Your AI Bill Has No Speed Limit. Set the Cap Before You Turn Anything On.

In plain terms: A lot of the AI tools you'll sign up for this year don't charge a flat monthly fee anymore. They bill by usage, by the token, like a utility meter, and that meter has no ceiling unless you set one. Big companies have already been blindsided by surprise bills, and the same trap is waiting for a one-person shop. Here is why it happens, and the single setting to find before you switch any AI tool on.

For years, software pricing was simple: a flat fee per user per month. You knew your bill before it arrived. AI is quietly breaking that, and the new model can bite.

The meter is already running

In June, TechCrunch reported what it called "the token bill coming due." The clearest example has names on it. Uber gave its engineers AI coding tools, usage took off, and the company burned through its entire 2026 budget for those tools in about four months. It responded by capping each employee at $1,500 a month. Uber's own CTO said he personally ran up around $1,200 in usage during a single two-hour demo.

(You may also have seen the eye-popping "$500 million AI bill" headline. Treat that one with caution: it traces back to a single anonymous consultant describing an unnamed client, and no company or vendor has confirmed it. The Uber numbers are the ones with real names attached.)

This is not only an enterprise problem. The same usage-based "credits" billing is now showing up in everyday tools. ChatGPT's Excel and Sheets add-on meters credits per task, and Microsoft's Copilot now bills usage in "Copilot Credits" on top of the monthly license. The model is coming to the tools you actually use.

Why the meter has no ceiling

Here is the part that traps people. The price per token keeps falling, so it feels like AI is getting cheaper. But an AI agent works far faster than a person, and the more capable ones happily burn tokens to finish a job. A falling price times soaring usage still equals a bigger bill.

The data backs this up. The engineering-analytics firm Jellyfish found that per-developer token use jumped roughly 18.6 times in nine months. And the heaviest users were not getting proportionally more done: they spent about ten times the tokens to get roughly twice the output. Faster and more does not automatically mean cheaper or better. A tool with no spending limit runs exactly as fast as you let it.

The one move to make

You do not need to avoid usage-based tools. You need to put a ceiling on them before you start. Three guardrails, about ten minutes:

• Set a hard spend cap on day one. Find the billing or limits setting and put in a real monthly cutoff, not just an email alert. Pick a number you would be fine paying twice, so a surprise can never wreck your month.
• Use a cheaper model for the routine stuff. Most tools let you pick which model runs. Save the expensive one for the hard jobs and let a smaller, cheaper model handle the high-volume, simple work.
• Watch the first week before you scale. Check the meter every few days for the first month. Once you know what "normal" costs, you can raise the cap with your eyes open.

The honest part

Usage-based AI is genuinely good for a small business. You only pay for what you use, and you are not buying ten seats you do not need. The danger is not the model, it is turning it on without a limit. And here is the part that favors you: you control your own card, which means this is a fix a 10,000-person company cannot make as cleanly as you can. Set the cap, then let it run.

This is the kind of unglamorous setup we handle for the businesses we work with: the AI that earns its keep, with the guardrails that keep it from surprising you. If you are about to wire an AI tool into your business and you are not sure where the spending limit lives, that is a quick thing to check, and a good conversation to have.