For dialler managers and ops leads who want detection to earn its keep: what AMD can genuinely do, what each mistake costs in the UK and US, and how we tune it on the floors we run.
How do you tune Five9 answering machine detection?
Start by choosing which error you can afford. Five9 classifies the first seconds of each answered call, so every profile trades hung-up humans against voicemails reaching agents. Tune per list in Call Analysis rather than globally, bias towards protecting live answers on regulated consumer lists, verify the error rates from your own dispositions and sampled recordings, and switch AMD off for small lists and expensive leads.
Two seconds of audio, a handful of signals, and a guess with money on it.
When an outbound call connects, Five9 has roughly two seconds of audio to decide whether the voice on the line is a person saying hello or a recording asking for a message. It cannot wait much longer, because a real human who says hello into silence twice will hang up, and in the US the clock on the two-second connection rule is already running. Everything about AMD follows from how little evidence fits into that window.
The detector reads a few signals from those first moments, none conclusive on its own.
Five9 describes the feature plainly enough: technology that works out whether a call was answered by a person or a voicemail system, configured through Call Analysis on outbound and autodial campaigns. What no vendor publishes is an audited accuracy figure for a real mixed list, and the honest answer is that a single figure would be meaningless anyway. A household answerphone with a long branded greeting is an easy classification. A commuter answering a mobile with one flat syllable is not. The same profile that tests beautifully on one list will embarrass you on another, which is why everything below keeps returning to per-list tuning.
One error interests a regulator. The other empties your occupancy figures.
Every AMD error lands in one of two buckets, and they are not equally priced.
The false machine first. A live person answers, the detector reads them as a recording, and the platform drops the line. From their side the phone rang, they said hello twice, and the call died. In the UK that is a silent call. Ofcom's persistent misuse statement names the technology directly, warning that AMD may generate silent calls by mistaking a call recipient for an answer machine and disconnecting, and silent calls are the category it treats most seriously, with penalties of up to £2 million. Since the current statement took effect in March 2017 there has been no safe percentage either: the earlier three per cent abandoned call threshold no longer operates as a safe harbour, the abandoned call rate is one factor Ofcom weighs when prioritising cases, and it reserves the right to act on any campaign. Our Ofcom dialler rules guide covers the full framework.
In the US the same event is priced in abandonment. Under the Telemarketing Sales Rule, a call answered by a person must reach a representative within two seconds of that person's completed greeting or hear a compliant recorded message; otherwise it counts as abandoned. The safe harbour in 16 CFR 310.4 caps abandonment at three per cent of calls answered by a person, measured per campaign over each 30-day period, and the FCC's rules carry the same limit. An AMD false positive is exactly a call answered by a person that received neither an agent nor the message, so every one of them spends that budget. We walk through the detail in our TCPA abandonment rules article.
While we are here: abandon rate is abandoned calls divided by calls answered by a live person. Not divided by dials. Both regimes measure against live answers, dividing by dials makes any floor look angelic, and we still find the wrong denominator in about half the compliance reports we audit. That last figure is our experience rather than a survey, but it has been depressingly consistent.
The false human is the other bucket: a voicemail greeting read as a live answer and delivered into an agent's headset. No regulator minds. Your operations lead should. The agent listens to the greeting, waits out the beep, wraps a disposition, and the best part of half a minute is gone. String enough of those together and occupancy sags while the pacing engine, convinced it has been finding connects, keeps making the same mistake. A dialler in this state looks busy and produces nothing, which is the most expensive kind of busy there is.
Our free Five9 Health Check is 17 questions and a few minutes of your time. It flags whether detection, pacing or dispositions are the settings dragging your numbers, before a regulator or a quarter-end does.
You cannot minimise both. Pick the mistake you can afford.
Tighten detection to catch more machines and it will catch more humans. Loosen it to protect humans and more voicemails reach agents. No platform escapes this, whatever the sales deck implies, so the only real tuning question is which error is cheaper on this campaign, on this list, in this jurisdiction.
For a UK consumer campaign the arithmetic is lopsided. The false machine risks attention from a regulator whose maximum penalty is £2 million and who no longer recognises a safe rate. The false human wastes seconds you can count and manage. That is why the UK setups we audit get biased conservative, accepting more voicemail into headsets in exchange for fewer dropped humans, and why we treat an aggressively tuned detector on a UK consumer list as a finding rather than a preference.
US campaigns have slightly more room, because the safe harbour tolerates abandonment of up to three per cent of live answers. But that allowance is shared. Predictive overdials and AMD false positives draw from the same account, and the pacing side is usually where the revenue lives. Spend the whole budget on a badly tuned detector and you end up running the dialler timidly just to stay inside the line, which costs more than the voicemails ever did.
Where you sit on the dial is therefore a commercial decision, and it belongs to whoever can see the compliance report and the occupancy report at the same time. On the floors we run, that is a weekly conversation rather than a go-live checkbox.
