WhatsApp and messaging agents that follow the rules

0:00 WhatsApp and messaging agents

In Karachi, Dubai and Riyadh, customers don't fill in contact forms. They send a WhatsApp message, often at eleven at night, and expect an answer. Automating WhatsApp and social DMs can turn response times from hours into seconds. But these channels have rules email doesn't: opt-in, message windows, pre-approved templates and policies about what kind of AI may run on them. Break them, and your number can be restricted. In this lesson you'll learn those rules, and how to design a messaging agent that respects them.

0:37 Analogy: shop etiquette

Here's an analogy. WhatsApp's rules are like the etiquette of a shop that customers invited you into. When a customer walks up and asks a question, you can chat freely for a while. Once they've left, you can't follow them down the street shouting offers; you can only send the kind of note they agreed to receive. The twenty-four-hour window and templates are that etiquette, enforced.

1:06 App vs platform

First, app versus platform. The WhatsApp Business app is for small teams replying by hand, with quick replies, labels and away messages. The WhatsApp Business Platform, through its Cloud API, gives programmatic access for automations and agents, usually via Meta directly or a Business Solution Provider or messaging platform. That's what automation tools and AI agent platforms connect to.

1:32 Rules 1–3

Now the rules, and check Meta's current policies because they evolve. One: opt-in. You need the person's permission to message them, collected clearly, for example through a checkbox or a click-to-WhatsApp ad they started. Two: the customer service window. When a customer messages you, you can reply freely, including with AI-written messages, for twenty-four hours from their last message. Three: outside that window, you can only send pre-approved templates, categorised as marketing, utility or authentication, and priced per message by category and country.

2:08 Rules 4–6

Four: quality. If many people block or report your messages, your number's quality rating and sending limits drop. Five: AI policy. Since January twenty twenty-six, Meta's business terms prohibit general-purpose AI assistants as the primary service on the Business Platform. Business-specific AI, like support, bookings, order tracking, lead qualification and sales, is still allowed. So keep your agent on your business's topics. Six: always offer a way to reach a person. Similar ideas apply to Instagram and Messenger DMs, including messaging windows and no unsolicited messages. Never auto-DM people who haven't contacted you.

2:48 Agent design

Designing the agent: scope it to business topics like FAQs, bookings, order status and lead capture. In the first message, greet, say it's an AI assistant, explain what it can help with, and tell people how to reach a person. Reply in the customer's language, whether English, Arabic, Urdu or Roman Urdu. Keep messages short and use lists or quick-reply buttons where available. For voice notes, transcribe and confirm understanding. Hand off complaints, payments and anything sensitive with the transcript. Track the twenty-four-hour window. And store consent records, honouring stop requests immediately.

3:28 Simple example: salon booking

A simple example. A salon in Dubai receives a WhatsApp message at ten p.m.: do you have a slot for a blow-dry on Saturday? The agent replies within seconds, says it's an AI assistant, checks the booking tool, offers two times and books one after confirming the name and phone. The next Friday, outside the window, the salon sends an approved utility template: reminder of your appointment tomorrow at eleven. Every message within the rules.

4:01 Worked example: Lahore fashion brand

Here's a worked example. A Lahore modest-fashion brand gets hundreds of WhatsApp messages a day during sales. Through a messaging provider, they connect an AI agent limited to size guidance, delivery times by city, order status via a lookup tool, exchanges and custom-order requests. Complaints and payment issues go to a human inbox. Outside the window, they send only approved utility templates, like your order has shipped, and marketing templates to customers who opted in. Routine questions now get answers in seconds, and the team spends its time on complaints and custom orders.

4:41 Business example (illustrative)

Illustrative numbers for the fashion brand. During its biggest sale, WhatsApp volume reached around two thousand messages a day. The agent answered about seventy percent end to end, mostly sizes, delivery times and order status. Complaints and payment issues, about one in ten, went to people with the transcript attached. The number's quality rating stayed high, because only opted-in customers received marketing templates and every stop request was honoured instantly.

5:11 Hands-on in the lesson

In the hands-on section you'll see the window logic in a few lines of Python: never message opted-out contacts, send free-form replies only inside twenty-four hours, and otherwise only an approved template with the right opt-in, or nothing at all. You'll see a template send with the Cloud API using curl, with the token in an environment variable, and an opening message you can adapt. Most messaging platforms handle this logic in their interface. This is what they're doing, and what you must respect if you build directly.

5:49 Common mistakes

Common mistakes. Building a general chatbot that answers anything, which Meta's business terms now prohibit as a primary service. Sending marketing templates to people who never opted in. Using free-form AI messages outside the window. Hiding the way to reach a person. Ignoring voice notes, which many customers prefer. And failing to record opt-outs, which damages your number's quality rating.

6:15 How you'll know it's working

How will you know your messaging agent is working? Routine questions are answered in seconds. Hand-offs for complaints and payments reach a person quickly. Your number's quality rating stays healthy. Opt-outs are honoured immediately. And the team spends its time on conversations that genuinely need a human, which you can see in the split between agent-resolved and handed-off chats.

6:41 Watch me do it: window-aware messaging

Watch me do it. I open the window logic. First, the window is twenty-four hours. Can send freeform checks whether the time since the customer's last inbound message is under that. Choose message first returns nothing for opted-out contacts. If the window is open, it returns a text message with the AI reply. If not, and the contact opted in to utility messages, it returns the order update template with the order number as a parameter. Otherwise, nothing: there's no permitted way to message them. Next, the curl command for the Cloud API: the Graph API version and phone number ID come from environment variables, the token is a bearer token from an environment variable, and the body names the template, its language and the parameter. I send it to a test number and the approved template arrives. Finally, the opening message: the greeting, the assistant's name, that it's an AI, what it can help with, and type agent to reach the team.

7:52 Recap

To recap: use the Business Platform for automation, collect opt-in, reply freely only inside the twenty-four-hour window, use approved templates outside it, keep AI business-specific, disclose it's AI, and make human hand-off easy. Your next step: design a messaging agent for your business, with its scope, opening message, hand-off triggers, window logic and two templates you'd submit for approval. Next, voice agents that answer your phone.

8:21 Try this now (30 minutes)

Try this now. Design a messaging agent for your business on one page: its scope, the opening message with AI disclosure and how to reach a person, hand-off triggers, the window logic and two templates you'd submit for approval, one utility and one marketing. Then write five test conversations, including a voice note, a Roman Urdu or Arabic message and an angry complaint.