What AI Actually Costs
A plain-English guide for operators who hate surprise bills.
Every operator asks the same question about AI, and almost nobody gives them a straight answer: what does this actually cost me? You can’t get a clean number because there isn’t one. AI is not priced like the software you already buy. It is priced two completely different ways, and the gap between them is where the surprise bills come from. Here is the whole picture in plain English, so you can budget for it before it budgets for you.
There is no single price
Ask “what does a truck cost” and you’d want to know which truck and how you’ll use it. AI is the same. A salesperson cleaning up emails and a developer writing software can run the very same tool and get bills that differ by a factor of a thousand. So stop hunting for the price. Start by figuring out which of the two pricing models you’re dealing with, because that decides everything else.
The two ways AI charges you
Almost every AI tool bills one of two ways.
The seat. You pay a flat monthly fee per person, the way you pay for a phone line or a seat of your accounting software. As of mid-2026, the standard plan for the big assistants sits right around $20 a month per person. ChatGPT, Claude, Google’s Gemini, and Microsoft’s Copilot are all within a dollar of each other. Heavier “power” plans run $100 to $300 a month per person for people who lean on it all day. The seat is predictable. You know the number before the month starts and it does not move. For most of your people, this is what you want.
The meter. Instead of a flat fee, you pay for exactly how much work you ask the AI to do, the way a taxi charges by distance or the power company charges by the kilowatt. What they count is words in and words out. Light use costs pennies. Heavy use does not. This is the model behind the tools that developers and automated systems use, and it is the one that bites.
| The seat | The meter | |
|---|---|---|
| How it bills | Flat monthly fee per person | Per unit of work — words in and words out |
| Typical cost | ~$20/month standard; $100–$300 for power plans | Pennies for light use; thousands for heavy automated work |
| Predictable? | Yes — you know the number before the month starts | No — invisible until the invoice lands |
| Best for | Most of your people, everyday chat-window work | Software integrations, automation, heavy development |
Where the surprise bills come from
Every horror story you’ve heard about a runaway AI bill came off the meter, not the seat. The finance-company employee who rang up $81,000 in a week building a side project was on the meter. Every time he asked the AI to work on his project it re-read the whole thing and the charge ticked up, and nobody was watching it climb. A seat cannot do that to you. The meter can, and it can do it over a weekend.
The meter is invisible until the invoice lands. There is no cashier and no “are you sure.”
The trap is that the meter is invisible until the invoice lands. There is no cashier and no “are you sure,” just a number growing quietly in the background while the work feels free. Producing words costs more than reading them, usually three to five times more, so the jobs that generate a lot, like long documents, software code, or anything running automatically without a person watching, are the ones that run up the tab fastest.
What it actually runs per person
Here is the honest range. A single seat for one of the main assistants is about $20 a month. A heavy individual on a power plan is $100 to $300 a month. On the meter, a light user might spend a few dollars a month and a heavy automated workload can spend thousands.
Two things most people miss. First, the headline price is often not the real price. Microsoft’s Copilot looks like $20 to $30 a person, but it only works if you also pay for the underlying Office subscription, so the true all-in cost can land at $50 to $80 a person. Always ask what else you have to buy to make the advertised price work. Second, the gap between companies is enormous. The spending tracker Ramp found the top one percent of companies spend about $7,500 per employee every month on AI, while the average company spends about eleven dollars. The leaders are not reckless. They have decided where AI earns its keep and they pay for it on purpose.
Five settings that keep the meter from running away
You do not need to understand the technology to control the cost. You need five things in place.
A dollar limit per person. Set it in the tool’s billing settings so no single account can run past a number you chose.
Spending alerts. Turn on the warnings that email you when usage spikes, so you find out in real time, not at month end.
One named owner of the AI bill. The same way someone owns every other line on your P&L.
Seat by default, meter by exception. Put general users on a flat seat; reach for the meter only where a specific job needs it.
A person in the loop on anything automated. Automation is where the meter runs fastest and quietest, so that is where a human checks first.
None of that requires a technical hire. It is the same spending discipline you already apply to travel, to company cards, to every other variable cost in the building.
Which model fits which job
A simple rule covers most cases. If a person uses AI through a normal chat window for everyday work, put them on a seat. It is predictable and almost always cheaper for steady use. If you are running AI inside your own software, automating a process, or doing heavy development, you are on the meter whether you like it or not, so that is where the limits and alerts go first. The danger zone is the meter plus heavy use plus nobody watching. Keep those three from lining up and you will never get the $81,000 surprise.
AI pricing will keep changing. The numbers here are accurate as of mid-2026 and some will move within months. But the two models will not change, and neither will the discipline: know which one you are on, put a ceiling on the meter, and decide on purpose where AI earns its keep. That is not a technology problem. It is an operating problem, and operating problems are the ones you already know how to solve.
If you want this mapped to your business — which of your people belong on a seat, where the meter is worth it, and what a sane monthly number looks like for a company your size — that is the work I do.
