Outbound
Answering Machine Detection: When It Helps and When It Hurts
How answering machine detection works in dialers, why it misclassifies calls, the trade-offs for predictive campaigns, and how to test AMD before using it.
- Author
- JH Dialer Engineering Team
- Published
- Reading time
- 1 min read
On this page
Answering machine detection (AMD) promises a simple win: agents only hear live people. In practice it is a trade-off worth testing rather than assuming.
How does AMD decide?
AMD listens to the first moments after the call is answered. A short “Hello?” followed by silence suggests a person; a long continuous greeting suggests a machine. The algorithm uses thresholds for greeting length, silence and word count.
Why does it get it wrong?
- People who answer with a long greeting (“Good afternoon, Sharma residence, how can I help?”) can be classified as machines.
- Short voicemail greetings can be classified as people.
- Background noise and line quality affect detection.
What does it cost you?
- Delay — the analysis time is silence for the person who answered. Some hang up.
- Lost conversations — misclassified people are dropped or sent to a message.
- Regulatory exposure — in some markets, how you treat a call after answer affects abandoned-call calculations.
When is AMD worth it?
- Lists where a large share of answers are voicemail.
- Campaigns where leaving a message has real value.
- Teams that have measured contact rates with and without it.
How should you test it?
Split a list. Run half with AMD and half without for long enough to compare contacts per agent hour, conversions and drop rate. Keep the version that wins on the metric you care about — not the one that merely sounds more efficient.