Every outbound operation running a predictive dialler makes the same judgement on every connected call: is the voice that just answered a person or a machine? Get it right and the call routes to an agent or drops a voicemail. Get it wrong and you either burn agent time on a voicemail greeting or, far worse, hang up on a real customer.
Most teams tune this once at deployment and never look at it again. That is a mistake, because the two failure modes have very different costs and they move in opposite directions.
Two failures, two very different costs
A false negative is a machine classified as a person. The call routes to an agent, who hears a voicemail greeting, sighs, and dispositions it. The cost is a few seconds of agent time, multiplied across the campaign. It is annoying and it is measurable in handle time, but it is bounded.
A false positive is a live person classified as a machine. This is the one that hurts. Depending on your configuration, the customer either hears silence and a disconnect, or gets a pre-recorded message aimed at an answering machine while they are standing there saying "hello?". You have lost a contact you paid to reach, and you have annoyed the person you were trying to sell to or serve.
Why tuning is a trade, not an optimisation
Detection works by analysing the first moments of audio after answer: the length of the initial utterance, the cadence, the pause that follows, and increasingly the acoustic signature of the greeting itself. Every system exposes some form of sensitivity threshold.
Push the threshold towards "assume machine" and you catch more voicemails, keep agents off dead calls, and drive up false positives. Push it towards "assume person" and you protect live contacts at the cost of agents hearing more greetings. There is no setting that reduces both. There is only the setting that is right for your call profile and your risk appetite.
That profile varies more than people expect. Mobile-heavy consumer lists behave differently from business landlines. Regional greeting conventions differ. Carrier-injected network announcements confuse detectors that were tuned on clean audio.
What to measure, and how
The awkward truth is that your dialler's reporting cannot tell you your accuracy. It reports its own classifications, which is exactly the thing in question. Measuring accuracy means comparing decisions against ground truth, and ground truth means someone listening.
The practical method is unglamorous and works:
- Pull a random sample of a few hundred calls from a single campaign and a single day.
- Listen to the first ten seconds of each and score it: person or machine.
- Compare your scores against what the system decided.
- Express the result as two rates: live calls wrongly dropped, and machines wrongly passed to agents.
Do it once and you have a baseline. Do it again after each threshold change and you have a feedback loop, which is more than most operations ever build.
| Approach | How it is tuned | Latency added | Audio leaves your estate | Best suited to |
|---|---|---|---|---|
| Dialler built-in detection | Usually a single sensitivity setting | Minimal | No | Low-volume campaigns where the default is good enough |
| Carrier or platform add-on | Tuned by the vendor across all its customers rather than to your list | Low | Usually yes | Teams without infrastructure to run their own |
| Self-hosted detection service | Can be tuned against your own scored calls | Low, if co-located with the dialler | No | Regulated sectors and high-volume outbound |
The data protection angle
Detection has to inspect call audio, which means the question of where that audio goes is a live one. If your outbound calls touch financial, health or other special category data, then any third party analysing the audio stream sits inside your processing chain, with the contractual, retention and audit obligations that implies.
Where to start
If you have never measured this, do not begin by changing settings. Begin by finding out what your current accuracy actually is. Score a sample by hand, get the two numbers, and decide which one you want to move. Most teams discover their detection is either far better or far worse than assumed, and either answer is more useful than the guess they were operating on.
Frequently asked questions
What is a reasonable false positive rate for answering machine detection?
There is no single figure to aim for: it depends on your call profile, your lists and your audio quality. What matters is knowing what yours actually is. Most teams have never measured it, because the dialler only reports what it decided, not whether it was right.
Does answering machine detection count towards the Ofcom abandoned call limit?
Yes. Under Ofcom's persistent misuse policy, a call where a live person answers and is then cut off is an abandoned call, and the abandoned call rate (no more than 3% of live calls on each campaign over 24 hours) must include a reasoned estimate of AMD false positives.
Can answering machine detection run without sending audio to a third party?
Yes. Detection can run on your own infrastructure alongside the dialler, which keeps call audio inside your estate. That matters where recordings carry personal or financial data and your retention policy has to cover every system that touches them.