Article

What does an AI agent cost per month for a business of ten people?

Ten seats run around $200 a month, but three pricing models decide what an AI agent really costs a team of ten.

BombosBombos team··10 min read

By Piet Baudoin · September 2026

Ten seats of Claude Team cost $200 a month with annual billing. That number says little, because per seat you hit a ceiling, per action your credit runs out, and if you build it yourself there is no ceiling and the bill keeps running. I work through the numbers below with the prices the makers publish themselves. For where AI agents stand today, see the overview of the agentic economy.

What three pricing models will you be offered?

The first is per seat. You pay per employee per month, and your use sits behind a ceiling of turns or sessions. Claude and ChatGPT work this way.

The second is per action. You buy credits in advance, and every step the agent takes draws down that balance. Microsoft and Salesforce work this way.

The third is per piece of text. At every step, an agent rereads everything that was said before, because otherwise it loses track of where it was. That rereading is billed per piece, and each piece is called a token.

With the first, you know what you pay but not what you get. With the second, you know what an action costs, not how many actions your work needs. With the third, you know neither, because nothing there holds you back with a ceiling. So the question is not what a seat costs, but what it costs to get the work done.

Pricing model What you pay for The ceiling Going over
Claude Team, per seat $20 per employee per month with annual billing, $25 with monthly billing Per session, 1.25 times the usage of Pro Not stated on the page
ChatGPT Plus, per seat $20 a month 5 to 45 turns of the heaviest model per five hours The task stops
Microsoft, per action $200 per 25,000 credits, 5 per action Your prepaid credits plus a quarter The agent shuts off
Salesforce, per action $500 per 100,000 credits, 20 per action Your credit balance You buy more
Zapier, per task From about $33 a month with annual billing, 1,500 activities 75 tasks per run The run pauses and asks you to take a look
Building it yourself, per token Every piece of text in and out None The bill keeps running

What do ten seats cost, and what do you get for that same money on the meter?

The math is short. These are worked examples with today’s prices, in dollars, and I have not checked what they come to with VAT.

Ten seats of Claude Team Standard cost 10 times $20: $200 a month with annual billing, $250 with monthly billing. On the pricing page I read, OpenAI lists only Plus, $20 a month, with 5 to 45 turns of the heaviest model per five hours. I have not checked the price of ChatGPT Business, so I leave it out of the math.

Now compare that to the metered pricing. At Microsoft, $200 buys you 25,000 credits, and an action costs 5. That is 5,000 actions a month: over twenty-one working days and ten people, about 24 per person per day.

Bringing in one invoice, finding the matching data and preparing the booking is already a handful of actions. So those 24 are a couple of tasks a day, not a day’s work. At Salesforce, that same $200 buys you 2,000 actions, because there, one action costs 20 credits.

What happens when the credit runs out?

Then the work stops, rarely at a convenient moment. Microsoft writes it this way in its documentation from 14 September 2026: “Enforcement is triggered when a tenant reaches 125% of their prepaid capacity.” You may go a quarter over your credit, and after that Microsoft shuts the agent off.

At ChatGPT it ends the same way, because Work and Codex share one credit balance. In a report from 10 September 2026, a running Work task stalled at the usage limit. Half the work was left standing.

At Claude, the Team plan gives you 1.25 times the usage of Pro per session. What happens when that runs out is not on the page I read.

On 23 August 2026, a business owner had two assistants transfer invoices out of Gmail and wrote on X: “the token usage was so high that both cut me off before I could finish a real workload.”

So the cost question with invoices that need to land against the right project is not whether it can be done, but whether it finishes within what you pay. Zapier does it differently from the other two: a run that goes over 75 tasks pauses and asks you to take a look, while Microsoft shuts the agent off and at ChatGPT the task stops.

Why is building it yourself not automatically cheaper?

