.png){kind=link}
A 24/7 voice AI receptionist that answers phone calls, checks live calendar availability, and books appointments into Google Calendar and a database — all through natural voice conversation.
Built for healthcare clinics (dental, physio, GP) that miss calls after hours, but the architecture works for any appointment-based business.
A caller phones the clinic. An AI receptionist named Sarah answers, talks like a real person, and handles the whole booking:
- Understands what service the caller wants
- Converts natural speech ("tomorrow", "next Tuesday at 2") into real dates
- Checks the calendar for that exact day before promising anything
- Collects the caller's name and phone number
- Books the appointment into Google Calendar and saves it to the database
- Confirms warmly and ends the call
If a slot is taken, it offers the nearest free time instead. It never claims something is booked unless the booking actually succeeded.
The system splits responsibility between two layers — this is the core design decision that makes it work in real time.
Caller (phone / web)
│ voice
▼
┌──────────────┐ VAPI = the BRAIN
│ VAPI │ • speech-to-text + text-to-speech
│ "Sarah" │ • holds the conversation
│ │ • decides WHEN to call a tool
└──────┬───────┘
│ webhook (tool call: checkAvailability / bookAppointment)
▼
┌──────────────┐ n8n = the HANDS
│ n8n │ • reads / writes the calendar
│ workflow │ • saves to the database
│ │ • returns a single-line answer to VAPI
└──────┬───────┘
│
┌────┴─────┬──────────────┐
▼ ▼ ▼
Google Airtable (response
Calendar database back to VAPI)
Why this split? Voice conversation needs sub-second responses. Putting a reasoning model inside the automation layer is too slow for a live call. So VAPI handles all the real-time thinking and voice, and n8n does the mechanical work — look at the calendar, write the record, answer. Fast where it needs to be fast, reliable where it needs to be reliable.
A single webhook receives every tool call from VAPI. A Switch node routes it by tool name into one of two branches.
Branch 1 — checkAvailability (read):
- Extract the requested date and time from the tool call
- Read that day's events from Google Calendar
- Decide free or busy
- Format the answer in VAPI's required response shape
- Respond to the webhook
Branch 2 — bookAppointment (write):
- Extract name, phone, service, date, time
- Create the event in Google Calendar
- Save the appointment record to the database
- Format the confirmation
- Respond to the webhook
| Layer | Tool |
|---|---|
| Voice + conversation | VAPI |
| Language model | Gemini Flash |
| Voice | Clara (VAPI) |
| Speech-to-text | Deepgram |
| Automation / orchestration | n8n |
| Calendar | Google Calendar API |
| Database | Airtable |
A few real problems this solves under the hood:
-
Date grounding. Voice models don't know today's date and will invent one. The system injects the live current date into the assistant's context so "tomorrow" resolves to the correct calendar date, with a server-side safety net that catches any bad date before it reaches the calendar.
-
Strict response contract. VAPI only accepts a specific JSON shape (
resultsarray, matchingtoolCallId, single-line string, HTTP 200 always). The workflow formats every reply to that contract so responses are never silently dropped. -
Graceful failure. If a tool errors, the assistant tells the caller there's a brief system issue and offers to take details — it never lies and says "booked" when the write failed.
-
Reusable template. The assistant prompt is modular: swapping the clinic info and services block adapts it to any new business without touching the logic.
This is a working demo built end to end and deployed live. The availability check and booking flow both run in production: VAPI holds the conversation, n8n reads and writes the calendar and database, and the caller hears a real answer.
Built by Usman Farooq — AI automation developer specializing in n8n, AI agents, and voice/WhatsApp booking systems.
- LinkedIn: linkedin.com/in/usman-rai
- GitHub: github.com/Usman-rai