The answering machine detection in your dialer costs nothing on the invoice — and it may be the most expensive line item in your operation. Here's the math.
The AMD built into Vicidial is free in the only sense that matters to a spreadsheet: it never appears on one. There is no per-call charge, no monthly fee, no contract. You flip it on in amd.conf and it works.
But "free" AMD is not free. It is a heuristic — Asterisk app_amd timing silences and greetings — that classifies the answer of a call by stopwatch instead of by sound. That trade-off has a price. It just gets paid out of buckets your accounting never connects back to AMD: dropped revenue, regulatory risk, idle agents, wasted carrier minutes, and DIDs you have to keep replacing. Here is each bucket, with the math.
1. Dropped live leads
Stock AMD's headline failure is the false positive: it decides a real person is a machine and hangs up. Heuristic Asterisk AMD runs around 70–85% accuracy and wrongly drops 5–15% of live humans. Every one of those is a prospect who answered the phone and got dial tone back.
Walk through it — the numbers below are illustrative only, plug in your own:
- You place 10,000 dials a day.
- Your list answers at 30% → 3,000 answered calls, of which the humans are the ones you care about.
- Stock AMD wrongly drops 8% of those humans → 240 live prospects vanish, every day.
- At a hypothetical $25 net value per connected lead, that is $6,000/day — roughly $180,000 a month — in opportunity walking out the door.
Change any input and the headline changes, but the shape does not: the dollars lost to dropped humans dwarf any per-detection AMD fee by orders of magnitude. The worst part is you never see it. The dropped call is logged as a machine; the prospect just never picked up again. There is no row in any report that says "human you hung up on."
2. FTC / TCPA exposure
The second cost is the one that does not show up because the first one is hidden. The FCC's safe harbor expects an abandonment rate at or under 3% of answered calls. When your AMD hangs up on a human it mistook for a machine, that drop is recorded as a machine — not as an abandoned call to a person.
So your dashboard shows a comfortable sub-3% abandon rate while your true human-abandon rate is higher, padded with all the live people the heuristic silently discarded. The tighter you tune amd.conf to kill machines aggressively, the more humans you drop — and the cleaner your reported number looks. That is the aggressive-tuning trap: the metric you watch to stay compliant is the same metric the failure mode hides behind.
This cost is regulatory risk, not a line item, which makes it easy to ignore until it is not. AMD supports your compliance program; it is not legal advice and it does not replace your own dialing policy or counsel. The point is narrower: a detector that can't distinguish a dropped human from a machine cannot give you an honest abandon rate.
3. Agent idle time
Tune amd.conf the other way — softly, to avoid dropping humans — and you trade the lead-loss bucket for the labor bucket. A cautious heuristic waits longer before it commits. It listens through 6–10 seconds of a voicemail greeting before deciding it was a machine, and it hands borderline calls to agents to sort out.
That waiting is paid labor. An agent sitting through a recorded greeting, or idling between real connects while the dialer second-guesses itself, is on the clock the whole time. Multiply a few wasted seconds per machine across thousands of machine-answers a day — and remember the network sees roughly 73% of answered calls as machines — and the burned agent-minutes add up to real payroll. By contrast, AMDY decides in under 200 ms, so the agent is only ever connected to a person.
4. Carrier billing for false-answers
Across roughly 2.3 billion answered calls a month, about 14% are carrier false-answers (FAS) — the line reports "answered" and starts billing, but no one ever picked up. Stock AMD has no concept of FAS; app_amd only knows HUMAN / MACHINE / NOTSURE. So those connects look real, your agents get routed to dead air, and your carrier bills you for time on a call that never happened.
At 14% of answered volume, that is one in seven "connects" you are paying for with nothing on the other end — both in carrier minutes and in agent attention. A detector that classifies FAS as its own bucket simply drops those calls before they cost you.
5. Caller-ID reputation
Some of the numbers your dialer hits are honeypots and spam-traps — numbers that exist to catch dialers. Stock AMD can't see them; acoustically a trap answers like anything else. So you keep dialing them, and the carriers and analytics networks that watch those traps flag your caller IDs.
