How to Automate WhatsApp with AI to Serve Customers Better
Learn how to automate WhatsApp with AI step by step: FAQs, conversation flows, a knowledge base, Meta's rules and when to hand off to a real person.
If your WhatsApp fills up with “price?”, “hours?” and “do you have availability?”, an AI assistant can answer those instantly and leave your team the conversations that truly need a person. In this tutorial you'll design it from start to finish: what it should answer, how it's organized, what Meta allows and when to hand the chat to a real human.
Picture a Friday at eleven at night. A tourist who just landed in Cancún messages a restaurant to book a table for eight the next day. Nobody replies until Saturday at noon, and by then they've booked somewhere else. It isn't carelessness: WhatsApp never closes, but your business does.
Automating with AI doesn't mean putting a cold robot in front of your brand. It means every message gets a useful answer at any hour, in your business's tone, and your team walks into every conversation with the context already in order. The difference between a customer who feels well looked after and one who feels like they're talking to a wall is in the design, not the tool.
Before you start: what you need to know
The WhatsApp Business app is not the same as the platform
There are two paths, and it pays not to mix them up:
- The WhatsApp Business app is the one you install on your phone. It's free and comes with greeting messages, away messages, quick replies and labels. For a small business it's a good first step, but it won't let you connect your own AI assistant that pulls from your information.
- The WhatsApp Business Platform is the business version that connects through an API. This is where you can integrate an AI assistant, your booking system or your CRM. You can access it directly through Meta or, more commonly for mid-sized businesses, through a solution provider that handles the technical side.
This tutorial focuses on the second path, but the design work (steps 1 to 4) applies even if you're using the app today.
Meta's rules you should know
These are the rules in effect according to Meta's official documentation as of October 2026:
- 24-hour customer service window. According to the platform's pricing documentation, when a customer writes to you, a 24-hour customer service window opens and resets with every new message from that customer. While it's open you can reply freely, including with your assistant's messages. Once it closes, you can only write using templates.
- Approved templates. Templates are predefined messages that Meta reviews before you can use them. According to the template guidelines, each one must be categorized as marketing, utility or authentication, and can only be sent once its status is approved.
- Consent (opt-in). According to the official opt-in guide, you need a person's permission before messaging them, you must make it clear that they're agreeing to receive messages from your business and name your business, and you must comply with applicable law. Meta recommends asking for consent separately for each type of message (for example, order updates and promotions).
- Paths to a person. The WhatsApp Business Messaging Policy allows automation in your replies, but requires clear, direct escalation paths, such as a transfer to a human agent within the chat, a phone number, an email address or a form. It also requires you to honor any request to stop receiving messages.
- No general-purpose AI as a product. Meta's terms for the WhatsApp Business Platform prohibit using the platform to offer AI technologies when they are the primary functionality being made available, rather than incidental or ancillary to it. In practice: an assistant that serves the customers of your restaurant, clinic or hotel is fine; a “ChatGPT on WhatsApp” that answers anything is not. The same terms limit the use of conversation data to train AI models, so ask your provider how they handle it.
Meta charges per message based on its category, and its billing rules change: before you estimate costs, check the current pricing page.
Step by step
1. Gather and group the real questions
AI will only be as good as the information you give it, and that information starts with knowing what people ask you. Open your WhatsApp and go through last month's conversations. Copy every distinct question into a document, in the customer's exact words (typos and abbreviations included: that's how they'll write to you).
Then group them. Most businesses end up with similar families:
- Information: hours, location, parking, payment methods, invoicing.
- Offer: menu or catalog, prices, promotions, availability.
- Action: book, schedule, order delivery, request a quote.
- Follow-up: “has my order shipped?”, “I want to change my appointment.”
- Problems: complaints, refunds, something that went wrong.
In each group, mark which questions are repetitive with a fixed answer (those go to the assistant) and which call for judgment, empathy or a decision (those go to a person).
The common mistake is designing from the office around what you “think” people ask.
2. Map out the conversation flow
Even if the AI understands natural language, you need a map. Sketch on a sheet of paper what happens from the very first message:
- Greeting and introduction. The assistant says who it is and that it's an automated assistant. Being transparent keeps anyone from feeling tricked.
- Intent detection. Does the customer want information, a booking, a follow-up, or do they have a problem?
- Resolution or data collection. If it's a common question, answer it. If they want to book, ask for the details one at a time (date, time, party size, name).
- Confirmation. Repeat back what was understood so the customer can confirm it.
- Wrap-up or handoff. Either it's resolved, or it goes to a person with a summary.
An illustrative greeting for a restaurant:
“Hi, I'm La Terraza's assistant. I can help you with the menu, hours and reservations any time of day. If you'd rather talk to someone on our team, just type ‘person’ and I'll connect you.”
Notice that the path to a human is already in that first sentence. Don't hide it.
The common mistake: endless button trees. With AI, customers can write the way they talk; use buttons only as shortcuts.
3. Write the knowledge base
The knowledge base is the document your assistant draws its answers from. It's the most important part of the whole project and the most neglected. It should be:
- Complete: hours for each day (holidays included), address with landmarks, menu or catalog with descriptions, and your booking, cancellation, payment and invoicing policies.