The controls are simple. Knowing where to set them is the job.
The switches live in the campaign profile. Call Analysis is where you enable detection for outbound and autodial campaigns, choose what happens when a machine is identified (drop the call, play a recorded prompt after the beep, or pass it through to an agent) and adjust how long the platform listens before committing to an answer. Five9's campaign documentation lists the options. It cannot tell you where to put them, because that depends entirely on who you are dialling.
The most useful mental model is that the listening window buys accuracy with dead air. Let the detector listen longer and it classifies better, but every live answer spends that time saying hello into silence, which is exactly the experience the silent call rules exist to prevent. Shorten the window and connects feel instant, at the price of more machines slipping through to agents. You are positioning the campaign on the same trade-off as the previous section, just expressed in seconds.
Lists do not behave alike, and mobile-heavy lists are the awkward ones.
| List profile | What tends to answer | How detection copes | Where we bias it |
|---|---|---|---|
| Mobile-heavy consumer | One-syllable answers, network-default voicemail greetings, background noise | Worst case. Short live answers imitate recordings, and default greetings vary by network | Conservative settings, shorter listen window, accept more voicemail reaching agents |
| Residential landline | Fuller greetings, household answerphones with longer messages | The classic signals mostly hold | Middle settings, reviewed weekly |
| B2B and switchboards | Receptionists, IVR menus, hold marketing | IVRs read as machines while a human sits one menu behind them | Often better in power or preview with agents deciding |
| Aged and re-dial data | A high share of voicemail, warier answerers | Sheer machine volume makes each percentage point matter | Detection on, message strategy decided deliberately rather than by default |
Run each list type as its own campaign with its own profile, change one setting at a time, and give any change a meaningful sample of connects before judging it. Detection also interacts with pacing, so tune it alongside your dialling mode rather than in isolation; our predictive dialler settings guide covers that half of the equation. One US-specific detail worth knowing: the Telemarketing Sales Rule also expects an unanswered call to ring for at least fifteen seconds or four rings before disconnection, so ring time is not a free variable either.
AMD is a scale tool, and not every campaign has scale.
Some campaigns should not run detection at all, which is a sentence you will rarely find in vendor material.
Small lists are the first case. Detection exists to save agent seconds across tens of thousands of dials; on a few thousand records the seconds saved never repay a single mishandled live answer, and the campaign finishes within days anyway. Let agents hear every pickup.
Expensive leads are the second. When a record has real acquisition cost behind it, a false machine is your media budget hanging up on its own prospect, and no occupancy gain covers that. On the mortgage and home improvement floors we run, fresh high-intent leads dial in power or preview with AMD off, and only the aged tail sees detection. If you are weighing those modes, our dialling modes comparison sets out the trade properly.
The third case is any compliance-sensitive segment: collections books, customers flagged as vulnerable, any base where a silent call turns into a logged complaint with your name on it. The occupancy argument does not survive contact with one upheld complaint in the wrong file. Scheduled callbacks belong here too, for a different reason: the person asked for the call, so let a human deliver it from the first syllable.
The feature is easy to switch on. The consent question is the hard part.
Leaving a message on every detected machine sounds free. The dialler already knows nobody live is there, so the temptation is to let the platform play thirty seconds of brand while agents take other calls. Two things argue against it: line time and law.
Line time first. Waiting through a greeting and a beep occupies a channel that pacing could have spent dialling, so message-heavy campaigns run measurably slower. Whether that trade pays depends on whether the messages actually generate callbacks, which is testable from your own inbound numbers. On the campaigns where we have measured it honestly, callback yield has been modest, and we say that as a firm that gets paid to configure it either way.
Now the law, and the US first. A voicemail drop is a prerecorded message. FCC rules under 47 CFR 64.1200 require prior express consent for artificial or prerecorded voice calls to mobiles, and prior express written consent where the content is marketing. Ringless voicemail was sold for years as the loophole around all of this; the FCC closed it in November 2022, ruling in FCC 22-85 that ringless voicemail to a wireless phone is a call made with an artificial or prerecorded voice and requires consumer consent. Courts have kept the pressure on since: in 2025 a federal court in Illinois held that near-identical voicemails were enough, at the pleading stage, to allege prerecorded calls. Rules and enforcement in this area keep moving, so verify the position against the current texts and take advice from qualified counsel before building a message strategy with real exposure.
The UK is stricter than most US operators assume. A recorded direct marketing message delivered by an automated system falls under regulation 19 of PECR, which requires the recipient's prior consent, and most outbound lists simply do not hold it. Separately, Ofcom's abandoned call rules describe what an information message should contain when no agent is available: identify who called, explain that contact was attempted, offer a basic rate number for opting out, and carry no marketing content. A voicemail drop that pitches product at a UK consumer list fails both tests at once, which is why operators here typically reserve recorded messages for the abandoned call scenario and keep the selling out of them.
