Chatbots · 2026-01-22 · 11 min read
WhatsApp AI Chatbot Architecture
Reference architecture for a production WhatsApp AI chatbot: BSP, orchestration, RAG, CRM write-back, human handoff, and analytics.
A WhatsApp AI chatbot that works in production needs more than a prompt. It needs channel compliance, CRM write-back, retrieval, and escalation.
## Reference architecture
1. **WhatsApp Business API** via an approved BSP
2. **Orchestration layer** (n8n or custom service) for session state and tools
3. **LLM + retrieval** grounded in approved FAQs and product docs
4. **CRM / database** for identity, orders, and lead fields
5. **Human handoff** with full transcript into your inbox or helpdesk
6. **Analytics** for containment, escalation rate, and conversion
## Critical design choices
- How do you identify the user (phone, login link, order ID)?
- What can the bot promise without a human?
- When must it refuse or escalate?
- Which languages matter for your market (especially Nigeria and multilingual audiences)?
- How do opt-outs and template messaging rules work?
## What breaks demos
- Hallucinated prices or policies
- No CRM sync, so sales cannot follow up
- No human takeover path
- Ignoring WhatsApp template and session constraints
## Our build pattern
We start with a narrow job: FAQs + lead capture, or order status, or appointment booking. Then we expand tools once containment and accuracy are measured.
## Next step
If WhatsApp is already your front desk, bring message volume and your CRM. We will propose a containment target and a phased rollout.