WhatsApp automation with AI: what can be automated, what cannot, and the October billing change

What you get from this article

• Legitimate automation requires the official API: a solution that drives the phone from outside gets the number banned.
• Since July 2025 billing is per message, with prices by category and by market.
• On 1 October 2026 service messages and utility templates stop being free.
• The 24-hour service window opens when the customer contacts you, and restarts with each new message.
• AI should solve the repetitive and hand the difficult to a person, not push the customer around.
• Billing and customer data together require a lawful basis under the data protection law and a record of consent.

First, the part that decides whether the project exists

Before talking about AI, one distinction that solves half the problems. There are two ways to automate WhatsApp, and only one is legitimate.

The official Meta API is the correct route. It requires a verified business account, a dedicated number that leaves the ordinary app, approval of message templates, and it is contracted through an authorized provider or directly with Meta. It takes more work to start and it is the only one that carries no risk.

The other is any solution that drives the app from outside, automating WhatsApp Web or the phone. That violates the platform terms of use, and the consequence is not a warning: it is the number banned, with the conversation history alongside it. Companies find this out the worst way, losing their main sales channel overnight.

A commercial warning is worth making: when a supplier offers cheap WhatsApp automation, with no business verification and no message template approval, it is almost always the second option. The initial saving is real and the risk is entirely yours.

The asset is the number, not the tool that automates it. Risking the asset to save on the tool inverts the arithmetic.

How billing works today

Since 1 July 2025, Meta charges per delivered message, no longer per conversation. The price depends on two things: the message category and the recipient country.

There are three categories, and the price difference between them is large. Marketing is the most expensive: promotions, news, win-back messages, anything with promotional intent. Utility covers order confirmations, delivery notices, appointment reminders, status updates. Authentication is verification codes.

The border between marketing and utility is where almost every company gets tangled, and Meta yardstick is promotional intent. Telling somebody their order is out for delivery is utility. Telling them the order is out for delivery and suggesting they use a coupon is marketing, and costs several times more. A coupon added by somebody on the sales team can multiply the cost of an entire campaign without anyone noticing.

There is also the service window: when the customer sends a message or calls you, a 24-hour period opens, and that counter restarts with each new message from them. Inside that window you reply in free text, with no approved template needed.

The 1 October 2026 change

This is the most important piece of information in this article for anyone already operating, and it has a date in Meta official documentation.

Today, two things are free: service messages, which are the free-text replies sent inside the 24-hour window, and utility templates sent inside that same window. It is on that free tier that most automated service projects built their numbers.

From 1 October 2026, both start being charged per message, at rates following the utility and authentication tariffs of each market. In other words: a service conversation, which today only costs for the first message when you start it, begins to carry a cost per exchange.

And there is an earlier change already in force: since 1 August 2026, messages from Meta own business agent are charged per token, at two dollars per million tokens, which is a billing model unlike anything the channel had before.

The practical consequence is direct: a long conversation now costs in proportion to its length. A bot that answers in five short messages what would fit in two starts costing more than double. Redoing the arithmetic with the new pricing, before October, is an afternoon of work that avoids a surprise in November.

Where AI genuinely pays off in this channel

With a cost per message in play, AI has to do two things: resolve more in fewer exchanges and stop a person spending time on the repetitive.

Qualification before the human. A clinic receiving fifty messages a day saying "how much is it?" can have the AI find out which procedure, whether it is insurance or private, and how urgent it is, handing the agent a completed record instead of a conversation starting from zero. The gain is bigger in team time than in the message itself.

Answering from in-house material. Opening hours, address, payment methods, lead time, returns policy, availability. These are questions whose answer already exists in writing and nobody can find. It is the classic RAG scenario: the AI answers from your documents instead of inventing.

Triage and routing. Reading the message, classifying the subject, checking whether that customer is already on file, and directing it to the right person with the context attached. That removes the worst friction in WhatsApp support, which is the customer repeating the same story to three people.

And one practice that started costing money with the October change: answering completely in the first message. That bot pattern of sending "hello", then "how can I help?", then "certainly, one moment" was irritating when it was free. Now it is irritating and expensive.

This text covers only WhatsApp. The other fronts, from content to CRM, are listed in AI for business.

• Qualify before handing over to a person, delivering a completed record.
• Answer questions whose answer already exists in a company document.
• Classify and route with the context attached, so the customer does not repeat the story.
• Confirm and remind about appointments, which is where utility pays for itself.
• Summarize the conversation for the human agent before they take over.

What should not be automated

Being specific here avoids the most expensive damage, which is to the relationship and not to the invoice.

