WhatsApp AI Receptionist with Live Booking

A WhatsApp receptionist I built for Mustaraka Properties, one of my own ventures, to book property viewings. The same pattern (qualify, check availability, book, hand off) maps directly to appointment requests at a clinic.

Enquiries arrive from ads and messaging at all hours, and each needs the same questions asked before a person can act.

Staff reply manually, ask qualifying questions one at a time, check a calendar and write notes by hand.

1. Channel

   WhatsApp conversation entering an n8n workflow.

2. AI agent

   Language model (gpt-4o-mini via OpenRouter) with tools to save lead fields, score and qualify, and hand off to a person.

3. Booking API

   A booking service exposes availability and booking endpoints, protected with a shared secret.

4. Data

   PostgreSQL stores leads, messages, slots and bookings. A unique constraint on the slot prevents double-booking.

5. Human

   The agent hands off with the conversation context when a person is needed, for example price negotiation.

1. A prospect messages on WhatsApp.
2. The agent collects qualifying details and saves them as structured fields.
3. It requests real availability from the booking API and offers slots.
4. It books the chosen slot through the API, which rejects a slot that is already taken.
5. On request for something outside its scope, it hands off to a human advisor with the full context.

• n8n
• OpenRouter (gpt-4o-mini)
• Next.js API routes
• PostgreSQL
• Prisma
• WhatsApp

• Time zones are handled in the API, which returns pre-formatted local times, so the model does no time-zone arithmetic.
• Availability and booking are tools backed by the database, so the model cannot invent a slot.
• Scoring and hand-off are explicit tools, which keeps the decision points visible.

• Double-booking is prevented by a database unique constraint.
• The booking API requires a Bearer secret; secrets are held in credentials, not in workflow files.
• Human handoff with context rather than the AI improvising on out-of-scope requests.

• The booking flow was tested end to end over a live WhatsApp conversation.

Outcome: Live on my own venture, with the booking flow verified end to end over WhatsApp. No performance metrics are published.

Limitations: Built for real estate, not a clinic. A healthcare version would use clinic-approved content, clinic scheduling rules, and a review of privacy and consent requirements first.