An AI chatbot on WhatsApp is an automated assistant connected to the WhatsApp Business Platform that answers customer messages, collects details and hands complex cases to a human. Moving from pilot to production requires an approved business account, message templates, clear opt-in, a tested handover path and tracking. Skipping any of these causes most failed launches.
How does a WhatsApp AI chatbot work?
Businesses connect to WhatsApp through the WhatsApp Business Platform, owned by Meta, either directly with the Cloud API or through a Business Solution Provider such as Twilio, 360dialog or Infobip. Incoming messages arrive at your system by webhook. Your chatbot logic, often a large language model with access to your FAQs, order data or booking system, decides the reply and sends it back through the API.
WhatsApp has rules that shape the design. A business can reply freely within a 24-hour customer service window after the customer’s last message. Outside that window, the business may only start conversations with pre-approved message templates, and pricing differs by conversation or message category. Check Meta’s current pricing page, since the model has changed more than once.
What should a pilot prove before scaling?
A pilot should answer three questions: does the bot resolve real requests, do customers accept it, and does the handover to staff work under load. Choose a narrow scope, such as order status or appointment booking, and run it with a limited number of customers or one region.
- Pick one or two intents that make up a large share of current messages
- Define what “resolved” means for each intent
- Set a target for handover time to a human agent
- Collect every conversation transcript for weekly review
- Run the pilot long enough to include a busy period, such as a campaign week
Implementation checklist: account and compliance
The administrative work takes longer than most teams expect. Start it before the bot is built.
- Verify the business in Meta Business Manager (business verification)
- Register a phone number that is not already used in the WhatsApp consumer or Business app
- Set the display name and check it follows Meta’s naming guidelines
- Write and submit message templates for order updates, reminders and re-engagement
- Record opt-in clearly: customers must agree to receive WhatsApp messages from you
- Review data handling under KVKK or GDPR, including where transcripts are stored and for how long
- Add an opt-out keyword and honour it immediately
Implementation checklist: the bot itself
The AI part needs limits more than it needs cleverness. An LLM that can say anything will eventually say something wrong about a refund policy.
Ground the answers
Give the model a curated knowledge base: policies, product details, delivery rules. Use retrieval so answers come from these documents, and instruct the bot to say it does not know when the documents do not cover a question. Keep prices and stock in live system lookups, not in static text.
Design the handover
Customers must be able to reach a person. Define triggers: the word “agent”, repeated failed answers, negative sentiment, payment problems. Pass the conversation summary to the agent’s inbox so the customer does not repeat everything. Tools like Zendesk, Freshdesk or HubSpot can receive these handovers.
Write for a chat screen
Messages should be short. Use WhatsApp’s interactive buttons and list messages for choices instead of asking customers to type. Test on a real phone, because a reply that looks fine in a dashboard can fill the whole screen.
Which metrics matter in production?
Track a small set every week and look at transcripts behind the numbers. A good containment rate means little if customers leave angry.
- Containment rate: share of conversations resolved without a human
- Handover rate and average wait time after handover
- Customer satisfaction from a one-question survey at the end
- Template message delivery and read rates
- Opt-out rate after outbound templates
- Conversions, such as bookings or recovered carts, attributed to the channel
Review a random sample of 30 to 50 conversations weekly in the first months. Patterns of confusion show up quickly, and fixing a knowledge base entry is usually faster than tuning the model.
How Alien Road applies AI chatbots on WhatsApp
A typical engagement begins with discovery: exporting a few weeks of existing WhatsApp or live chat messages, grouping them by intent and estimating which intents a bot can answer reliably. This tells the client where the volume is before any build starts.
Setup covers Business Manager verification, the provider choice, template writing and the knowledge base. The team connects the bot to the client’s CRM or order system for live lookups and designs the handover into the support tool the client already uses. A pilot runs with one intent and a limited audience, with weekly transcript reviews.
Reporting tracks containment, handover time, satisfaction and opt-outs next to the pre-launch baseline from the old chat data. New intents are added only when the current ones meet their targets.
Which mistakes stall WhatsApp chatbot projects?
Most delays come from process, not technology. Templates get rejected because they read like promotions when submitted as utility messages. Phone numbers get stuck because they were already registered in the WhatsApp Business app. Bots launch without a handover path and frustrate the first customers who need a person.
- Submitting templates with vague wording or missing variables
- Sending marketing templates to people who never opted in
- Answering in one language when customers write in two
- Letting the bot guess order details instead of looking them up
- Forgetting to staff the handover queue outside office hours
Each of these is cheap to prevent and expensive to fix after a bad week of customer complaints.
Where to start this week
Export the last month of customer messages from your WhatsApp Business app or chat tool and tag 100 of them by topic. The two biggest topics are your pilot scope.
In parallel, start Meta business verification, because approval can take time. Draft three message templates you will need on day one, such as an order update, a booking reminder and a follow-up, so they are approved before the bot is ready.
Keep reading: AI Marketing Agency · AI Marketing