How NuFace Dental Hospital automated inbound reception with Anaya — a Gujarati-default trilingual voice AI integrated with their on-premise HMS
NuFace Maxillofacial & Dental Hospital in Surat covers Maxillofacial Surgery, General Dentistry, Orthodontics, Implants, Endodontics, Periodontics, Pediatric Dentistry, Cosmetic Dentistry and Prosthodontics — nine specialty departments served through a single inbound patient-call channel. Patients in this catchment call primarily in Gujarati, with Ahmedabad/Surat-style code-mixed Gujarati-Hindi-English the natural way they actually speak. Caller Digital deployed Anaya — a 15-stage finite-state-machine voice AI receptionist — integrated with the hospital's on-premise HMS/CRM for real-time appointment booking, rescheduling, cancellation and lookup.
High-volume inbound calls across 9 specialty departments, with Gujarati-Hindi-English code-mixing the norm, peak-hour drop-off, and rescheduling that forced reception to flip between the HMS and the live caller
NuFace's reception team needed to handle a high-volume inbound call channel for appointment scheduling, rescheduling and cancellation across nine specialty departments, while answering patient FAQs (hours, location, doctors, indicative cost, departments).
Patients call primarily in Gujarati with Surat-style Ahmedabad code-mixing across Gujarati, Hindi and English. Manual reception capacity at peak hours dropped calls; rescheduling required hopping between the HMS/CRM doctor calendar and the patient on the line; after-hours bookings could not be served at all.
Peak-hour calls dropped
Manual reception capacity could not absorb peak inbound volume. Dropped calls translated directly into lost appointments and frustrated patients.
Rescheduling required HMS-flip mid-call
Reception had to switch between the on-premise HMS doctor calendar and the patient on the line — slowing call wrap, frustrating patients, and producing data-entry errors.
After-hours and weekend bookings unserved
Patients calling outside reception hours had no path to book or check appointments. They went to competitors or skipped the appointment entirely.
Gujarati-Hindi-English code-mixing handled inconsistently
Different reception staff applied different code-mix defaults. Patients who naturally code-switch sometimes felt the call register did not match how they actually speak.
Anaya — a 15-stage finite state machine on Gemini Live with a BASE/STAGE two-layer prompt architecture that eliminates the most common healthcare voice-AI failure modes
Caller Digital deployed Anaya as the hospital's inbound AI voice receptionist. Anaya runs a 15-stage finite state machine on top of a real-time Gemini Live model: Greeter, BookingReason, BookingPhone, BookingSlotPick, BookingPatientInfo (NEW patients), BookingConfirm, ReschedulePhone, RescheduleNewSlot, CancelStart, PickAppointment, CancelConfirm, LookupPhone, InfoProvider, EmergencyHandler, EndCall. Each stage owns one focused sub-task and hands off cleanly to the next.
Anaya's instructions are deliberately split into two layers. The BASE layer is sent once at session start and contains the persona, voice and accent rules, conversational guardrails, objection handling, and full hospital reference information. The STAGE layer is pushed at every state transition and contains the current stage name, a snapshot of everything captured so far, and a single-sentence task plus the tool to call next. This split is the engineering that stops the bot from re-asking captured information, narrating its own guardrails aloud, or looping on stale instructions — the three failure modes that historically made automated reception unusable in healthcare.
A shared session-state object (NufaceData) is the single source of truth across all stages: language, phone, patient_type (NEW or OLD via /crm/patient), patient_name and patient_age (auto-filled for OLD patients from the HMS response, asked for NEW patients), patient_city, appointment-specific fields. Stages never re-ask for a field that is already filled. A dedicated update_known_data tool lets the caller fix any field mid-call without restarting the flow.
15-stage finite state machine on Gemini Live
Greeter, BookingReason, BookingPhone, BookingSlotPick, BookingPatientInfo, BookingConfirm, ReschedulePhone, RescheduleNewSlot, CancelStart, PickAppointment, CancelConfirm, LookupPhone, InfoProvider, EmergencyHandler, EndCall — each owns one focused sub-task.
BASE / STAGE two-layer prompt architecture
BASE layer at session start (persona, voice, guardrails, hospital reference). STAGE layer at every transition (current state, snapshot, single-sentence task, tool to call). Eliminates re-asking, guardrail leakage, looping.
