If your business handles customers on WhatsApp, the bill changed twice this half-year. On August 1, Meta began charging per token for replies from the Meta Business Agent. On October 1, it also began charging service messages and utility messages sent in reply to the user inside the 24-hour window. The window still exists; what ended was serving customers for free inside it.
This text is about the WhatsApp Business Platform, the API used by anyone who automates customer service, integrates a catalog, or sends a notification. If you use only the WhatsApp Business app on the phone, the charging model is different; the decision part — what to automate and what to keep human — still applies, but check the amounts in your dashboard.
This post is an analysis of public documentation and of my experience with WhatsApp flows, not a test of the Meta Business Agent. The numbers are the ones Meta publishes, with a date. Where the math depends on a hypothesis, I say which one.
Update note (October 30, 2026): the previous version said that ZapOrca sends the quote as a PDF over WhatsApp. Checked in the code, what goes out on WhatsApp is the proposal text, with the Pix copy-and-paste code; the PDF comes from printing in the browser, and the provider attaches it by hand when they want to. The ZapOrca technical case describes the flow.
What changed, in order
The first change was one of category. On July 1, the Meta Business Agent became its own type of non-template message, alongside service messages. Meta reports that more than 1 million businesses were already using a Business Agent on WhatsApp and on Messenger in the June presentation. Source: Meta, June 3, 2026.
The second was one of price. Since August 1, Business Agent messages are charged per token: US$2.00 per 1 million tokens, in a single charge that includes the agent’s processing and message delivery. Meta estimates 20,000 to 25,000 tokens per message, which comes to about 4 to 5 cents of a dollar per message. A simple question, with four replies, used about 80,000 tokens in Meta’s own example (16 to 20 cents). A complex interaction, with ten replies, reached 250,000 tokens (40 to 50 cents). Source: Meta for Developers, updated in 2026.
The third was the one that most surprised people who already used the API. Since October 1, service messages — free-form replies inside the 24-hour window, written by a person or by a third-party agent — began to be charged per message. They had not been charged since November 1, 2024. Utility messages sent in reply to the user inside the window also began to be charged; they had not been since July 1, 2025. Source: the same pricing page.
There is an operational detail that has already passed: Meta warned that anyone who did not have a payment method on file by September 30 would stop receiving delivery of service messages starting October 1. If your Billing Hub has no card or credit line, it is worth checking before you open customer service.
And there is an exception that still holds: conversations that start from an ad with a click to WhatsApp, inside the free entry-point window, still have no delivery charge. But Business Agent messages are still charged per token. In other words: the ad can open the door for free; the AI reply cannot.
The bill Meta publishes
Meta itself publishes a comparison for Brazil. The table below is Meta’s, with an important caveat: it is not an apples-to-apples comparison. The cost of a service message includes what you pay your AI provider and your operation; the cost of the Business Agent includes the token and delivery. Meta says this explicitly.
| Scenario (1,000 reply messages in Brazil) | Estimated cost |
|---|---|
| Service with low-complexity AI | ~US$27 |
| Service with high-complexity AI | ~US$97 |
| Meta Business Agent, high complexity | ~US$40 to 50 |
The delivery fee Meta uses for Brazil is 0.68 cent of a dollar per message (on the date of the table, July 1, 2026). The rest of the service cost is the AI model and operations. The Business Agent replaces all of that with a single per-token charge, which includes delivery.
In everyday terms: a conversation does not cost one message. It costs the number of replies it generates. A request for a time slot can end in two messages. An assembly question can turn into ten. The cost per conversation is messages × cost per message, and that is what has to enter the bill — not the unit price that appears in the platform’s announcement.
The plain arithmetic
The formula is simple, but it has to include what almost nobody includes:
Cost per conversation = (messages × cost per message) + cost of exceptions + cost of upkeep.
The first term is what Meta charges. The second is what comes back to the human: a complaint, a refund, a negotiation, a question outside the catalog. The third is what does not appear on the invoice: an updated catalog, a revised FAQ, monitoring, testing, correcting a wrong answer, and the time of the person who looks after that.
An illustrative example, with hypothetical amounts to show the method: suppose an agent that uses 8 messages per conversation at US$0.05, or US$0.40 per conversation. Suppose an attendant who costs US$5 per hour (example) and resolves 6 conversations per hour: US$0.83 per conversation. If 20% of conversations need a human, the average cost comes to 0.8 × 0.40 + 0.2 × 0.83 = US$0.49. The savings exist, but they are US$0.34 per conversation — and they disappear if the handoff rate rises or if catalog upkeep consumes a few hours a week. Across 1,000 conversations, that is US$340 before those costs.
If your attendant costs R$20 per hour and resolves 6 conversations per hour, each conversation costs about R$3.33 in time — not counting labor charges, training, and turnover. Convert Meta’s dollars at the day’s exchange rate before comparing. The conclusion is rarely “the agent is always cheaper”; it is “the agent is cheaper when volume is high, ambiguity is low, and the exception is rare.”
