Building a WhatsApp AI Chatbot for Business
By Darus S Alfatah · 2025-09-11 · Implementasi · 9 menit
Author note: Reply fast, then complete. Holding a reply until perfect makes customers think the bot is dead.
Architecture in Three Parts
A WhatsApp bot has a webhook that receives messages, a logic layer that calls the model, and a sender that replies. Keeping these separate makes each part testable, and lets you replace the channel later without rewriting the logic.
Store conversation state per phone number: which menu the customer is in, what information is missing, and whether a human has taken over. Without state, every exchange restarts and the bot feels broken.
Handling Real Messages
Customers write fragments: "how much?", "still available?", or send a sticker. Decide early what you support and what you route to a human. Guessing on ambiguous messages produces confident wrong answers, which is worse than admitting uncertainty.
When a human joins, disable automatic replies for that conversation so the customer does not receive two competing answers.
Quick Replies
Use buttons or numbered options for flows with limited answers: order status, opening hours, payment methods. Choice-based interaction is more accurate than free-text understanding and faster for the customer.
Compliance and Etiquette
Follow the platform's messaging rules, including template requirements for messages sent outside the customer service window. Tell users they are talking to an automated assistant and give them a fast way to reach a person.
Data Handling
Chats contain phone numbers, addresses, and payment details. Store the minimum needed, redact the rest before sending anything to a model, and limit staff access to conversation history.
Frequently Asked Questions
Which provider should I use? The official business API; unofficial libraries risk bans.
How do I test before launch? Run a private number with real staff as customers for a week.
What about voice notes? Transcribe them if the model supports audio, otherwise ask for text.
How do I measure success? Self-service rate, first response time, and escalations.
What is the biggest risk? Answering confidently when the answer is unknown.
Conclusion
A WhatsApp bot is mostly conversation design and data discipline. Get the state machine and escalation right, and the model choice becomes secondary.
Improving Over Time
Review fifty conversations a month and look for repeated failures: the same question answered poorly, or moments where the customer repeated themselves. Those patterns point directly at missing content rather than model limitations.
Update the answer bank and the menu structure accordingly. Small, regular edits keep the bot useful; large rewrites every six months lose the knowledge accumulated in between.
Handling Escalations Well
When handing over, pass context to the human: what the customer asked, what the bot already answered, and any order number collected. Agents who must start from scratch will ask the customer to repeat everything, which is worse than having no bot at all.
Measure escalation rate weekly. A rising rate is an early warning that your content or flows no longer match what customers need.
Launch Checklist
Test with ten real customers, confirm the handover works, verify that opt-out requests are honoured, and check that the bot never sends two replies to one message. Run through the list again after any change to the flows.
Example: reply fast: Reply got it, checking now within a second, then send the full answer once data is available. The user feels served from the start.
Frequently Asked Questions
What do I get from this implementation guide?
Concrete steps to connect an AI API to your system, including example configuration and pitfalls to avoid.
Are there code samples I can use?
Yes, the article includes examples you can adapt to your stack.
How do I handle errors during integration?
Use the API error messages for diagnosis; BankAI docs explain common error codes and fixes.