For most of 2025 and into 2026, many companies told employees to use AI for everything, betting that heavier use would look like progress on a spreadsheet. Consultants have a name for this now: tokenmaxxing, which just means running up AI usage as high as possible to appear productive. The bill for that habit has landed, and it is bigger than most finance teams expected.
The deeper problem is not how much AI got used. It is that the newest and most useful kind, AI that can complete multi-step tasks on its own instead of just answering one question, needs far more computing power per task than the basic chat tools most people started with.
Gartner's research backs this up: worldwide spending on AI models and platforms is expected to hit 64 billion dollars this year, up 63 percent from last year. More capable models are getting more expensive to run, not less, because each task requires processing far more information.
This breaks a basic assumption a lot of executives made, that AI would get cheaper over time the way most technology does. Instead, the cost of running the advanced systems businesses actually want keeps climbing, since a single request handled by an autonomous AI agent can burn through many times more computing resources than an older style query. That gap between rising capability and rising cost is the real reason so many companies are now pulling back from blanket AI rollouts and asking harder questions about which tasks genuinely need it.
Governments are stepping in at the same time. The US now runs a voluntary system asking AI developers to let federal regulators review their most powerful new models before public release, partly to screen for national security risks. Europe's AI Act already requires companies operating there to be transparent about how they use AI and to train staff on it.
Together, these rules add legal and reputational risk right on top of the cost problem. Then there is the workforce gap: research from Forrester found that only a small share of office workers actually understand how to use AI tools well, and that share has barely moved in a year. The mismatch between what executives expect from AI and what employees can actually do with it is quietly driving a lot of the disappointment companies are now reporting.
This is worth remembering the next time a company announces AI driven layoffs. Many of those cuts are financial decisions dressed up as AI stories, since AI still cannot do most jobs from start to finish. The practical move for any business right now is not to quit AI, it is to measure it.
Track what each tool actually costs, tie that cost to a real result like faster work or better customer service, and cut anything that fails the test. The companies that get through this phase in good shape will be the ones that treat AI like any other line item: worth keeping only when it proves it pays for itself.