Meta's agent is billing now: what your first invoice says

If you have an AI agent answering on WhatsApp, the meter has been running for twelve days. On 1 August Meta started charging for its Meta Business Agent, and at the end of the month you'll get your first real invoice — priced in a unit you've probably never seen on a bill before: tokens. Let's turn that into money, work out which numbers you need to be able to calculate yourself, and be honest about when the right move is to do nothing at all.

A token isn't a message, and that's where the confusion starts

A token is a chunk of text — roughly a short word, sometimes half a word. When a customer writes in, your agent reads the message, pulls what it knows about your business (catalogue, opening hours, past orders), reasons about the request and writes a reply. All of that reading and writing is counted in tokens.

Meta's published rate is $2 per million tokens. A typical exchange burns 20,000 to 25,000 tokens, which Meta itself translates to roughly 4-5 cents per message. It's in their pricing documentation.

Here's the part that actually matters: you're no longer paying per message sent, you're paying per complicated conversation. Someone asking "what time do you open on Saturdays?" costs you pocket change. Someone who makes the agent search three products in the catalogue, check the calendar, offer two slots and close the booking costs considerably more — even though the chat thread looks about the same length on screen.

One detail that gets overlooked and works in your favour: that rate bundles the AI processing and the message delivery together. Build your own agent and those are two separate bills — one from your model provider, one for WhatsApp delivery. Here it's a single line. Not necessarily cheaper, but far easier to read.

The maths, with two real businesses

Round numbers help nobody. Two concrete cases, using the published rate and counting conservatively at 4-5 cents per agent reply.

A dental practice or a hair salon. Around 25 conversations a day, five and a half days a week — call it 600 a month. Most are short: confirm an appointment, ask a price, move a slot. Say three agent replies each. That's 1,800 replies, or $70 to $90 a month. Less than one afternoon of front-desk cover per week, but it's no longer zero.

An ecommerce shop during a campaign week. 150 conversations a day for seven days, with longer queries — where's my order, size swap, return. Easily five or six replies per conversation. That's about 5,500 replies in a week, on the order of $250 for those seven days alone. At that level the number deserves a look before the campaign goes out, not after.

To be clear: these are estimates built on the average consumption Meta publishes, not on your business. Your real usage depends on how big the knowledge base you gave the agent is and how much it waffles. Which is why the next section is the one that counts.

Four numbers you should be able to work out

If all you look at is the invoice total, you'll learn nothing. These four tell you something:

1. Tokens per conversation. Total tokens divided by conversations handled. If it lands well above 25,000, it isn't that your customers are difficult — it's that your agent is reading too much context or replying in paragraphs. You fix that by pruning what you fed it, not by switching provider.
2. Cost per resolved conversation, not per conversation handled. If the agent replies and the customer phones you anyway, you paid twice for the same job.
3. Share that ends up with a human. This is your quality thermometer. 20-30% escalation is healthy for a well-built agent; 60% means it's missing information, not intelligence.
4. Cost per conversation against what that conversation is worth. Five cents to handle someone booking a €200 treatment is a bargain. Five cents to tell someone whether you have parking is money down the drain — and the fix there is probably putting your address and hours somewhere visible.

The seven-week window almost nobody is using

There's a gap here worth spotting. Since 1 August, replies from Meta's agent are billed — but replies typed by a human on your team inside the 24-hour service window stay free until 1 October. After that date those get charged too, per the change schedule Meta published alongside the new rate.

So between now and October you have a lab with asymmetric prices: you can see exactly what it costs when the machine answers and compare it against what it will cost when a person answers and that's billable too. It's the only window this year where that comparison can be made with data rather than gut feel. From October, however you reply, everything counts.

You don't need to build anything for this. For two weeks, log how many conversations the agent carries end to end and how many land with your team. That's half the decision made.

When Meta's agent isn't your best option

Now the uncomfortable part, because price per message isn't the only criterion.

You don't control the model. Meta hasn't publicly said which model sits under the Business Agent. If they update it tomorrow and the tone or the shape of the answers shifts, you'll hear about it from your customers. For a salon that's fine. For a firm answering questions with legal consequences, it really isn't.

It only lives inside Meta. The agent works on WhatsApp, Instagram and Messenger, and that's the end of it. It doesn't touch your email, your phone line, your website chat or SMS. If your support is spread across channels you'll need something else anyway — and then the question stops being "what does Meta's agent cost?" and becomes "how many separate systems do I want to maintain?"

And the most common case: this may simply not apply to you. If you send ten or fifteen messages a day, everything above adds up to about three euros a month. Reorganising how you handle customers over that is a waste of an afternoon. Bookmark this, set yourself a reminder for September, and get on with your day.

Some context so nobody gets carried away: Gartner projects that more than 40% of enterprise agentic AI projects will be abandoned before the end of 2027, mostly over unclear value and escalating costs. The difference between being in that 40% and not usually comes down to precisely this: having measured from month one.

What I'd do this week

Three things, half an hour total. Ask your provider — or find it in your dashboard — for usage from 1 to 12 August and multiply by 2.5 to project the full month. Note how many conversations closed without anyone stepping in. Then decide, with that number in front of you, whether you want the machine or your team doing more of the answering come October.

If you're at exactly that point of deciding who replies and under what rules, our receptionist agent brings WhatsApp, your website and email into one place, handles the repetitive stuff and hands you what needs judgement. You approve, it executes — and you draw the line.