How We Automated WhatsApp Customer Support with AI

A client came to us with a concrete problem: their 3-person team was drowning in WhatsApp messages 14 hours a day, answering the same questions over and over. Prices, hours, availability, order tracking. The same 20 questions, repeated thousands of times.

We built them an agent. Here’s what we learned along the way.

The stack: less is more

After trying several combinations, we settled on:

• WhatsApp Business API (official, via Meta Cloud API — no unofficial workarounds).
• n8n as the workflow orchestrator.
• An LLM (Claude or GPT-4, depending on the case) for understanding and response.
• Vector DB (Qdrant) for RAG over the client’s catalog and FAQs.

All running on a $20/month VPS. You don’t need Kubernetes for this.

What worked

The key wasn’t the LLM. It was prompt engineering grounded in real business context. We loaded FAQs, tone of voice, clear boundaries (“never promise discounts”, “if they ask about refunds, escalate”), and examples from real conversations.

Result: 80% of queries resolved without a human. Response time: <3 seconds. The client recovered 20 hours per week of team time.

What broke (and how we fixed it)

• Context window: long conversations lost coherence. Fix: automatic summarization every 10 messages, stored in the DB, injected as context on each turn.
• Escalation logic: the agent tried to answer things it shouldn’t (complaints, legal claims). Fix: an intent classifier in front that marks messages as “human required” and pings the team on Slack.
• Price hallucinations: the LLM invented numbers. Fix: strict RAG — the agent can only quote prices coming from the vector store, never from memory.

What we’d do differently

If we started over, we’d kick off with a much smaller scope. Our initial MVP tried to solve everything. Better: 5 well-defined intents, polished, then expand.

Also: measure from day one. Resolution rate, escalation rate, satisfaction. Without metrics, you’re flying blind.

The lesson

AI doesn’t replace the team. It replaces repetitive work so the team can focus on what matters: complex cases, relationships, sales. That’s the kind of automation worth building.