AI Voice Receptionist

The build

A dental practice loses the call that lands while both chairs are busy, and the one at 9pm. Those callers do not leave voicemail — they ring the next practice on the list, so a callback queue solves the record-keeping and not the booking. Vapi runs the voice agent; this is the back end it calls out to mid-conversation, one webhook routed by tool name: read real open slots from Google Calendar, create the appointment, answer a practice question from an Airtable knowledge base the practice edits itself, or score urgency and page the on-call dentist over Telegram. The constraint that shaped it is that a voice agent cannot wait and cannot see an error, so every branch — failure paths included — returns a Vapi-shaped response rather than letting a node fail, and the agent never dead-ends mid-sentence. The end-of-call report is turned into a structured record, logged, and fanned out to SMS, email and Telegram; a 06:00 branch sends next-day reminders, an owner digest and a 48-hour nudge to callers who never booked.