Live HMS/CRM integration
Verified against live on-premise HMS API on 14 May 2026. Real-time doctor calendar fetch, patient classification via /crm/patient, appointment create/update/lookup/cancel as real persisted HMS operations.
Gujarati-default trilingual with Surat-style code-mixing
Greets in Gujarati. Switches silently to Hindi or English based on caller. Ahmedabad/Surat-style code-mixing welcomed — it's how patients in the region actually speak.
Privacy-gated lookup, identity-verified cancel/reschedule
LookupPhone returns appointment details only after phone read-back verification. Cancel and reschedule require explicit haan/yes confirmation gates before calling /crm/cancel or /crm/reschedule.
Emergency keyword router
Safety triggers (chest pain, accident, emergency, bleeding) route to the hospital's emergency line without further questioning — safety prioritised over conversation flow.
Inbound calls served continuously by Anaya across the 9 specialty departments, with bookings, reschedules and cancellations writing back to the live HMS as real persisted operations
The Service Agreement is dated 29 April 2026. The integration with NuFace's on-premise HMS/CRM was verified against the live HMS API on 14 May 2026 and covers (a) real-time doctor calendar reads, (b) patient classification via /crm/patient, (c) appointment create/update/lookup/cancel writing back to the HMS in real time, and (d) trigger of appointment confirmation through the hospital's existing SMS/WhatsApp workflow.
| Dimension | Before | After |
|---|---|---|
| Peak-hour inbound calls | Dropped due to reception capacity | Served continuously by Anaya |
| Rescheduling workflow | Reception flipped between HMS and caller | Real-time HMS calendar fetch inside the call |
| After-hours bookings | Not served | 24×7 inbound coverage |
| Gujarati-Hindi-English code-mixing | Varied by reception staff | Gujarati-default with consistent code-mixing |
| Lookup-style calls | Consumed disproportionate reception time | Privacy-gated read-out via phone verification |
| Cancel / reschedule identity verification | Ad-hoc | Explicit haan/yes gate + phone read-back |
Frequently Asked Questions
An inbound voice AI receptionist (Anaya) for appointment booking, rescheduling, cancellation, appointment lookup and patient FAQ — covering all 9 specialty departments at NuFace Maxillofacial & Dental Hospital, integrated with the hospital's on-premise HMS/CRM via secure API access.
Gujarati (default), Hindi and English, with natural Ahmedabad/Surat-style code-mixing welcomed because that is how patients in the region actually speak. The agent greets in Gujarati and from the second turn onward switches silently to whichever language the caller is comfortable in.
Anaya integrates with the NuFace on-premise HMS/CRM via secure API access (with VPN or required network access provided by the hospital) to (a) read doctor calendars and slot availability in real time, (b) classify the caller as NEW or OLD via phone number on /crm/patient, (c) create, update, lookup and cancel patient appointment records, and (d) trigger SMS/WhatsApp confirmations via the hospital's existing integrated workflow.
After phone capture and digit-by-digit read-back, Anaya calls /crm/patient. If the patient is OLD, the response auto-fills name, age and city — Anaya never asks for them again. If the patient is NEW, Anaya captures name and age in the BookingPatientInfo stage. A dedicated update_known_data tool lets the patient correct any field mid-call without restarting the conversation.
Anaya's instructions are split into two layers. The BASE layer is sent once at session start and contains the persona, voice, language and accent rules, conversational guardrails, objection handling, and full hospital reference information. The STAGE layer is pushed at every state transition and contains the current stage name, a snapshot of everything captured so far, and a single-sentence task plus the tool to call next. This split eliminates the most common voice-AI failure modes — re-asking captured fields, guardrail leakage into spoken audio, looping on stale instructions.
The EmergencyHandler stage routes safety-keyword triggers (chest pain, accident, emergency, bleeding) to the hospital's emergency line without further questioning. Cancellation and reschedule flows require identity verification via phone read-back plus explicit haan/yes confirmation gates before any HMS-persisted action.
Run a similar voice AI rollout for your team
We work with Indian enterprises across BFSI, D2C, retail SaaS, and consumer services. Book a 30-minute consult to see what voice AI can do for your call queue.
© 2025 Caller Digital | All Rights Reserved