence and word count. A person who answers with a long greeting, says several words, or waits silently can trip those rules and be marked as a machine, with causes like LONGGREETING, MAXWORDS or INITIALSILENCE.
How do I measure my AMD false positive rate?
Take a random sample of calls your AMD marked as machines and have a person listen to the first seconds of each. Count the live people, scale that to your monthly volume, and compare it with the live people that did reach agents. That gives you the share of live people your AMD sends away.
Can AMD false positives cause compliance problems?
They can. Under the US Telemarketing Sales Rule and FCC rules, a call counts as abandoned if a person isn't connected to a live agent within 2 seconds of their completed greeting, and abandoned calls are limited to 3% of answered calls per campaign per 30 days. This is not legal advice; check with counsel.
How accurate is nomoVM at keeping live people?
In a test on 400 real outbound calls checked by a person, none used to build nomoVM, 94.9% of live people were sent to agents and 88% of calls sent to agents were live people. Results on your own lists will vary, which is why we suggest testing on one campaign.
How do I estimate what false positives cost my center?
Use the nomoVM ROI calculator. Enter answered calls a day, dialing days, the share of live people on your lists, the share your AMD sends to agents today, your conversion rate and the value of a conversion. It runs in your browser and shows the monthly difference.