Because then you hold the meter yourself, rereading included. On 16 September 2026, someone on the OpenAI forum worked out the numbers for one night. Two agents working together used 158,137,319 tokens, and 95.4 percent of that was rereading something the system had already seen. The bill came to $99.80.

That coordinator spent a lot of time waiting: 66 of its 89 wait steps timed out. You end up paying while nothing happens.

Anthropic says so itself: the announcement from 17 September 2026 states that a project with several concurrent threads hits the usage limit faster. So whether you actually want several agents under one coordinator with a team of ten or twenty is first a cost question: see what a coordinator with sub-agents costs.

What can you not know in advance today?

What a month is going to cost you. No vendor puts that on paper, because no one has measured your workload.

Microsoft writes that Copilot, the AI in its accounting and business software Business Central, “is included with your Business Central license at no extra cost.” In the same answer it says: “However, fair-use policies, quotas, or pricing might be introduced later.” When, and at what price, the answer does not say. The AI already inside your software has the same unknown meter as the rest.

For agents built in Copilot Studio, Microsoft publishes that one action costs 5 credits. Nowhere does it say how many actions one invoice takes. What Claude does on its own for a small business and what ChatGPT manages to do in your accounting software differs per task, and so does the bill. I cannot tell you in advance how much work is in your tasks either, because no one has measured it. Whoever claims to know is guessing.

Where does the bill go if this pace continues?

Up. OpenAI’s heaviest model from 3 September 2026 costs $10 per million input tokens and $50 per million output tokens. Its predecessor from 9 July 2026 is priced at $4 and $20, and OpenAI calls that a promotion running until at least 21 November 2026. So in two months the price of the strongest model went up two and a half times, while agents use more per task, because they keep working longer and reread their instructions every time.

My hypothesis: within two years, the big makers will bill your work mostly per action, and a fixed amount per seat will be the exception.

A fixed amount per seat only works if the provider knows what an average user uses. That usage is now climbing on two sides: the price per piece of text is going up, and so is the number of pieces per task. Microsoft and Salesforce already bill per action. Anyone who doesn’t track their own usage against that ends up paying for their heaviest user out of pocket.

Two things could break this. The price could fall, because that $4 and $20 is a promotion OpenAI can extend. And finishing faster means using less: at OpenAI, the new model scores 72.6 percent in forty minutes per task, against 65.7 percent in seventy-five minutes for its predecessor.

For a business, that means one question before you sign, with the answer on paper: what happens when the credit runs out, in the middle of a task? Follow up by asking what one of your tasks costs in actions, measured over a week. Also agree that nothing goes out the door without approval.

How do you start small, and what does it cost with us?

Pick one task that comes back every day and that someone now does by hand. Put it on the cheapest plan that can handle it for a month and count what it uses. Then you have your own number.

Bombos reads what comes in: email, chat messages, receipts and invoices, finds the matching data in the software you already use, and prepares the work with the reason behind it. No message goes to a customer and no payment goes out the door until someone at your business presses Approve, and that is technically enforced and always on.

With us, the model is a part that can be swapped: if a better model comes along, it goes in underneath without you having to set anything up again. We can put the strongest and the most cost-efficient model under it right away, and switch per client to what that client wants. Bombos runs on its own servers in Europe, and so do the models. What Bombos learns about your business does not sit in that model, but in what your people have approved and corrected, and that stays in place when the model switches. We read back what happens in the field every day and write it up in our daily newsletter.

A pilot costs €2,000 for three months, then €1,000 a month, excluding VAT and the same for everyone. That is more than ten seats at $20. With those seats, you buy a tool that you set up yourself; here, you buy work that gets set up for you.

Check the comparison yourself against the sources below. Leave your number, and we will call you back.

Your own measurement over a month tells you more than three price lists side by side.

Sources

Every source was opened on 20 or 23 September 2026 and every quote appears in it word for word.

Prices per seat

Prices per action

Usage and ceilings in practice

Prices per token, and the software