Most VICIdial rooms have an AMD config that is either too aggressive (drops real humans) or too soft (inflates dial ratio on machines that never transfer). Almost nobody is in the narrow middle where AMD actually does what it is supposed to do.
The two failure modes
Too aggressive (amd_send_message=HANGUP, drop_call_seconds=4-5):
- AM accuracy in the 88-94% range
- Real-human false-positives drop 2-4% of valid connects
- FCC abandon rate looks great (because AMD drops are not counted) while real-human abandon might be 5-6%
- Classic signature: great-looking dashboards, slowly dying conversion
Too soft (amd_send_message=CONTINUE or drop_call_seconds=10+):
- AM accuracy collapses to 70-78%
- 12-18% of calls burn 6-10 seconds on machines before the system gives up
- Dial ratio inflates because the predictor "sees" more available lines than the agent pool can actually handle
- Classic signature: agents wait 15+ seconds between real connects while dial stats show a healthy 3x ratio
The middle
The right AMD config depends on the carrier, the lead vintage, and the time of day:
- Teleforwarding + fresh leads at 10am local: can push
drop_call_seconds=6safely - Aged leads (>60 days) at 7pm local: needs
drop_call_seconds=8-9or you eat real humans
We document the full AMD tuning matrix, including sampling methodology for measuring real false-positive rate against recordings, at predictive-dialer-roi.com/answering-machine-detection-tradeoffs.
The sampling method nobody does
The cheap way to measure AMD accuracy: pull 50 random calls dispositioned as AMD, listen to 8-10 seconds of each recording, count humans. If >3 of 50 are real humans, your AMD is eating valid connects. Typical operators never do this; they trust the dialer's reported AMD rate.
Companion reading:
If you run VICIdial and your conversion is slowly bleeding without obvious cause, AMD is usually where to look first.
Jason Shouldice
Founder and CEO, ViciStack
Houston
