Why your AI receptionist should never say "I didn't catch that"
Most AI phone systems fail in the same place: the moment a caller says something slightly unexpected. Here is what we changed.
Most AI phone systems fail in the same place: the moment a caller says something slightly unexpected. Here is what we changed.
The failure mode nobody demos
Vendor demos use clean audio and cooperative scripts. Real calls have a truck idling, a toddler, and someone saying "yeah hi is this the place that does the AC thing".
When the model is unsure, the lazy behaviour is to ask the caller to repeat themselves. Do that twice and the caller hangs up.
What we do instead
- Commit to the most likely intent and confirm it in passing
- Keep the caller moving toward a booking
- Escalate to a human the moment confidence stays low
A caller should never be asked to repeat themselves twice.
Measuring it honestly
We track containment alongside booking_rate — a bot that contains every call but books nothing is worse than useless.
intent_confidence: 0.71 -> confirm inline, keep going
intent_confidence: 0.38 -> escalate, do not re-promptSee how it works for the full pipeline.
The result
The calls that do escalate reach a human already knowing what the caller wanted, which is the part that changes how the conversation goes.