When AI usage becomes your largest cost

A large technology organization recently pulled a popular AI coding tool away from its own engineers. The tool was not broken. The engineers liked it and used it every day. That was the problem. The work ran on usage, the usage ran through a meter, and the bill outpaced the budget faster than anyone had planned for.

The story reads like a warning about one product. It is really about a cost that never showed up in the business case and only appeared once the work was underway. More AI tools now charge for what you use rather than for a seat you own, and the newest ones use a great deal.

It was cheap to pilot and expensive to adopt. That gap is the part most organizations never price.

The bill is set after you choose, not before

Flat pricing put the cost decision at the moment of purchase. You knew the number before you signed. Usage pricing moves that decision into daily work, where nobody is watching the meter and the number is different every day.

A seat has a known price. A workflow that calls a model hundreds of times, retries when it fails, reloads its context at every step, and keeps working while people are asleep does not. EY puts the cost of a single agentic interaction at around thirty times that of a simple chat exchange, because the agent does far more work to reach an answer. That cost is a function of how the system is built and how heavily it is used, and both are decided long after the contract is signed, by people who were never shown the meter.

Where the cost hides

The number that convinced everyone is rarely the number the business ends up paying. The real cost surfaces later, and it surfaces in places the pilot was never going to show.

• The pilot ran on a small team and a light load, so the figure everyone saw was the smallest figure the tool would ever produce.
• The tool does not ask one question and stop. It plans, calls other systems, retries, and reloads context, and every step spends. A single task can cost many times what anyone assumed.
• The heaviest users are the ones getting the most value, so spend rises exactly where the tool is working best. Success and cost climb together.
• The budget was set against a seat price or a pilot invoice, not against the workflow the tool would eventually run inside.

EY estimates that organizations tend to plan for only three of the seven cost categories an agent creates, and meet the rest during scaling rather than at the pilot.

A tool that is cheap to try and expensive to use has not saved you money. It has moved the cost to a place you were not watching.

Design for the bill before the pilot succeeds

A cost that grows with use has to be planned for, not discovered on an invoice. That means treating the spend as part of the design, not as a surprise the finance team absorbs later.

1. Price the workflow, not the seat. Before you scale, work out what one real task costs when the tool plans, retries, and runs at full volume. Model the busy day, not the demo.
2. Put the meter where people can see it. Usage that no one owns is usage no one controls. Cost teams now rank managing this kind of spend as their leading challenge. Give the spend an owner, a ceiling, and a signal that fires before the budget is gone, not after.
3. Keep a way out of a single model. A tool you cannot leave is a price you cannot argue with. The exit you never priced becomes leverage the vendor has and you do not. Build so that moving to a second model is a switch, not a rebuild.
4. Match the model to the task. Not every step needs the most capable and most expensive model. Send the simple work to something cheaper and keep the strong model for the work that earns it. Cost here is something you engineer, not something you accept.

The advisory point

The value of these tools is real, and the answer is not to hold them back. Adoption is still the point. What has changed is that adoption now carries a bill that grows with use, and a bill that grows with use has to be designed for rather than found. Gartner expects more than forty percent of agentic AI projects to be abandoned by the end of 2027, with cost among the leading reasons.

The organizations that struggle are not the ones that spend on AI. They are the ones that let the spending decide itself. Price the workflow before it scales, and the cost becomes a choice. Wait for the invoice, and the cost becomes the decision that was made for you.