- Current: if the menu or seasonal hours change, update it the same day. An outdated answer delivered with confidence is worse than no answer at all.
- Explicit about what it doesn't know: write down which topics the assistant should not answer and what it should do in those cases (for example, “if someone asks about severe allergies, don't confirm anything; hand off to the manager”).
Write it in clear sentences, one idea per line. If a new server wouldn't understand the answer, the AI won't deliver it well either.
4. Define the assistant's voice and limits
Your assistant speaks on behalf of your brand, so it needs tone guidelines just like an employee would. Write a short profile:
- How it addresses customers: casual or formal, warm or more reserved.
- Reply length: on WhatsApp, keep messages short. Two brief messages beat one giant paragraph.
- What it never does: make up prices or availability, promise discounts that don't exist, weigh in on topics unrelated to the business, ask for sensitive data it doesn't need.
- What it does when it doesn't know: admit it and offer to connect the customer with a person, instead of improvising.
This profile also keeps you on the right side of Meta's rules: an assistant limited to your business is clearly a customer service assistant, not a general-purpose AI.
A good assistant knows how to answer, and also when to stay quiet.
5. Design the handoff to a person
This is the step that separates automation that takes care of customers from automation that loses them. Define precisely when the conversation goes to your team:
- When the customer asks for it (“I want to talk to someone,” “person,” “agent”).
- When there's a complaint, a refund or strong emotion.
- When the assistant can't find the answer in the knowledge base.
- When the sale is large or complex (a private event, a group, a custom quote).
And define how it happens:
- The assistant lets the customer know: “I'm connecting you with someone from our team. Our hours are 9 a.m. to 6 p.m.; if you write outside those hours, we'll get back to you as soon as we open.”
- Your team receives the conversation with a summary: who the customer is, what they want and what details they've already given. Nobody should have to ask the customer the same thing twice.
- While a person is handling it, the assistant stays quiet in that conversation.
Remember the 24-hour window: if your team takes more than a day to pick things back up, you won't be able to reply freely anymore and will need an approved template. Set up shifts or alerts so that doesn't happen.
6. Test before going public
Before connecting it to your main number, have your team try to break it. Ask them to write like real customers: in a hurry, with typos, with two questions in one message, angry, with absurd questions. Note every wrong answer and fix the knowledge base or the instructions, not just that one reply.
At a minimum, check that it answers the most common questions correctly, says “I don't know” instead of making things up, hands off to a person the first time it's asked, respects anyone who asks to stop receiving messages and keeps its tone with an upset customer.
7. Measure and adjust every week
Once it's live, set aside some time each week to read conversations. Look for the questions the assistant couldn't answer (that's your improvement list), the handoffs that could have been avoided and the ones that came too late.
Common mistakes when automating WhatsApp
- Hiding that it's an automated assistant.
- Not offering a clear path to a person, which WhatsApp's policy also requires.
- Letting the AI make up prices, availability or promotions.
- Not updating the knowledge base when the menu or hours change.
- Sending promotions to contacts who never gave consent.
- Forgetting the 24-hour window and getting stuck unable to reply.
- Launching without testing and discovering the errors with real customers.
- Using the assistant as an “anything goes” AI instead of focusing it on your business.
Pre-launch checklist
- I have the real questions grouped and know which ones the assistant handles.
- The flow is mapped out, with a transparent greeting and a path to a person.
- The knowledge base is complete, current and says what not to answer.
- The handoff to a person works and arrives with a summary.
- I have consent from the contacts I'll send templates to.
- My team tested the assistant and we fixed what failed.
- One person is responsible for reviewing it every week.
Automated service that keeps the human touch
If you want to see how this looks for specific businesses, in AI applied to marketing we walk through other practical uses. And if your assistant will be getting traffic from ads, our guide to the anatomy of a landing page helps your page and your chat tell the same story.
At BDB, a marketing agency in Cancún, Mexico, we design AI-powered automated service for restaurants across WhatsApp, Instagram and reviews, in each brand's own voice and with clear paths to the team. If you'd like to explore how it could work for your business, get in touch.
Frequently asked questions
Do I need the WhatsApp API to automate with AI?
For an AI assistant connected to your systems, yes. You use the WhatsApp Business Platform, either directly or through a solution provider. The free WhatsApp Business app offers greeting messages, away messages and quick replies, which are fine for getting started, but it won't let you connect your own assistant to your knowledge base.
Does Meta allow AI chatbots on WhatsApp?
Yes, for serving your business's customers. What Meta's terms prohibit is offering a general-purpose AI as the main product through the platform, like an assistant that answers anything at all. An assistant that answers questions about your services, hours and orders counts as customer service.
Can I message any customer whenever I want?
No. You need their prior consent to message them, and outside the 24-hour customer service window that opens when the customer writes to you, you can only send approved templates. Inside that window you can reply freely, including with messages your assistant generates.
Will AI replace my customer service team?
That's not the idea, and WhatsApp's policies require clear paths for customers to reach a person. AI handles the repetitive questions and organizes the information; your team keeps the complex sales, the complaints and the cases that call for judgment.