The numbers worth trusting come out of your campaigns, not the defaults.
Platforms ship sensible defaults, and defaults have never met your lists. Both error rates are measurable with tools you already have, and neither takes more than an hour a week.
The false human rate is the easy one. Give agents a one-click voicemail disposition, then report its share of connects per list per week. When that share climbs, either the detector has drifted, the list is decaying towards voicemail, or your calling windows are wrong. Any of the three is worth knowing on Tuesday rather than at month end.
The false machine rate takes slightly more effort because those calls never reach a headset. Pull a weekly sample of calls the detector classified as machines and dropped, listen to the recordings, and count the humans saying hello. Do it per list, since an acceptable rate on aged landline data can coexist with an ugly one on a fresh mobile file. Twenty minutes of listening tells you more about your real silent call exposure than any dashboard, and in the UK this is the number you would be asked to justify if Ofcom ever asked questions.
Agent voicemail dispositions as a share of connects, per list, per week. Rising means detector drift, list decay or wrong calling windows.
Humans found in a weekly sample of AMD-dropped call recordings. This is silent call exposure in the UK and abandonment spend in the US.
Abandoned calls divided by calls answered by a live person. Never divided by dials, whatever the report template suggests.
If you leave voicemails, the inbound callbacks attributable to them. This decides whether drops earn the line time they consume.
We run the same 30, 60 and 90 day review rhythm on detection changes that we apply to any implementation work, because real call volume always drifts away from assumptions. A profile tuned in January meets different voicemail greetings, different carrier behaviour and a different list mix by June. If your AMD was configured once at go-live and has not been looked at since, it is not tuned. It is merely still switched on.
AMD listens to the first seconds of an answered outbound call and classifies it as a live person or a recording, using signals such as greeting length, cadence and audio energy. Five9 exposes it through Call Analysis on outbound and autodial campaigns, where you choose whether detected machines are dropped, played a message after the beep, or passed to an agent. Its purpose is to keep agents talking to humans instead of listening to greetings.
No vendor publishes audited accuracy figures for real mixed lists, and honest operators will not give you a single number either. Accuracy depends on the list: landline answerphones with long greetings classify well, while mobile-heavy consumer lists with one-syllable answers and carrier-default voicemails are genuinely hard. The useful question is not a headline percentage but which error dominates on your list, measured from your own dispositions and sampled recordings.
Yes, when AMD drops a call that a live person answered. Under the Telemarketing Sales Rule, a call answered by a person must reach a representative within two seconds of the completed greeting or hear a compliant recorded message; otherwise it is abandoned. The safe harbour caps abandonment at three per cent of calls answered by a person, per campaign over 30 days, and AMD false positives draw from the same budget as pacing abandons.
Ofcom's persistent misuse statement identifies AMD as a cause of silent calls when it mistakes a live person for a machine and disconnects. Silent calls are a stated enforcement priority, with penalties of up to £2 million, and since March 2017 there has been no percentage threshold below which a campaign is automatically safe. UK operators typically bias detection conservatively and monitor false positives by sampling recordings of dropped calls.
The FCC ruled in November 2022, in FCC 22-85, that ringless voicemail to a mobile is a call made with an artificial or prerecorded voice, so it requires consumer consent under the TCPA, and courts have continued applying that position in litigation since. In the UK, recorded marketing messages fall under PECR regulation 19 and need prior consent. Treatment keeps evolving, so verify the current position and take qualified advice before relying on it.
Small lists, expensive leads and compliance-sensitive segments. On a few thousand records, the agent seconds saved never repay a mishandled live answer. On fresh high-intent leads, a false machine is your marketing budget hanging up on its own prospect. On regulated or vulnerable-customer books, one silent call can become a logged complaint. In those cases run power or preview dialling and let agents hear every answer from the first syllable.
Decide it with numbers rather than instinct. Messages occupy lines while the platform waits for the beep, so campaigns dial slower, and callback yield on the campaigns we have measured has been modest. Legally, prerecorded marketing messages need prior express written consent in the US, and recorded marketing calls need prior consent under PECR regulation 19 in the UK. If you drop messages at all, do it on segments where consent is documented and measure the callbacks honestly.
Track two numbers per list per week. Give agents a one-click voicemail disposition and report its share of connects; that is your false human rate. Then sample recordings of calls AMD classified as machines and dropped, and count humans saying hello; that estimates your false machine rate, which is silent call exposure in the UK and abandonment spend in the US. Review both on a 30, 60 and 90 day rhythm after any change.
We operate outbound floors on Five9 in both the UK and US regimes, so detection tuning is weekly maintenance for us, not theory. If you want your AMD, pacing and dispositions reviewed properly by a dedicated pod, talk to us.
Talk to us