An AI receptionist on a real WhatsApp number: architecture, routing logic, and failure modes

Description and use

A self‑hosted AI receptionist on a regular WhatsApp number uses Evolution API, n8n, and an LLM to process messages with filtering, human handoff, and office‑hours logic.

Problem

WhatsApp is the primary front desk in Gulf and Levant regions; human operators must constantly monitor messages, leading to high labor costs and missed opportunities.

Outcome

The architecture delivers a reliable, low‑cost AI receptionist that handles routing, human handoff, and office‑hours logic, while persisting sessions to avoid QR rescans and mitigating version drift and unofficial‑client risks.

Workflow

• 1. WhatsApp messages received via Evolution API webhooks
• 2. n8n workflow filters out non‑text, outgoing, and group messages
• 3. Detects human handoff keyword and stops bot replies, notifying team
• 4. Checks office hours; if closed, sends automated out‑of‑office reply
• 5. For remaining messages, calls LLM with system prompt containing business facts and language preference
• 6. Sends LLM reply back to WhatsApp
• 7. Persist session and message data in Postgres; cache in Redis

Evidence

Four services: WhatsApp ⇄ Evolution API → webhook → n8n workflow → LLM (Claude/GPT) ↑ │ └── send reply ──────────┘ Postgres (sessions, messages) · Redis (instance cache)

Source

The naive version — "forward every message to an LLM" — fails in production within a day.

Source