Two years ago, 31% of FinOps practitioners were managing AI spending. This year it's 98% (FinOps Foundation). By that measure, tracking AI costs is now standard practice.
The overruns haven't stopped. In a February survey of 500 finance leaders at US and UK companies with more than 1,000 employees, 79% said they had AI cost overruns in the past year (DoiT). And the most counterintuitive finding: the companies with the most mature cost-management practices overran more often, 89% compared with 69% for early-stage teams, and by more, an average of 30.9% over budget compared with 16.1%.
That survey was commissioned by DoiT, which sells cost-management software, and better measurement probably explains part of the gap: mature teams catch overruns that others miss. But at minimum it shows that tracking spend doesn't stop it.
Visibility isn't control
A separate survey of CIOs by Futurum Research found 46.9% of enterprises running over budget on AI, with another 10% having no formal AI budget at all (Futurum). The two figures come from different groups and different questions, so they shouldn't be combined. Together they say overruns are common, not rare.
What organizations do when they overrun is revealing too. Among Futurum's over-budget respondents, 47.6% went looking for supplemental funding and 43.3% absorbed the overrun into the next planning cycle. About one in six reduced or paused the AI work. In other words, most overruns are handled after the fact, by finding more money, rather than prevented at the point the cost is created.
The more telling number is where controls sit. KPMG's Q2 2026 AI Pulse, as reported by PointFive (another cost-tool vendor), found that 66% of large US organizations run AI cost dashboards but only 36% have direct controls over token or usage consumption (PointFive). Most organizations can see the bill. Far fewer can cap it at the point it's incurred.
Why AI costs behave differently
Traditional software spend is mostly decided up front: licenses, seats, capacity. AI spend is increasingly decided by usage. Gartner forecasts that 2026 is the first year inference, running AI models, overtakes training in AI-optimized cloud infrastructure spend: $23.3 billion against $19 billion, rising to 59% of the total next year (Gartner).
Gartner expects that market to grow another 56.5% in 2027, to $66.1 billion. More of the bill each year comes from running models, not building them, and running them is a cost that rises with every user, every prompt and every automated workflow that calls a model.
Futurum lists the drivers of overruns: rising prompt and inference volume, model changes, new services reaching production, wider rollout across the business, and annual budgets set against costs that vary month to month. Pricing has also shifted from seats to usage, so consumption alone can push the bill up.
That means the cost of AI is set by how many people run how many requests through which models. Those are product and engineering decisions, made daily, often by people who never see the invoice.
What it looks like in practice
Uber is the clearest public example. Its CTO, Praveen Neppalli Naga, confirmed to The Information that the company had used up its entire 2026 AI budget by April (Forbes). Use of an AI coding agent among its roughly 5,000 engineers went from 32% in February to 84% in March. Average cost was $150 to $250 per engineer per month, with power users at $500 to $2,000.
Uber didn't do anything wrong. It encouraged adoption, adoption worked, and the budget was built on assumptions about usage that the rollout overtook. Naga said the company was going "back to the drawing board" on those assumptions while continuing its AI push. A budget that's annual and fixed was never going to hold a cost that depends on how enthusiastically 5,000 engineers use a tool.
The DoiT survey points to the organizational side. Only 15% of finance leaders can calculate AI ROI without significant bottlenecks, and accountability is split almost evenly between technology (55%) and finance (53%), meaning no single owner. Senior leaders are also more optimistic than the people doing the work: 93% of C-suite respondents rated their FinOps maturity as mature or better, compared with 60% of managers.
What to do about it
Put limits where the spend happens. Usage caps per team, per tool or per environment, enforced in the platform, not in a monthly report. Stack Overflow's 2026 survey found 38% of developers' organizations already set usage limits (Stack Overflow).
Route work to the right model. Not every task needs the most capable and most expensive model. A routing rule that sends routine work to a cheaper one is a cost control engineers can own.
Budget for usage as well as licenses. Treat a rollout of AI coding tools as a capacity plan: expected users, expected usage per user, a ceiling, and a review point when adoption passes a threshold.
Measure cost per outcome. Total spend tells you little. Cost per task, per feature or per customer interaction tells you whether the spend is worth it. Only 26% of the finance leaders DoiT surveyed had implemented AI unit economics.
Give it one owner. Someone needs to be accountable for both the value and the cost of each AI deployment. Shared accountability between technology and finance tends to mean neither acts in time.
We build AI-assisted engineering with structured review, and cost is part of that structure. If your AI bill is growing faster than you can explain it, let's talk.