WhatsApp is where many users already live, which makes it a powerful channel — and an easy one to ruin. A bot that misunderstands, loops, or spams will be muted in seconds. Building one people tolerate, let alone like, takes restraint.
Start with the webhook, not the model
The foundation is a reliable integration with the WhatsApp Business API: verified webhooks, idempotent message handling, and clear state per conversation. Get delivery and ordering right before you add any intelligence. Most bad bot experiences are plumbing failures, not model failures.
Be useful in one message
People do not want a chat; they want an answer. Lead with the most likely action, confirm understanding in plain language, and always offer an obvious way to reach a human. An assistant that knows its limits earns trust.
Let AI handle phrasing, not policy
Use the model to understand intent and write a clear reply. Keep the rules — eligibility, limits, what is allowed — in code you control. The model interprets; your system decides.
Built this way, automation can run at national scale with zero manual overhead and still feel respectful of the person on the other end.