What an agent does well
The Business Agent was designed to answer questions about the business, recommend products from the catalog, schedule, qualify leads, close a sale, and hand off to a person when needed. Source: Meta, June 3, 2026. In practice, it shines at what is repetitive and verifiable:
- hours, address, payment methods, and delivery time;
- “do you have it in another color?”, “what size is it?”, “how much does it cost?” — as long as the catalog is correct;
- order and appointment status;
- a first reply outside business hours;
- qualifying whoever arrives: what the person wants, for when, and in which region.
The criterion I use is reversibility. If the wrong answer costs a cheap correction — “the hours are different,” “the color is this one” — the agent can take the risk. If the wrong answer costs money, a right, or reputation — a refund, a cancellation, a warranty period, sensitive data, legal, medical, or financial guidance — the answer is human.
The second criterion is ambiguity. If the question depends on context that is not in the catalog or in the history, the agent will make it up. And invention in customer service has a name: a promise the company did not make.
The handoff is part of the bill
“Hand off to a human” is not a courtesy button; it is a design decision. The agent needs to know when to stop, how to stop, and what to hand to the person who takes over. Without that, the customer repeats the story three times and the cost rises on both sides.
In practice:
- Say that it is an assistant. Transparency is not a legal detail; it is what avoids the feeling of talking to a fake person.
- Define clear triggers: “talk to a person,” a complaint, a refund, a cancellation, a negotiation, sensitive data, explicit frustration, a repeated question with no answer.
- Pass the context. The human should receive the history and what the agent already tried. A customer who repeats everything is a customer who gives up.
- Have business hours and a fallback. Outside business hours, the agent can record the request and promise a return; it cannot promise what nobody will fulfill.
- Count the handoff message. Since October 1, the human reply inside the window is also a charged service message. The handoff is not free.
Checklist before you turn it on
- A payment method in the Billing Hub. Without one, delivery of service messages can stop. Check before you publicize the number.
- A correct catalog and FAQ. The agent learns from the page, the catalog, old conversations, and the site. If the source is wrong, the wrong answer is only a matter of time.
- Written limits. What it does not answer, what it does not promise, and what it forwards. That is more useful than any pretty prompt.
- Test with an allowlist. Meta lets you test with a restricted audience before opening to everyone; opening to everyone requires the payment method. Use that stage.
- Measurement. Meta itself exposes the Business Agent Usage Insights API, with billable messages, tokens, and charges. Also track the containment rate (how many conversations end without a human), the handoff rate, and the cost per resolved conversation.
- Privacy and the LGPD. The business is responsible for what it collects and what it sends to the agent. Say what it collects, what it uses it for, and how the customer talks to you. Do not send sensitive data to the agent, and do not use the conversation for a purpose the customer did not authorize.
- A weekly sample. Read real conversations. An aggregate metric hides a wrong answer with a good score.
- Turning it off. Know how to pause the agent and who answers when it stops. An agent with no plan B is a channel that is down.
What I saw in Lika and in ZapOrca
Both projects are mine, and both use WhatsApp as the channel where the deal closes — not as a Business Agent. The lesson I take from them is not about tokens; it is about what the customer asks before they arrive.
At Lika Toalhas e Buquê, the site publishes prices, measurements, and colors. That reduces the first question and makes the contact arrive more qualified. WhatsApp closes the conversation; it does not need to guess the catalog. The automation that works best there is not a robot answering: it is the right information being available beforehand.
In ZapOrca, the quote and the Pix code for the deposit are built on the phone, and the proposal text, with the Pix copy-and-paste code, goes out through the provider’s own WhatsApp. The PDF, when they want to send it, is the browser printout, attached by hand. The artifact is automatic; the negotiation stays human. The customer receives something professional, and the provider does not lose time building a proposal by hand. Again: automate the artifact, not the relationship.
If I were to put an AI agent in these two businesses today, I would start with what is repetitive and verifiable: hours, availability, deadline, order status, a first screening. I would not start with a wedding price, a contract exception, or a discount negotiation. The cost of an error there is higher than the cost of the token.
What remains
The 24-hour window still exists. What changed is that it stopped being a free space. That holds for Meta’s agent, for your own agent, and for the person who replies on the phone. The decision stopped being “AI or human” and became “which mix resolves the conversation at the lowest total cost, including the error.”
The best use of the agent is the one that removes the trivial question and frees the human for the exception. The worst is the one that tries to save money by hiding the price, pushing the sale, or pretending it understood. The math works out in the first case; in the second, it arrives later, as a complaint, a refund, or a lost customer.
If you cannot measure the cost per resolved conversation, you are not ready to turn the agent on. Start with the spreadsheet, not the chatbot.
Sources
- Meta for Developers, Upcoming pricing updates for Meta Business Agent, service and utility messages, with the dates of August 1, 2026 and October 1, 2026, the cost table for Brazil, and the service × Business Agent comparison: developers.facebook.com, verified on 2026-10-23.
- Meta, Conversations 2026: Introducing Meta Business Agent, June 3, 2026, with the figure of more than 1 million businesses and the agent’s functions: facebook.com/business/news, verified on 2026-10-23.
- The labor amounts in the example are hypothetical and serve to show the method; they are not a wage survey. Converting dollars to reais depends on the day’s exchange rate and the payment method used in the Billing Hub.
