From 1 October 2026, Meta starts charging per message for the ordinary replies a business sends back to a customer on WhatsApp, the kind that used to be free once someone messaged you first. I read the announcement expecting to feel worried on behalf of clients. I did not. Most of the WhatsApp systems I have built were never going to run up much cost under this pricing, and the ones that would be expensive were already the wrong kind of automation.
What Meta is actually changing
Right now, if a customer messages you first, you get a 24 hour window to reply for free, no matter how many messages you send back. According to Meta's own documentation, that stops being free on 1 October 2026. Every reply inside that window, what Meta calls a service message, gets charged per message from that date. There is also a new category, Meta Business Agent, which Meta charges based on how much the reply actually has to process and generate, meaning a short answer costs less than a long, multi-turn one.
So the free ride on replies is ending. Utility templates sent in reply inside that window, such as an order confirmation, also become chargeable on the same date. Marketing and authentication templates were never free to begin with. The change is specifically about the back-and-forth after the customer writes in.
Why this does not change what I would build
The first WhatsApp automation I built replaced hospital staff who were manually replying to and chasing every enquiry that came in. In the first week of real use, patients replied in ways the flow had never been designed for. My fix was not a cleverer flow that tried to handle every case. It was a fallback to a human at any point the bot was not confident. That one change dropped manual follow-up effort by 80 percent with zero missed enquiries, and it also means the bot was never sending endless speculative replies trying to cover every branch. It asks, it answers what it can, and it hands off. Fewer messages, not more.
On a separate build, I put two or three quick questions in front of any human touch: what the person needs, when, and where. Staff only look at the enquiries that answered. That filter was designed to save staff time, but it has the side effect of keeping message volume down too, because you are not running a long automated conversation with every browser who pings you at midnight. You are asking three short questions and routing. Under the new pricing, a system built to minimise unnecessary back-and-forth costs less to run, not because I designed it for that, but because a good automation and a cheap one turn out to be the same automation.
Where this actually bites
A client once asked me to send an offer to everyone on their contact list over WhatsApp in one go. I talked them out of it, because messaging people who never asked to hear from you gets reported, and enough reports gets your number blocked, taking down every message the business sends, not just the offers. We used opt-in, targeted messages instead. That decision was about keeping the number alive. Under the new charging, it also happens to be the cheaper choice, because you are not generating a flood of unsolicited outbound that then triggers a wave of confused inbound replies, each one now billable.
The businesses that will feel this pricing change are the ones running WhatsApp like a call centre with no filter: every enquiry gets a full multi-turn AI conversation regardless of whether it was worth having. A per-message charge sounds trivial until you multiply it by a bot that is happy to go back and forth many times with someone who was never going to buy anything. Meta's own worked example in the documentation shows a single customer interaction going from three charges to five once service and utility replies stop being free. On their own those are small numbers. It is the multiplier across volume that matters.
Meta's documentation contrasts a short, simple question that gets resolved in a handful of messages with a longer troubleshooting conversation that runs to many more messages back and forth. A bot that answers clearly and hands off costs less to run than one that keeps chatting.
What I would check in an existing setup
- How many replies does the bot send per conversation before it resolves the enquiry or hands off to a person
- Is the bot having full conversations with people who never asked a real question, versus people who opted in and want a reply
- Does every uncertain reply fall back to a human, or does the bot keep guessing and generating more messages trying to get it right
- Are templates written as plain utility messages (reminders, confirmations) rather than promotional ones that get rejected and resent
- Is broadcast volume opt-in and targeted, or a blast to the whole list that risks getting the number reported
None of this is really new advice. I once rewrote a rejected template as a plain utility message because it read like an offer instead of something the customer had asked for, and it got approved on the second try. The lesson was that the messages worth automating are the boring, useful ones, not the exciting promotional ones. Meta pricing service messages per message from October just puts a cost on something that was already true: a WhatsApp automation that talks too much, to too many people, for too little reason, was never a good automation. It is just about to become a visibly costly one.
