After years of championing artificial intelligence across their organizations, Chief Information Officers and other C-suite technology leaders are now tightening the reins as expenses climb and returns remain uncertain. The pullback marks a notable shift from the adoption phase—marked by hackathons, broad tool access, and experimentation—toward a more disciplined operational phase reminiscent of how enterprises eventually applied cost governance to cloud computing after an initial land-grab period.

At Samsara, Chief Information Officer Stephen Franchetti has enthusiastically adopted a broad suite of AI tools for the tech firm's 4,100 employees. He has authorized Anthropic's Claude, Google's Gemini, OpenAI's ChatGPT, and the AI coding agent Cursor, while also building an internal system that tracks AI expenses on a daily basis. The need for such granular monitoring underscores how usage-based, token-driven pricing—still relatively novel in enterprise software—differs fundamentally from the per-seat subscription model that most companies are accustomed to budgeting for. Recently, however, Samsara has imposed usage caps for some non-technical employees, while groups such as research and development—which rely on AI for intensive coding and data analysis—retain greater latitude to experiment.

"It took us a while to settle on the right caps, to make sure everyone was well served," says Franchetti. "But it puts people in the position where they're kind of in control and they can make choices as to which models they use."

The shift reflects a broader trend across corporate America. Global AI spending is projected to reach $2.5 trillion this year, a 44% increase from prior-year levels. After years of promoting AI adoption through training courses, hackathons, and broad employee access to coding tools, agents, and chat assistants, some technology leaders are now limiting how frequently these tools can be used. They are also retraining staff to work with smaller, cheaper AI models capable of handling many routine workplace tasks.

A number of companies have reported that their 2026 AI budgets have already exceeded initial projections without producing corresponding business value. AI hyperscalers have taken notice, responding with cheaper models and price cuts. The competitive dynamics among the leading model providers—OpenAI, Anthropic, Google, and others—have intensified as enterprise customers signal that cost efficiency, not just capability, will determine which platforms win long-term contracts.

"2026 is the year of everyone finding out that AI is actually really hard," says Will Sommer, a quantitative modeling and economic forecasting expert at research firm Gartner. "It's not a free lunch. It requires a lot of thought and effort to get right." Sommer cautions that companies can easily spend thousands of dollars per employee on AI tools whose output is essentially useless and fails to improve productivity.

In June, Gartner issued a bearish report warning that AI coding costs would overtake the average developer's salary by 2028, driven by rising token consumption and a shift toward consumption-based pricing models. The projection highlights a tension at the heart of enterprise AI adoption: the tools most tightly integrated into developer workflows—coding agents that read, write, and debug code—are also among the most token-intensive applications, making them a primary driver of cost overruns.

At Docusign, the electronic-signature software company, Chief Technology Officer Sagnik Nandy says the organization has taken the cost challenge seriously. "We've taken this very seriously, both in the internal use case and we are propagating those learnings externally," he says. Internally, Nandy notes that every engineer has adopted AI tools and that 75% of the code they develop is initiated by AI.

However, Nandy quickly discovered that AI code agents were designed to access Docusign's entire code base for context before executing a task. "That's a lot of tokens, because you're trying to read everything," he says. Nandy reconfigured the coding agents so that the default setting pulls only the context relevant to the specific task a developer is working on, reducing token usage by nearly 50%. The optimization illustrates how engineering teams are learning that prompt architecture and context management—not just model selection—can dramatically affect the unit economics of AI deployment.

Jim Dausch, chief digital and technology officer at Yum Brands—the operator of KFC and Taco Bell—says AI token usage is not yet "a material number, but the trajectory was one that we're watching." Earlier this year, he noticed both AI token usage and expenses rising within the restaurant company.

Dausch contends that a large majority of tasks assigned to AI tools—perhaps as high as 95%—can be handled by more basic, less expensive models. Yum has since promoted additional training on AI model selection and advised business leaders to manage digital spending with the same rigor they apply to departmental headcount budgets. "We're trying to kind of democratize where the costs live and how they're managed, so it isn't just an IT line item," says Dausch.

At Cigna Group, the healthcare giant has cast a wide net by partnering with multiple AI hyperscaler vendors, enabling employees to use small language models or earlier, cheaper versions for tasks that do not demand significant reasoning capabilities. "The way you really run up costs is you use the most expensive models with no guardrails around them," says Katya Andresen, Cigna's chief data, digital, and AI officer. Cigna has authorized more than 70 different AI models for internal use. While Andresen acknowledges that compute and AI token usage has increased, total spending is not rising at the same pace thanks to this multimodal approach. The strategy reflects a maturing view of AI tools as a tiered portfolio—matching model capability and cost to the complexity of each task rather than defaulting to the most powerful option.

Shay Artzi, CTO at real-estate brokerage Compass, says his AI investments have targeted three groups: engineers, the company's AI Assistant tool that automates tasks for real estate professionals, and the broader corporate workforce. Engineers were the first to use AI extensively, but Artzi says he was cautious about token spending from the outset and never mandated that all code be written with AI assistance.

Compass piloted several AI coding tools before settling on partnerships with two industry leaders, Anthropic and Google, with certain financial constraints in place. "We also put budgets for every engineer, so they are aware of how they're spending," says Artzi.

By John Kell. This story was originally featured on Fortune.com.