Malaysia had 35.4 million internet users at the end of 2025, 98% of the population, according to DataReportal. In the Reuters Institute’s 2026 survey of 2,076 Malaysians, WhatsApp was the most widely used platform for news, at 50%, narrowly ahead of Facebook.
For an aircon servicer, a clinic, a tuition centre or a property agent, that makes the chat the shopfront. Enquiries arrive at all hours. The ones answered late tend to go to a competitor.
An AI agent can answer them. It does so on Meta’s platform and under Meta’s rules, and two of those rules changed in 2026. This article covers customer service and enquiries. Order entry for distributors is a separate subject.
The rules Meta sets
An agent connects through the WhatsApp Business Platform, Meta’s interface for business software. Five rules shape what it can do.
The 24-hour window. When a customer messages you, a customer service window opens for 24 hours and resets with each new message from them. Inside it, you may reply freely. Outside it, you may only send message templates that Meta has approved.
Opt-in. You may contact people only if they have given you their number and agreed to hear from you. Meta also recommends clear instructions for opting out.
A way to reach a person. Automation is allowed, but Meta’s policy says a business “must also have available prompt, clear, and direct escalation paths”. It lists transfer to a human agent in the chat, a phone number, email, web support, a visit in store, or a support form.
No general-purpose chatbots. Since 15 January 2026, Meta’s terms bar AI providers from using the platform to offer an AI assistant as its “primary (rather than incidental or ancillary) functionality”. Meta told TechCrunch that this “doesn’t affect businesses that are using AI to serve customers on WhatsApp”. A clinic’s booking agent is fine. An ask-me-anything bot is not. Keep the agent on your own business.
Your chats are not training data. The same terms say a business must not let its WhatsApp data be used to build or improve AI models, with a narrow exception for a model tuned for that business’s exclusive use. Ask any vendor to confirm this in writing.
Meta last updated these terms on 23 September 2026. Check them before you build.
What it costs since 1 October 2026
Meta has charged per delivered message, not per conversation, since 1 July 2025. Messages from customers to you are free. What changed in October is the price of replying.
- Replies inside the window are charged again. Meta calls these service messages. They were free from 1 November 2024. From 1 October 2026, each business phone number gets 1,000 free a month, and Meta charges from the 1,001st. Unused messages do not carry over.
- Without a payment method, replies stop. Once the free 1,000 are used, Meta will not deliver further service messages for a number with no billing set up.
- Templates are charged by category and country. Marketing, utility and authentication each have a rate. Meta lists Malaysia as its own market and has published rates in ringgit since 1 April 2026. Rates change, so read Meta’s current rate card for exact figures.
- Meta’s own agent is priced differently. Replies generated by Meta’s Business Agent Platform are billed at US$2.00 per million tokens. Meta says a typical message uses 20,000 to 25,000 tokens, or about 4 to 5 US cents.
One thousand a month is about 33 replies a day from one number. A business sending more will pay for the rest, on top of what its AI model costs per reply.
The practical effect: a chatty agent is now an expensive one. One complete answer is better than five short ones.
What an agent does well
- Answers common questions from your own information. Opening hours, location, prices, what a service includes, how to prepare for an appointment. At 11pm and on public holidays.
- Collects the details before a person steps in. Name, address, the problem, a photo of the unit.
- Books, moves and cancels appointments against a real calendar.
- Looks things up. Is my order ready, what do I owe, when is the technician coming.
- Reads what customers send. A photo, a PDF, a screenshot of a payment.
What it should not do
- Decide refunds, exceptions or complaints. These are judgment, and a wrong answer costs more than a slow one.
- Give medical, legal or financial advice.
- Invent a price. Prices come from your system by rule. The model does not work them out.
- Chase people who never agreed to be messaged.
- Pretend to be a person. Say in the first reply that an assistant is answering and that staff are available.
Handover is the product
The difference between an agent customers tolerate and one they like is how it gives up.
When to hand over. The customer asks for a person. The agent cannot find the answer in your information. The message is a complaint. Money above a set limit is involved. Anything medical or legal.
What staff see. The whole conversation, with a short summary and the details already collected. Nobody should have to ask the customer to repeat themselves.
What the AI does next. It stops. When staff take over a chat, the agent pauses until it is told to continue.
After hours. The agent says when a person will reply, and the chat is waiting in the queue in the morning.
We are building a platform of this kind, AllSet, with Northridge Technologies of Singapore. It is in development and runs on demonstration businesses for now. Its agents answer from an indexed knowledge base and report the gaps when they cannot find an answer. The staff inbox shows when the 24-hour window has closed and offers an approved template. Most of the code is not the chat at all. It is the calendar, the customer records and the invoicing the agent acts on.
Language
Your customers write in Malay, English, Mandarin and all three at once, with short forms no textbook contains. Published tests show AI models are much weaker on this than on formal language. We set out the evidence and a way to test a model on your own messages.
Personal data
A chat history is personal data: a phone number, a name, an address, sometimes a health detail or a bank slip.
- An AI model is a third party. Sending a chat to a hosted model discloses it, often to another country. Send only what the task needs, and know the provider’s region and retention terms. We covered this in the amended PDPA for software teams.
- Keep proof of opt-in for anyone you send marketing templates to.
- Decide how long chats are kept, and delete them when that time is up.
- Tell customers. Link your privacy notice where they will see it.
A sensible first project
One number, one job: after-hours enquiries for one branch.
- Read a month of chats and list the twenty most common questions.
- Write the answers down. This is most of the work, and it is useful whether or not you ever deploy an agent.
- Let the agent answer only those. Everything else goes to a person.
- Start in draft mode. For two weeks the agent proposes and staff press send.
- Measure four things: time to first reply, share of chats handed over, wrong answers staff caught, and messages billed.
- Then let it send on its own for the question types it got right.
If the numbers are poor, you have lost two weeks and gained a written set of answers. If they are good, add bookings next.
Sources
- Digital 2026: Malaysia (DataReportal)
- Social media overtakes TV, news websites as Malaysians’ main source of news (Malay Mail, on the Reuters Institute Digital News Report 2026)
- Pricing on the WhatsApp Business Platform (Meta for Developers)
- WhatsApp Business messaging policy (Meta)
- Meta Terms for WhatsApp Business Platform, section 4.7 on AI providers (Meta)
- WhatsApp changes its terms to bar general-purpose chatbots from its platform (TechCrunch)
This article is general information and is not legal advice.