Building a WhatsApp AI Agent for Business in 2026
WhatsApp has 2.9 billion users and a 90%+ open rate. Here is the architecture, the template rules, and the real per-conversation cost.
Email open rates sit around 20%. WhatsApp sits above 90%. For businesses in LATAM, South Asia, the Middle East, and southern Europe, a WhatsApp agent is not a channel experiment — it is the primary support and sales surface.
The platform rules you must design around
- 24-hour session window. After a customer messages you, you can reply freely for 24 hours. Outside it, only pre-approved templates.
- Template approval. Outbound templates are reviewed by Meta. Marketing templates get rejected for pushiness; utility templates almost always pass.
- Per-conversation pricing. You are billed per 24-hour conversation, priced by category and country — roughly $0.005–$0.09 for utility, higher for marketing.
- Quality rating. Blocks and reports throttle your sending tier. One bad campaign can halve your daily limit.
Reference architecture
- Webhook receiver — a public endpoint verifying Meta's signature, acknowledging within five seconds, and pushing to a queue. Never process inline; timeouts cause duplicate deliveries.
- Conversation store — Postgres with one row per message, threaded by phone number hash, plus a rolling summary per contact.
- Agent runtime — an LLM with tool access: order lookup, availability check, booking, ticket creation. Tools do the truth-telling; the model only phrases it.
- Guardrails — refuse pricing promises, refuse medical or legal advice, never invent order status. Every factual claim must trace to a tool result.
- Handoff — explicit trigger words, low confidence, or two failed resolutions route to a human with full transcript context.
Cost model for a real deployment
| Component | Monthly at 10k conversations |
|---|---|
| WhatsApp conversation fees | $120–$700 (country dependent) |
| LLM inference (8 turns avg) | $40–$180 |
| Infrastructure | $25–$80 |
| Build (one-off) | $6k–$18k |
What makes these agents fail
Three things, in order of frequency: answering from model memory instead of a tool call, refusing to hand off to a human, and writing paragraphs. WhatsApp is a two-line medium. Cap responses at roughly 300 characters and use quick-reply buttons for anything with fixed options.
Media, voice, and language
Customers send photos of receipts and 40-second voice notes. Handle both: transcribe audio, run images through a vision extraction step. Auto-detect language per message rather than per contact — code-switching mid-conversation is the norm, not the exception.
Measuring success
Track containment rate (resolved without a human), median time to first response, handoff satisfaction, and block rate. Containment above 60% with a block rate under 0.5% is a healthy deployment.
FAQ
Frequently asked questions
How much does a WhatsApp AI agent cost to run?+
At 10,000 conversations a month expect roughly $185–$960 all-in, dominated by Meta's per-conversation fees which vary heavily by country.
Do I need the WhatsApp Business API?+
Yes for any automated agent. The consumer and Business app do not support programmatic messaging at scale; use the Cloud API or a BSP.
Can I message customers first?+
Only with a pre-approved template outside the 24-hour session window, and only to contacts who opted in.
How do I stop the agent hallucinating order details?+
Give it tools for every factual lookup and instruct it never to state a fact that did not come from a tool result. Log any response that lacks a tool call.
Building something similar?
Let's talk in 30 minutes.