CIOs and CTOs spent years lauding AI. Now, with costs rising, they're putting limits on how it's used

After spending years cheering on AI, enterprise tech leaders are starting to put the brakes on how much employees can use it. The shift is not anti-AI. It is anti-surprise-bill.

At Samsara, CIO Stephen Franchetti has made a broad bet on the main tools from Anthropic, Google, OpenAI and Cursor, while also building an internal system that tracks AI costs daily. But the company has now capped usage for some non-technical staff, while research and development teams still get more room to experiment. Franchetti says it took time to find the right balance, but the point is to leave people with choices rather than hand them an unlimited buffet.

That kind of restraint is spreading because the numbers are getting awkward. Global AI spending is projected to hit $2.5 trillion this year, up 44% from last year, and some companies say their 2026 AI budgets are already running past expectations without delivering clear business value. Gartner’s Will Sommer puts the mood bluntly: AI is not a free lunch, and companies can burn through thousands of dollars per person on output that amounts to junk. In June, Gartner even warned that AI coding costs could surpass the average developer’s salary by 2028 as token use rises and pricing shifts toward consumption.

Docusign’s Sagnik Nandy found a practical fix inside his own shop. Engineers there have leaned hard into AI, with 75% of code reportedly initiated by AI, but the company changed its coding agents so they pull only the context needed for a narrow task instead of reading the entire code base. That move cut token usage by almost 50%. It is a neat little reminder that efficiency is not a slogan; it is usually a settings change.

Others are arriving at the same conclusion from different angles. Yum Brands says most AI tasks, maybe as many as 95%, can be handled by smaller and cheaper models, so it is pushing more training and treating digital spend more like headcount. Cigna has authorized more than 70 AI models and uses cheaper or smaller ones for tasks that do not need heavy reasoning. Compass, meanwhile, has set budgets for every engineer so the cost does not just vanish into IT. The era of “use AI everywhere” is giving way to “use the right model, or pay for the privilege.”