An angry customer. The moment a message carries a complaint, urgency or frustration, the automation needs to step aside and call somebody. A bot insisting on collecting information from an angry person turns a small problem into a bad public review.

Non-standard pricing and negotiation. If the amount depends on analysis, on a discount or on special terms, a person decides. AI can quote the price list, it cannot negotiate.

Anything with legal or financial effect. Confirming a cancellation, promising a deadline, authorizing a refund, accepting contractual terms. The house rule applies in full here: amounts are not calculated or promised by a language model.

Pretending to be human. This deserves its own paragraph because the temptation is strong. You do not have to announce "I am a bot" in the first line, and you also cannot lie when the person asks. The honest yardstick is simple: the service can be natural and pleasant, and when the customer asks directly, the answer is the truth. Besides being right, it avoids the feeling of being deceived when the person finds out later.

The data part, which is not optional

A WhatsApp conversation is personal data, and frequently sensitive data. Name, phone, what the person bought, the health problem they mentioned, the delivery address. All of that is under Brazilian data protection law, and the responsibility sits with the company operating the channel.

Three practical points. Lawful basis: a marketing message requires consent; serving somebody who contacted you has a different basis. Mixing the two in the same list is a common and avoidable problem. Record: keep when and how each person agreed to receive, because when it is time to prove it, whoever did not record has nothing to show.

Where the conversation ends up: if the customer text is sent to an AI provider, check in the contract what is done with that content and whether there is a commitment not to use it in training. That matters especially when the conversation touches health, financial situation or any sensitive category. The detail is in LGPD and AI.

And the basic thing many companies forget: an easy way out. Every marketing communication needs a clear path for the person to stop receiving, and that request has to be honoured quickly, not at the next list cleanup.

How to build it without getting tangled

The order that avoids waste starts with measuring before automating, and almost nobody does that.

One week of reading. Before choosing any tool, read the messages that came in over the last week and classify them by hand. You will discover that a handful of questions accounts for most of the volume, and that is where the project should start. Automating what appears once a month is effort with no return.

Start with the answer, not the bot. Write the correct answer to the five most common questions, the way your best agent would answer. That material is what feeds the AI, and its quality sets the ceiling on the result. Without it, you automate a bad answer, faster.

Launch with an easy handover to a human. In the first weeks, the threshold for passing to a person should be generous. It is better to pass too often and tighten up with data than to save on support and lose a customer.

Redo the arithmetic with October pricing. Before scaling, project the cost under the new billing: service messages and utility templates charged per message. If the project only works on the old pricing, it does not work.

If you want that design done with the costs projected and the human handover defined before it becomes code, that is what ROO3 AI consulting delivers in the first stage.

Frequently asked questions

Can I automate my ordinary WhatsApp?

Not safely. Legitimate automation requires the official Meta API, with a verified business account and a dedicated number. Solutions that drive the app or WhatsApp Web from outside violate the terms of use, and the typical consequence is the number being banned, along with the conversation history.

What changes in WhatsApp billing in October 2026?

From 1 October 2026, service messages, which are the free-text replies inside the 24-hour window, and utility templates sent in that window start being charged per message. Today those sends are free, and the tariffs will follow the utility and authentication rates of each market.

What is the difference between a marketing and a utility message?

The yardstick is promotional intent. Confirming an order, giving notice about delivery or reminding about an appointment is utility. The same message with a coupon or an invitation to buy again becomes marketing, which costs several times more. A discount added without thinking can multiply the cost of an entire campaign.

What is the 24-hour service window?

It is the period that opens when the customer sends a message or calls your company, and that restarts with each new contact from them. Inside it you reply in free text, with no approved template needed. From October 2026 those replies start being charged per message.

Can the AI pretend to be a person in customer service?

It must not lie. The service can be natural and pleasant without announcing that it is automated in every message, but when the customer asks directly, the answer has to be truthful. Besides being right, it avoids the feeling of deception when the person finds out later.

Do I need consent to send a WhatsApp message?

For marketing communication, yes, and you need to be able to prove when and how the person agreed. Serving somebody who contacted your company rests on a different lawful basis. Mixing the two situations in the same list is a common and avoidable mistake, and responsibility for the processing sits with the company.

Sources

• Meta for Developers: pricing updates for service and utility messages
• Meta for Developers: WhatsApp Business Platform pricing
• Meta for Developers: service messages and the service window

Rodrigo Fávaro

Founder of ROO3, a marketing and technology agency in São José do Rio Preto, Brazil. Builds AI products running in production (Tobia, gerar.app, Pense Mercado) and maintains the AI Benchmark, a public ranking of AI models. See ROO3 AI consulting.

X @rodmf LinkedIn rodrigofavaro