Why build customer support on WhatsApp?
In India, WhatsApp is the main messaging app for most people, including when they contact a shop, a clinic or a courier. Customers already know the interface, do not need to install anything and expect quick replies on their phone. For support teams, that means a channel with high engagement and a natural place for order updates, bookings and simple service requests.
An AI chatbot lets you answer at any hour in the customer's language, handle the repetitive requests automatically and give your human agents the harder conversations with full context.
What do you need before you start?
Businesses build chatbots on the WhatsApp Business Platform, the official API from Meta, rather than the free WhatsApp Business app. Prepare these basics first:
- A Meta business account and a verified business, since verification affects what your account can do, including messaging limits.
- A phone number that is not in use on a WhatsApp or WhatsApp Business app account (it must be deregistered first).
- A provider or direct Cloud API access, with a webhook endpoint that receives incoming messages and sends replies.
- Customer consent: users should opt in to receive messages from your business, and you need a record of it.
- Approved message templates for conversations that your business starts, such as order updates or appointment reminders.
Platform rules change over time, so confirm the current terms in Meta's documentation before you plan.
How does a WhatsApp AI chatbot work?
A modern WhatsApp chatbot has five parts working together:
- The channel layer receives each WhatsApp message through a webhook and sends replies back, handling text, images, documents, buttons and lists.
- Understanding uses a large language model to work out what the customer wants, in English, Hindi, Gujarati or a mix, including messages written in Roman script.
- Knowledge retrieves the answer from your approved content, such as FAQs, policies and product documents, so the reply is grounded and can be traced.
- Actions connect to your systems so the chatbot can check an order, update an address or book a slot, each through a tool with narrow permissions.
- Handoff passes the conversation to a person, with a summary, when the request is outside policy or the chatbot is unsure.
For the engineering view of how agents use tools safely, see our guide what is an AI agent.
The 24-hour window and message templates
WhatsApp distinguishes between two kinds of conversation, and the chatbot has to respect both:
| Situation | What you can send | Design implication |
|---|---|---|
| Customer messages you | Free-form replies during the customer service window, which is 24 hours from their last message | The chatbot can answer naturally and take the conversation wherever it needs to go |
| You message first, or the window has closed | Only pre-approved templates | Write clear templates for order updates, reminders and follow ups, and get them approved in advance |
A good design keeps important steps inside the window where possible, for example collecting every detail needed to resolve a request in the same conversation.
Designing conversations that actually resolve requests
WhatsApp is a small screen with short attention. A few design principles help:
- Start with your top intents. Review real tickets and chats, then automate the most frequent, rule-based ones first: order status, delivery changes, appointment booking, invoice copies.
- Use buttons and lists for choices. They reduce typing, errors and ambiguity, especially for authentication and menu steps.
- Keep replies short. One idea per message, in the customer's language and in your brand voice.
- Ask for one thing at a time. A single clear question is answered far more reliably than three.
- Verify before sharing personal data. Confirm identity with a one-time code or known details before showing account information.
- Never guess. When confidence is low, say so and offer a person.
Handing over to a human agent
The handoff is where customer trust is won or lost. A good one has three properties. The chatbot recognises the moment early, either because the topic is out of policy, the customer is frustrated, or confidence is low. It passes a summary and the key facts so the customer never repeats themselves. And it tells the customer what happens next, including expected response time. Connect the handoff to your existing helpdesk, whether Zendesk, Freshdesk, Intercom or an in-house tool, so agents work in one place.
Languages: English, Hindi, Gujarati and mixed messages
Customers write in their own mix of languages, often Hindi or Gujarati typed in English letters. Modern language models handle this well, but every language and style should be tested on real conversations before launch, including spelling variations, voice notes and screenshots. Decide up front which languages you will support fully, which you will answer with a handoff, and how the chatbot should reply when a customer switches language mid-conversation.
Testing, launch and what to measure
- Build an evaluation set from real past conversations with the correct outcome for each.
- Run in shadow mode. The chatbot drafts, your agents approve, and you measure agreement before customers see anything.
- Launch to a small share of traffic and widen as the numbers hold.
- Review failed conversations weekly. Most improvements come from reading the conversations that went wrong.
Measure resolution rate (requests fully solved without a person), handoff quality, customer satisfaction, first response time and wrong-answer rate. A chatbot that deflects without resolving has not helped anyone.
To see the approach in action, chat with our live support agent demo, or read about our AI chatbot development services and contact us to plan a WhatsApp pilot.