AI arrived with a straightforward financial promise. Businesses could get more done with fewer resources, and the person signing off on next year’s technology budget took notice. A leaner cost base and less pressure on headcount looked like the natural result of building AI into daily work.
Nine months into that experiment, many professional services firms are watching AI tool costs multiply faster than the original business case predicted, while AI reliability keeps shifting by task and by tool, so more checking gets built into everyday workflows than anyone planned for. The saving that was supposed to land on the P&L has not always shown up, and where it has, it often sits next to costs the original budget never accounted for. A law firm billing in six-minute increments notices that kind of shortfall fast, because it becomes a number the finance lead has to explain to the partners.
Why the subscriptions pile up
AI tools do not always give the same answer twice, and different tools can handle the same task very differently. Stanford’s 2026 AI Index found that hallucination rates across 26 frontier AI models ranged from 22 percent to 94 percent on one widely used evaluation, with some models sitting within a few points of each other on one benchmark before diverging sharply on the next. A confident answer from one tool is not automatically correct just because it came from AI, which is why a partner reviewing an AI-drafted client letter checks it anyway, without calling that a formal process.
That unpredictability changes how people work in practice. A confident output on a client memo gets checked against another source or against a second AI tool that seems to handle that type of task better, and the habit spreads through a firm quickly. One employee ends up paying for the tool that drafts well while a colleague pays for a different one that handles spreadsheets more reliably, and between them the business is now funding two subscriptions to do one job.
The subscriptions accumulate because people are compensating for a technology that has not yet earned unconditional trust. A business that set out to buy one AI tool can end up running three or four, each justified on its own terms and each adding to AI subscription costs that were never sized up in advance.
The bill doesn’t work like a subscription
Subscription sprawl is one layer of the cost problem. Usage-based pricing is another, and it behaves differently from any software cost most finance teams have budgeted for before. Many AI tools charge for the volume of text processed, so AI costs for businesses can grow automatically as adoption spreads through a company, independent of how many seats sit on the original contract. A traditional software license renews at a fixed amount a finance team can set and forget, but an AI usage bill tracks how much the tool gets used, so the same tool can cost twice as much in a busier month even though nothing in the contract changed.
The FinOps Foundation’s 2026 State of FinOps survey of nearly 1,200 finance and technology practitioners found that AI has become the fastest-growing cost category most of them manage, and nearly three-quarters of the organizations they represent had already spent past their original AI budget for the year. Uber’s experience makes the same pattern concrete. After giving thousands of engineers access to an AI coding tool in December 2025 and tracking usage on internal leaderboards, the company burned through its entire planned AI budget for 2026 within the first four months.
AI can still deliver real value even when spending runs past the plan. If it was brought in partly to control costs, what counts now is whether current spend sits below what it replaced, or whether usage has grown into its own line item, tracked separately from the saving it was supposed to produce.
Cutting people too soon
Workforce planning carries its own version of this risk. Some firms have looked at AI’s promised productivity gains and decided they can operate with fewer people, without waiting to see whether AI can reliably absorb the work those people did. That sequencing is the real problem, because a staffing decision made before the technology has proven itself at scale is a bet, and in a professional-services firm, the person let go is often the junior associate whose caseload looked automatable on a spreadsheet.
By Gartner’s own projection, three in ten workers whose roles are eliminated by AI over the next few years will be back on a payroll by 2029, often at a higher wage than the one that was cut. The reasoning is clear because workforce cuts drain institutional knowledge and shrink the pipeline of people who understand how the business runs, and both become expensive to rebuild once AI turns out to cover less of the job than expected.
Removing a salary shows up immediately as a saving. If the business later finds AI cannot reliably carry the work that person did, recruitment, onboarding, and a higher starting salary all re-enter the calculation, well after the original saving was already reported to the board.
Adding up the real cost
The real cost of AI spreads across more than one place in a business. A firm that opened with a single subscription can be funding several tools within a year, once employees have worked around each one’s blind spots, while usage charges add to that total as adoption grows, whatever the contract originally assumed about how many people would use the tool. Underneath both sits a staffing plan built on assumed productivity, and that becomes expensive to reverse once the assumption turns out to be premature. The number a vendor first quotes rarely reflects any of this.
AI can still make a business more productive, but that does not mean every dollar spent on it is producing a saving. Productivity and cost control are two different things to measure, and most businesses are only measuring one of them.
Is what your business is spending on AI right now less than what it replaced?
