WhatsApp Sales Qualification Bot

The Problem

A real-estate agency was missing leads that came in outside business hours — roughly 40% of their total inquiry volume. Their agents worked 9-to-6; WhatsApp messages sent at 11pm sat unanswered until morning, by which time the prospect had messaged three competitors.

What I Built

A WhatsApp Business API chatbot that handles the full top-of-funnel qualification flow:

1. Greeting — Natural language welcome, detects language automatically
2. Qualification — Asks 5 structured questions (budget, timeline, property type, location preference, financing status) in conversational style
3. Scoring — Lead score 1–10 based on qualification answers + response time
4. CRM handoff — Writes qualified lead to HubSpot via API with full conversation transcript
5. Routing — Hot leads (score ≥ 7) trigger an immediate SMS to the assigned agent

The bot uses GPT-4o for response generation but runs all qualification logic through deterministic state machines — I don’t trust LLMs to reliably extract structured data from freeform conversation, so I use the LLM for language and a finite state machine for logic.

Technical Stack

• Meta Graph API for WhatsApp webhook integration
• Node.js + Express for the webhook receiver
• Supabase for conversation state and lead records
• OpenAI GPT-4o for natural language generation
• Bull queue for reliable message processing (no dropped webhooks)

The Qualification State Machine

Each conversation has a state (greeting → q1 → q2 → q3 → q4 → q5 → handoff). The LLM generates the message text; a separate extraction step parses the user’s reply into a structured answer. If extraction fails, the bot gracefully re-asks. This made the system reliable across messy, abbreviated WhatsApp-style responses.

Results

• 62% of after-hours inquiries now captured and qualified (was 0%)
• Average qualification time: 3 minutes (human agents took 12–15 minutes per call)
• Lead-to-meeting conversion up 28% — agents only call pre-qualified prospects
• Zero complaints about bot-feel — most prospects didn’t know they were talking to an agent