Once a number is tagged "Spam Likely," its answer rate falls — people don't pick up a flagged number. To keep your volume, you buy and rotate fresh DIDs, which get flagged in turn, and the cycle repeats. That is a recurring, compounding cost — DID churn — driven entirely by detection you don't have. AMDY classifies honeypot/spam-trap answers as their own bucket so you can stop dialing them and protect the numbers you already own.
6. Engineering time
Finally, the cost of keeping the "free" thing working. Heuristic AMD is a pile of timing thresholds — initial_silence, greeting, after_greeting_silence, total_analysis_time — that someone has to tune for your lists, your carriers, and your call patterns.
And it drifts. The mix of lists changes, carriers reroute, voicemail greetings shift, and the thresholds that worked last month start dropping humans or letting machines through again. So an engineer re-tunes amd.conf, validates it, and watches it drift right back out. Those hours are salary spent maintaining a detector that is structurally never quite right — because it is guessing from a stopwatch instead of listening.
Totaling up "free"
Add the buckets together and the invoice line that reads $0 is actually one of the most expensive things in your operation. You pay for free AMD in dropped revenue (live humans hung up on), in regulatory risk (an abandon rate that lies to you), in idle agents (labor burned on machines and dead air), in carrier minutes (FAS you can't see), and in DID churn (caller IDs flagged by traps you keep dialing). None of it is on a bill, which is exactly why it persists.
Put a real detector against that. AMDY is an AI/ML model that classifies the acoustic signature of the answer audio — the sound of the pickup, not transcribed words — at 99% accuracy in under 200 ms, sorting human, voicemail, FAS, honeypot/spam-trap, fax, and silence into separate buckets. It installs on Vicidial, Asterisk, FreeSWITCH, or Issabel with one bash command in about five minutes, and it is telco-agnostic, so you keep your carrier.
The pricing sits in context against everything above: $0.00010–$0.00025 per detection, with 50,000 detections a month free on the Sandbox plan. A single recovered live lead pays for hundreds of thousands of detections. Put differently — in the worked example above, one day of dropped leads was hypothetically worth $6,000; 50,000 detections cost between $5 and $12.50.
FAQ
Isn't the AMD in Vicidial free?
The license is free, but the detection is not. Vicidial's AMD is Asterisk app_amd configured in amd.conf — a heuristic that times silences and greetings. It costs nothing to enable, but it misclassifies 5–15% of live humans as machines and can't see carrier false-answers or spam traps. You pay for it in dropped leads, idle agents, wasted carrier minutes, and DID churn — none of which show up on an invoice, which is exactly why it feels free.
How much revenue does bad AMD cost?
It depends on your dial volume, answer rate, and lead value, so the only honest answer is to run your own numbers. As an illustration only: on 10,000 daily dials with a 30% answer rate, that is 3,000 humans reaching the dialer; if stock AMD wrongly drops 8% of them, that is 240 live prospects vanishing every day. At a hypothetical $25 net value per connected lead, that is $6,000/day in opportunity gone — far more than any per-detection AMD fee.
Does free AMD create TCPA risk?
It can mask it. When AMD hangs up on a human it thinks is a machine, that drop is usually logged as a machine, not as an abandoned human call. Your reported abandon rate can sit comfortably under 3% while your true human-abandon rate is higher. The risk is regulatory, not a line item — and it is worse precisely because the metric you watch looks clean. AMD supports your compliance program but is not legal advice; pair it with your own dialing policy and counsel.
How is dialing spam traps expensive?
Stock AMD can't tell a honeypot or spam-trap number from a real person, so your dialer keeps calling them. Carriers and analytics providers use those hits to flag your caller IDs as "Spam Likely", which drops answer rates across all your numbers. Lower answer rates mean you buy and rotate more DIDs to keep volume up — a recurring cost driven entirely by detection you don't have.
How do I calculate my AMD ROI?
Add up the hidden buckets — recovered live leads, reduced FAS minutes, agent time saved, and DIDs you stop burning — then subtract the AMD cost. AMDY runs $0.00010–$0.00025 per detection with 50,000 free every month, so the cost side is tiny next to a single recovered lead.
Top comments (0)