In October 2026 Meta changed how it charges for the WhatsApp Business API. Service messages, the replies a business sends inside the customer service window after a customer writes in, now have a cost. Incoming messages are still free. Marketing, utility and authentication templates were already charged per message and still are.
If a team of people answers your WhatsApp, the change shows up on the invoice. If you are about to deploy an AI agent, it changes how you design it, because an agent that answers in five messages costs more than one that answers in one. With more than 2.3 million messages handled in Recaudo, this is not something we can watch from a distance.
I will not quote prices here. They vary by country and message type, and the only table that counts is Meta’s. What I can share is how to run the numbers before you decide.
Start by counting what already happens
Before talking about the agent, pull three numbers from a normal month of your current operation:
- How many conversations your customers open.
- How many messages your business sends per conversation, on average.
- How many templates you send on your own: payment reminders, appointment confirmations, shipping notices.
If you use an inbox like Recaudo or any other API provider, this data is already there. If your team answers from the WhatsApp Business app on a phone, the change does not affect you, since it applies only to the API, but you also have no way to measure.
With those numbers and Meta’s table you know what today’s operation would cost under the new rules. That is your baseline. Without it, any comparison with an agent is a guess.
A chatty agent gets expensive
Language models tend to split their answers: a greeting, then the information, then a closing question. In a chat it feels natural. With per-message charges for service messages, that is three charges where one was enough.
When we design an agent we check three things about how it answers:
- One complete answer per turn. If the customer asks how much they owe, the agent replies with the balance, the due date and the payment link in the same message.
- No filler messages. “Let me check” followed by the real answer is one message too many. If the lookup takes a while, it is better for the reply to take a while and arrive complete.
- Follow-ups as templates, with a reason. If the customer needs a message the next day, it is a utility template with a concrete reason, not “are you still there?”.
None of these rules is specific to agents. A human rep who sends six short messages also costs more. The difference is that you can fix the agent’s habit once, in its instructions, and it stays fixed.
Compare against a person, not against zero
An easy mistake in these calculations is comparing the agent’s cost against nothing. The right comparison is against what it costs for a person to handle that same conversation: their time, the hours when nobody is around, the conversations left unanswered over the weekend.
And there are conversations where the agent should not even try. A serious complaint, a payment negotiation or a customer who has already written three times, annoyed, goes to a person from the first message. What counts there is keeping the customer.
What I would measure in the first month
If you have already decided to deploy the agent, these are the figures I would review every week during the first month:
- Messages sent per resolved conversation, compared with your baseline.
- Share of conversations the agent hands to a person, and at what point.
- Templates sent that nobody answered. If that is most of them, you are sending too many.
That lets you adjust the instructions before the cost piles up. If you want to see how we build these agents, our AI agents page has the details.
Have you already received your first invoice under the new rules? I would like to know how much it moved and in which message type.