SAN FRANCISCO — The spreadsheet landed on the chief financial officer’s desk with a number that did not look right. Uber’s internal data showed employees had used up the company’s full-year budget for AI tools in the first four months of the year, according to people familiar with the matter, and the spending was not slowing down. The tools were popular, the usage was real, and the bill was roughly ten times what the company had planned for.
Uber is not alone. The pattern, in which AI tools become essential so quickly that budgets cannot keep pace, is repeating across corporate America, and it is forcing a reckoning that technology executives have been anticipating since the tools went mainstream. The comparison that captures the problem: a typical software subscription costs a few hundred dollars per user per year, while AI tools, with usage-based pricing and expensive models underneath, can cost ten times that once employees actually use them.
The economics of the old software world favored the vendor. Companies bought licenses, and the cost was predictable, capped by the number of seats. The economics of AI tools are different: the bill grows with usage, and usage grows with enthusiasm. A tool that saves an analyst three hours a week gets used more, not less, and every additional use costs money. The result is a new category of expense that finance teams did not budget for and did not know how to forecast.
Uber’s response, according to people familiar with the matter, was to begin managing the problem the way it manages every other major cost: inventory what employees are actually using, negotiate with vendors, and set guidelines for which tools are approved for which tasks. The company has not stopped employees from using AI, and it has no plans to, but the era of unbounded experimentation is over. Every tool now needs a business case.
The phenomenon has a name inside technology departments: shadow AI, a cousin of the shadow IT that plagued companies in the early cloud era, when employees signed up for services without telling the IT department. Shadow AI is harder to control because the tools are so cheap to start using, a credit card and an email address, and so valuable that employees will not give them up. Blocking the tools is not a realistic option; managing them is the only strategy.
The budget pressure has implications beyond individual companies. The AI industry’s largest revenue streams run through usage-based pricing, with OpenAI, Anthropic, and their competitors charging per token or per seat with consumption on top. If the companies paying those bills start enforcing budgets, enforcing caps, and negotiating discounts, the growth rates that investors have priced into AI companies will face a new headwind. The demand is real, but so is the bill, and the people paying the bills are starting to push back.
CIOs have begun to talk about AI cost control as the defining IT problem of the second half of the year, according to executives who attend the industry’s budgeting meetings. The questions are practical: which AI spending is transformative, which is convenient, and which is waste? The answers are shaping up to be hard-nosed, with finance teams applying the same return-on-investment tests to AI tools that they apply to everything else, a standard the tools have not always had to meet.
The negotiating power is also shifting. Software vendors once sold to IT departments; AI vendors sell directly to employees, who adopt the tools and then ask IT to pay. As budgets tighten, the pendulum swings back: procurement teams are demanding enterprise agreements, usage caps, and volume discounts, and vendors who built their pricing on viral adoption are discovering that the buyers are not as enthusiastic as the users.
The internal debate that followed the spending revelation was as instructive as the number itself. Engineers argued that the tools were already paying for themselves in productivity, and finance countered that productivity was precisely the problem, since it could not be measured on a spreadsheet the way a license fee could. The disagreement exposed a deeper issue: companies are trying to manage AI costs with accounting systems built for a different kind of software. Until the metrics catch up, the argument over whether the AI bill is too high will be settled by whoever controls the budget.
For Uber, the episode is a case study in the gap between AI’s promise and AI’s bill. The tools deliver, employees love them, and the productivity gains are real. They also cost more than anyone expected, in a way no one knew how to predict. The company’s finance team has learned the lesson that every CIO is learning this year: the question is no longer whether AI is worth it, but how much of it the business can afford, and who gets to decide.
