What an AI agent costs a business, and the four bills nobody adds up

“What does an AI agent cost?” is one of those questions with no single answer, but four. And the problem is not that there are four: it is that almost no quote separates them, so the buyer ends up comparing one number against another number that measures different things.

Here they are, in the order they show up on the invoice.

1. The platform

This is the monthly fee: the workspace, the channels, the CRM, the users, the place where everything lives. It is the easy part to compare because it is a published number. At Meteor plans start at US$99 a month and are sized on two things: how many people on your team will use it and how many contacts you carry.

Two warnings about this bill, because it is where most bad comparisons happen:

• Look at the limit that will bind first. It is usually not the one you focus on. A team of three with a base of thirty thousand contacts needs a bigger plan than a team of fifteen with two thousand.
• Check whether the annual price is a discount or free months. They are not the same. We publish that a year equals ten months, which is another way of saying two are free, and that arithmetic can be checked.

2. Model usage

This is where comparisons break.

An agent that executes calls tools, reads what your systems return, reasons about it and calls again. That consumes, and what it consumes depends on how much work it does, not on how many seats you have. A company with two people and a thousand daily conversations consumes far more than one with twenty people and fifty.

Which is why this cost is almost never fixed, and when someone sells it to you as fixed it is worth asking what happens when you exceed it. The two possible answers are very different: either service stops, or an invoice arrives at the end of the month.

At Meteor this is called Energy and it is prepaid. It runs out and executions stop. That is less comfortable than an open invoice, and it is what makes an end-of-month surprise impossible.

What you will not find anywhere, and should distrust when offered, is a precise estimate of your consumption before anything is switched on. It depends on how many conversations arrive, how complex they are, and how many tools get called to resolve each one. You find out by measuring.

3. What the channel charges

This is the bill that surprises people most, because it does not come from the platform. It comes from Meta.

And its shape changed. Until mid-2025 WhatsApp billed per conversation: you paid once and inside that 24-hour window you could send several templates of the same category at no extra cost. Since 1 July 2025 billing is per delivered template message. Three marketing templates to the same number in one day are three charges, not one.

The practical consequence reshapes campaign design:

Who starts the conversation	What gets sent	Who bills it
The customer writes to you	Free-text replies from the Met	Nobody bills the message
You write first	An approved template	Meta, on every send

In other words: answering is cheap and chasing costs money. The day-before reminder, the overdue invoice nudge and the abandoned cart recovery pay for a template every time. That is not an argument against doing them, it is an argument for budgeting and measuring them separately: a collections campaign pays for itself if it recovers more than it costs to send, and that arithmetic is doable.

The exact rate depends on the country of the receiving number and the template category, and Meta publishes it. We do not copy it here because it changes.

4. The rollout, and everything after it

The fourth bill has two halves, and the second is the one almost nobody quotes.

The first is setup: connecting systems, writing the Met’s instructions, registering templates, testing. It can be estimated reasonably and depends on how many systems are involved and how tidy they are.

The second is upkeep. Reading real conversations, finding the answers that sounded right and were wrong, adjusting. That work does not disappear: it gets small. In the first weeks it can be a few hours a week; after that it drops a lot, but it never reaches zero, because your catalog changes, your prices change and your customers ask new things.

A proposal that leaves this half out is not cheaper. It is the same proposal with a part unwritten.

How to read a quote with this in hand

When someone hands you a number, split it into the four bills and ask for whatever is missing:

1. How much is fixed per month, and which limits come with it?
2. How is AI usage billed, and what exactly happens when it is exceeded?
3. Is the channel cost included or separate? If they say included, ask up to how many sends.
4. What does the rollout include, and what happens the month after it goes live?

If all four have answers, you can compare. If one does not, you are not comparing prices: you are comparing what each vendor decided to tell you.

A shortcut for the first bill

If all you want is to know which plan covers you and what it costs, without writing to anyone, the calculator on this site answers with two inputs: how many people on your team will use it and how many contacts you have. And when the operation exceeds the largest published plan it returns “let us talk” instead of inventing a price, which is the honest answer at that size.