A voice agent that books 1,200 patient appointments per week across a 14-clinic network

Fourteen clinics, one shared call center, and a phone tree from 2009. The client was losing appointments to voicemail every evening and weekend. Their answering service cost $14k a month and only forwarded messages — patients still had to wait until the next morning to actually book. The dropout rate from "left a message" to "appointment kept" was 40%.

The internal scheduling team handled 4,000 inbound calls a week. Two thirds were the same five intents: book new appointment, reschedule, confirm, ask for directions, ask if a provider takes a specific insurance.

Sprint one (2 weeks) — single-clinic pilot. Built a voice agent on Twilio + Deepgram + GPT-4.1 + ElevenLabs. Custom EHR adapter to read provider availability, insurance acceptance, and existing patient records. Strict guardrails: the agent will book, reschedule, cancel, and confirm — anything else (clinical questions, billing disputes, provider-specific requests) routes to a human within 15 seconds.

Sprint two (2 weeks) — fleet-wide rollout. Multi-location routing. Each clinic gets its own number and provider schedule but shares the agent layer. Added language detection and a Spanish lane (the network serves a heavily bilingual market). Built an admin dashboard so each clinic manager can see real-time call volume, conversion, and any escalations needing a callback.

We designed the agent voice with the client's marketing team — warm but not sticky-sweet, clearly an AI but never robotic. Patients can interrupt at any time. The agent confirms back every booking detail before committing to the EHR.

1,200+ patient calls per week handled end-to-end with no human in the loop. Booking conversion of 78% (industry average for human call centers is around 65%). The answering service was retired, saving $168k/year against an annual operating cost for the agent of about $14k including telephony and model usage.

The unexpected win was after-hours: 31% of bookings now happen between 6pm and 8am — appointments that previously would have been voicemails dropped overnight. Patient satisfaction scores on post-call surveys came in higher than the human call center, mostly because there was no hold time.