For the month of July, Meta's Business Agent — the AI that answers customers inside WhatsApp, Instagram, and Messenger — worked for free. On August 1, the meter switched on.
The rate is $2.00 per million tokens, which sounds abstract until you translate it. A typical customer interaction burns 20,000 to 25,000 tokens of reading, reasoning, and replying, which works out to roughly 4 to 5 cents per message. A business whose agent handles 5,000 conversations in a month pays about $200 to $250. And that one charge covers both the AI and the message delivery — no separate WhatsApp bill, no separate model bill.
If you switched the agent on when it went global in June and never looked at it again, this is the month it quietly becomes a line item. The first invoices land at the end of the billing period. Better to know your number before the envelope does.
October is the real deadline
The August change only touches AI-handled conversations. The bigger shift comes on October 1, and it touches every business on WhatsApp, agent or not.
Since late 2024, the free-form replies you or your staff send inside the 24-hour customer service window have been free. That ends October 1. Utility templates sent inside that window get billed again too. The practical effect: from October on, there is no free way to answer a customer inside an open conversation. AI reply, human reply, template — every response carries a cost.
The question is no longer "should we pay for AI replies?" It's "every reply now costs money — which kind earns more back?"
You have a two-month measurement window
Between now and October 1, there's a strange and useful gap: AI replies are billable, but human replies inside the service window are still free. That makes these two months the cheapest experiment you'll ever run on your own customer conversations.
Here's what to actually measure. Volume: how many conversations come in per month, and how many the agent can resolve without a human. Cost per conversation: at 4–5 cents a message, a five-message exchange runs about a quarter. Value per conversation: what a booked appointment, a recovered quote, or a completed order is worth to you. If a 9-cent exchange books an $85 job, the cost is a rounding error. If your agent burns tokens answering questions that never lead anywhere, that's worth knowing before the human replies start billing too.
For context, Meta's nickel is on the cheap end of the market. Intercom's Fin charges about 99 cents per resolved conversation. Salesforce's Agentforce runs roughly $2 per conversation. The trade-off is coverage: Meta's agent only works on Meta's surfaces. Your email, your website chat, your phone line — those still need their own answer.
The honest caveat
None of this math works if the agent answers badly. An agent trained on a thin, outdated knowledge base gets things wrong — confidently — and that costs more than tokens. Gartner projects that over 40% of agentic AI projects across the industry will be abandoned by the end of 2027, mostly for unclear business value. The difference between the projects that survive and the ones that don't is rarely the model. It's whether anyone fed it good information, set real guardrails, and measured what it produced.
The businesses that track cost per conversation and value per conversation starting now will walk into October with data. The ones that don't will be guessing at their own numbers when the bill arrives. Guessing was fine when replies were free. That era has about eight weeks left.