To set up an AI chatbot on WhatsApp, you connect a business phone number to the WhatsApp Business Platform, give the bot a limited set of jobs (answer common questions, capture lead details, route the rest to a person), and test it on real conversations before going live. Decide first what the bot is allowed to do, what it must hand to a human, and where the lead data should end up.
What should I decide before building anything?
Decide the bot's job, its limits and its destination for data. Most failed chatbots skip this and start from the tool. Answer these five questions on paper first:
- Which questions repeat most? Open your last 50 to 100 WhatsApp conversations and count. Opening hours, prices, availability, order status and "how do I book" are typical.
- What is the one main goal? Answering faster, capturing leads outside office hours, or taking bookings. Pick one for version one.
- What must never be answered by the bot? Complaints, refunds, custom quotes, medical or legal questions, anything where a wrong answer costs you money or trust.
- Who takes over, and when? Name a person and the hours they are reachable.
- Where do leads go? A CRM like HubSpot, a Notion database, a Slack channel or a spreadsheet. If nobody looks at it, the capture is pointless.
If you are unsure whether WhatsApp is the right first thing to automate, this prioritization guide helps you compare it with other options.
What can a WhatsApp chatbot realistically do for a small business?
It can handle repetitive, well-defined conversations. It is weaker at anything that needs judgment or access to information you have not given it.
| Use case | Fit | Notes |
|---|---|---|
| Answering FAQs (hours, location, services) | Good | Needs an accurate, up-to-date source of answers |
| Capturing lead details (name, need, contact) | Good | Ask few questions; send to your CRM |
| Booking or qualifying appointments | Good, with integration | Requires access to your calendar or booking tool |
| Order status | Possible | Needs a connection to your store, such as Shopify |
| Complaints and refunds | Poor | Hand off to a person |
| Custom quotes and negotiation | Poor | Collect the details, then hand off |
If you wonder how an AI-based bot differs from the old menu-style ones, agents versus traditional chatbots explains it.
How do I connect a chatbot to WhatsApp?
You use the WhatsApp Business Platform (the API), either directly through Meta or through a provider that sits on top of it. The regular WhatsApp Business app is fine for manual replies and simple auto-messages, but a real AI chatbot generally needs the API. Check Meta's WhatsApp Business Platform documentation for current requirements, since they change.
- Create a Meta Business account and verify your business if asked. Have your legal business details ready.
- Choose a phone number. It usually can't be registered at the same time in the regular WhatsApp app, so plan for a dedicated number or a migration.
- Set up API access directly or through a provider. Providers add their own fees on top of Meta's.
- Prepare message templates. Messages you initiate outside the 24-hour customer service window must use templates approved by WhatsApp.
- Connect the bot's brain. This is the AI model plus your business information (FAQs, services, policies).
- Connect your tools so conversations create records in your CRM, calendar or store.
- Test, then launch to a small group first.
Costs depend on Meta's conversation pricing, your provider, and the AI model usage. Rates change, so check the current pricing pages rather than relying on figures from articles.
How should the bot hand off to a human?
Define clear triggers and make the handoff visible to the customer. This matters more than the AI quality. A bot that traps people is worse than no bot.
- Explicit request: if the customer asks for a person, hand off right away.
- Repeated failure: if the bot misunderstands twice, stop trying.
- Sensitive topics: complaints, payments, cancellations, anything you listed as off-limits.
- Frustration signals: angry wording or very long back-and-forth.
When it hands off, the bot should say so, tell the customer when to expect an answer (based on your real hours), and pass along a short summary so nobody has to ask the same questions again. Notify your team where they already work, for example a Slack channel. For how much freedom to give the bot, see autonomous vs. semi-autonomous agents.
How do I capture leads without annoying people?
Answer first, ask second, and keep it to two or three questions. Someone asking about a price wants the price, not a form.
- Answer the question they asked.
- Ask what they need in their own words.
- Ask for the one contact detail you really need (WhatsApp already gives you their number, so often you only need a name).
- Confirm what happens next and when.
Be careful with personal data. Tell people the chat is handled by an automated assistant, state briefly how you use their data, and follow the privacy rules that apply in your country. Security and ethics in agent deployment covers the broader principles.
How do I test the bot before customers see it?
Test with real past questions, not invented ones. Take 30 to 50 actual messages from your history, including the messy ones with typos, voice notes and several questions in one, and check each answer.
- Does it answer correctly from your information, and say "I don't know" when it should?
- Does it invent prices, policies or availability? Fix this before launch.
- Does the handoff work at every trigger?
- Do leads actually land in the right place with the right fields?
After launch, read conversations weekly for the first month. You will find gaps, and each one is a quick fix. Decide up front what you will measure, such as response time, share of conversations resolved without a person, and leads captured. Measuring the ROI of AI agents goes deeper.
What usually goes wrong?
- Outdated information. The bot repeats old prices or hours. Assign someone to update its sources whenever the business changes.
- Too broad a scope. A bot meant to do everything does most things badly. Start narrow.
- No owner. Nobody reviews conversations, so errors persist.
- Template rejections and number issues. Allow time for approvals and number setup; don't plan a launch for tomorrow.
- Ignoring voice notes and images. Many customers send them. Decide how the bot responds, even if only by routing to a person.
Should I build it myself or get help?
If your needs are simple (a few FAQs and an out-of-hours message), a no-code provider may be enough. If you need integrations with your CRM, store or calendar, reliable handoff, and ongoing tuning, help saves time. At Kuan AI we build chatbots for WhatsApp, Instagram and the web, working in stages: diagnosis, design, implementation, then measuring and scaling what works. We connect to tools you already use, such as HubSpot, Slack, Notion, Zapier or Shopify. Scope and terms are agreed per project. You can tell us about your case through the contact form.
WhatsApp chatbot: common questions
Can I use the regular WhatsApp Business app for an AI chatbot?
Usually not. The app suits manual replies and simple auto-messages. An AI chatbot generally needs the WhatsApp Business Platform (API), directly or through a provider. Check Meta's current documentation.
Do I need a separate phone number?
Often yes. A number used on the API typically can't stay active in the regular WhatsApp app at the same time, so many businesses use a dedicated number or plan a migration.
Will the chatbot replace my staff?
No, it takes the repetitive questions. People still handle complaints, quotes, exceptions and anything sensitive. A clear handoff is part of the design.
How do I stop the bot from making things up?
Limit it to your own approved information, tell it to say it doesn't know when unsure, test with real past questions, and review conversations weekly after launch.
Where should captured leads go?
Wherever your team already works: a CRM like HubSpot, a Notion database or a Slack channel. Choose it before building, and make sure someone follows up.
Start small: one channel, one goal, a clear handoff. Once that works, expanding it is much easier than repairing a bot that tried to do everything.