If your customers are in Nigeria, much of Africa, India, Latin America or large parts of Europe, there is a good chance they would rather message you on WhatsApp than fill in a web form. Meeting them there is one of the highest-value automations I build.
It is also the one where cutting corners can cost you your phone number. So here is how to do it properly.
Why WhatsApp
The reason is not novelty. It is friction. A customer who has to find your website, locate the contact form and wait for an email is a customer who often gives up. A customer who can send a message from an app they already have open is a customer who asks the question.
For a business, that means more enquiries answered, faster, in a channel where people expect a quick reply. An AI assistant covers the repetitive questions at any hour and passes the rest to you.
The app vs the API
There are two products with similar names, and confusing them is the most common early mistake.
| WhatsApp Business app | WhatsApp Business Platform (API) | |
|---|---|---|
| Who it is for | One person on a phone | Systems that send and receive at scale |
| Automation | Basic auto-replies only | Full programmatic control |
| Multiple agents | Limited | Yes, through connected software |
| Fits an AI assistant | No | Yes |
An AI chatbot needs the Platform. You reach it either directly through Meta's Cloud API or through an approved provider such as Twilio or 360dialog, which handle some of the setup and support for a fee.
The rules that matter
The 24-hour window
When a customer messages you, a window opens in which you can reply freely. After it closes, you can only send pre-approved message templates. Your design has to account for this, because a follow-up sent two days later is a template, not a free-form message.
Templates need approval
Templates are submitted to Meta and reviewed. Write them plainly, keep them genuinely useful, and expect a round or two of revision.
Opt-in
You need a person's permission before you message them first. A customer who writes to you has opened the conversation. A scraped list of numbers is the fastest route to complaints and restrictions.
Do not use unofficial automation tools. Libraries that drive a normal WhatsApp account through a hidden browser session are tempting because they are free and fast. They breach the terms of service and numbers get banned, sometimes permanently. For a business, losing the number customers already know is a serious outage.
How the system fits together
- A customer sends a message to your WhatsApp number.
- The platform forwards it to a webhook, a web address your automation listens on.
- An n8n or Make.com workflow receives it, looks up the customer, and loads the conversation so far.
- An AI step decides what is being asked and drafts a reply, grounded in your own information. How to ground an assistant is covered in RAG explained.
- The workflow sends the reply back through the API and records the exchange.
- If the message is sensitive, or the AI is unsure, the workflow hands over to a human and tells them what was said.
If you want the agent to decide which tool to use, such as an order lookup or an escalation, building your first AI agent in n8n walks through exactly that.
Designing the conversation
- Introduce yourself honestly. "I am the assistant for [Business]. A person is available any time you ask."
- Keep replies short. WhatsApp is read on a phone, in a hurry. Three sentences beats a paragraph.
- Ask one question at a time. Multi-part questions get half-answered.
- Make the human route obvious and instant. If someone types "agent", they get one.
- Escalate on emotion. Anger, a complaint or a refund request goes to a person immediately.
- Hand over with context. The human should see a summary, not start from "hello".
Mistakes that get numbers banned
- Messaging people who never opted in. The single biggest cause of blocks and reports.
- Sending identical bulk messages. Personalise, and only message people who expect to hear from you.
- No easy way to stop. Respect "stop" immediately and permanently.
- Pretending to be human. Disclosure builds trust and avoids the awkward discovery.
- Letting the bot answer everything. Refunds, complaints and anything legal belong with a person.
Build it staged: let the assistant draft replies for a human to approve for the first week, then let it handle one narrow category such as order status alone, and widen from there. It is slower than switching everything on at once, and it is how you avoid explaining a bad conversation to a customer who screenshotted it.