As companies mature in their AI deployment and double down on their previous investments into the technology, pressure on executives has reached a fever pitch.
Early 2026 was the era of tokenmaxxing, or ramping up use of AI compute units as much as possible to appear productive. But momentum from AI providers to transition from flat-rate subscriptions to consumption-based pricing has ramped up in the last few months.
Maximizing token use went from a novel concept to the flashiest strategy and then an unmanageable scenario, all in the span of six months, said Nicholas Merizzi, principal at Deloitte Consulting.
The “use-AI-for-everything” mindset many companies adopted in 2025 and into 2026 now bears a much larger price tag.
“Before CFOs could even get it on their radar, the bills, the damage had been done,” Merizzi said.
Rising AI spend is just one factor leading companies to take a more pragmatic approach to AI deployment in recent months, consultants and experts told CIO Dive. Shifts in federal and global tech policy, coupled with less-than-ideal workforce adoption of AI, has made the technology’s deployment more difficult in 2026.
“Not everyone needs access to these systems,” said Will Sommer, a senior director analyst at Gartner. “And not every workflow needs to [use] AI.”
Exponential costs
Enterprises are reporting benefits from AI, but not the cost savings they might have expected. A recent SAP report found that AI use wasn’t saving organizations money, but it did help employees create insights, make decisions and interact with customers.
“The models produced by leading AI labs are getting more token-hungry faster than they are getting cheaper,”
Will Sommer
senior director analyst, Gartner
It’s a tough reality for companies that are spending millions to embed the technology into their workflows, while knowing the cost for AI will only grow — Gartner recently projected spending on AI models and platforms will increase 63% from last year, reaching $64 billion.
Although AI providers report foundational model costs are improving, and their operations are getting more efficient, it’s not the full picture, Sommer said. As models advance, they also get more expensive to operate. Complex uses, such as agentic models or agents, can require three to five times more tokens for a single query than earlier models.
“The models produced by leading AI labs are getting more token-hungry faster than they are getting cheaper,” Sommer said.
Token costs are not falling, they are widening, he said. Cost increases from escalating model capability and token consumption will soon outweigh any cost savings companies find from AI efficiencies — if they haven’t already, Sommer said.
Vendor fragmentation, inconsistent pricing models and constant price volatility are additional cost challenges enterprises face, Merizzi said. Executives are debating how to define AI value, but finding it has become increasingly important as the true cost of the technology reveals itself.
“The ability to link the spend to business outcomes is still evolving for a lot of our clients right now,” Merizzi said.
At this point in the AI lifecycle for most companies, CIOs should be able to connect some amount of their tokens to an increase in revenue, a positive difference in the client experience or speed and quality of their team’s work.
“If you can’t do that, you should be making adjustments, and use lower cost models and alternative means,” Merizzi said.
Shifting AI policy
Although American AI providers and the companies that use them were mostly spared from government regulation of the technology for most of its lifecycle, 2026 ushered in a new era of oversight.
President Donald Trump signed an executive order in May establishing a voluntary review of frontier AI models to assess safety vulnerabilities before they’ve been released to the public. It aims to screen AI models for national security concerns, such as those raised by Anthropic’s Claude Mythos unveiled in April.
“The Trump administration has largely been all gas, no brakes, on advancing innovation somewhat in this lens of anti-China,” said Jeff Le, managing principal at tech consultancy 100 Mile Strategies. “And now it’s realizing, ‘Oh we have these tools that we don’t know how we can control.’”
AI has presented national security risks in a way prior technology didn’t. It’s given governments and foreign actors more access and options for both digital campaigns and physical operations, Le said. In the process, it also introduced more risk for enterprises.
The core transparency rules and general oversight powers of the EU’s AI Act, which includes prohibitions on unacceptable use cases and AI literacy obligations, took effect earlier this month and applies to all AI companies that wish to operate in Europe.
“It’s about whether an institution will, by surprise, be in the crosshairs of policymakers for doing what seemed to be like the normal course of business.”
Miranda Bogen
chief technologist at The Center for Democracy and Technology and director of the organization’s AI Governance Lab
Some AI labs have been trying to get ahead of regulations. OpenAI in May released its Frontier Governance Framework, a look into the company’s safety and security practices. The company said it planned to align with emerging state and global AI regulations.
The policy landscape isn’t one many companies have had to navigate before, according to Miranda Bogen, chief technologist at The Center for Democracy and Technology and director of the organization’s AI Governance Lab. It’s costly and time consuming to make sound decisions, she said, and evolving conditions present a risk factor that many companies don’t want to deal with.
“The policy landscape is so volatile,” Bogen said. “It’s not only about uncertainty. It’s about whether an institution will, by surprise, be in the crosshairs of policymakers for doing what seemed to be like the normal course of business.”
Enterprises have an appetite for an institutionalized review of AI models, because they’re risk-averse, Bogen said. But Trump’s executive order reviews a narrow set of the newest models and risk factors that most enterprises don’t adopt immediately. The EU’s AI Act will likely present more immediate ramifications for companies as there’s still a lot of uncertainty around what compliance looks like, Bogen said.
“It would make sense for institutions to be treading carefully there,” Bogen said.
Upskilling the workforce
Humans are a key part of successful AI adoption, said J.P. Gownder, VP and principal analyst at Forrester.
But there’s a large gap between CIOs’ expectations of their workforce and the reality of their abilities, he said. In a 2025 survey of how organizations are using AI, Forrester found only 16% of information workers had a high understanding of AI tools, compared with 43% who were in jeopardy of misunderstanding it.
“The lack of understanding, skills and ethics that employees writ large have is a huge barrier to not only adopting AI successfully, but to driving good outcomes, positive return on investments and all the things that we are challenged by at the moment,”
J.P. Gownder
VP and principal analyst, Forrester.
“That means you’re not really ready to use AI at work, but when you do use it, it’s generating risks for the organization,” Gownder said.
Forrester also found that fewer than half of surveyed information workers said they know when to question the output of AI, and 42% said they sometimes refrain from using AI due to ethics, privacy or business risk.
Employees are reporting that they feel personal risk from autonomous systems because of their lack of control over how and when AI is deployed. As a response, they’re limiting their use. When they do use AI, business tasks represent just one-quarter of total AI usage by professionals, according to CompTIA data.
“The lack of understanding, skills and ethics that employees writ large have is a huge barrier to not only adopting AI successfully, but to driving good outcomes, positive return on investments and all the things that we are challenged by at the moment,” Gownder said.
The imbalance of executive boards’ expectations and their workforce’s skills is usually institutional, Gownder said. Within an enterprise, a majority of employees are not technologists and are not given the proper resources to learn this new part of their role. Most companies are skipping the organizational overhaul needed to truly embed AI and train their employees on it.
The technology is at the center of scores of layoffs this year, but executives who think they can replace roles with AI agents have outsized expectations, Gownder said. Most employees use AI to help accomplish some of their tasks, but not to do their whole job.
“AI is being blamed for financialized layoffs that have nothing to do with AI,” Gownder said. “AI isn’t really ready to do that work.”
Where IT leaders go from here
CIOs who are overwhelmed by changing AI cost structures, regulations and workforce needs can find some solace in knowing there are ways to regain control.
To start, tech executives need to get a clear picture of what their token consumption looks like, Sommer said. To keep costs in check, they can swap AI models for more domain-specific ones, or open source models. They can also change how the models run, playing with factors such as latency to get more efficient use.
CIOs can also change who or what gets to run the models, adjusting the number of loops it makes, or setting permissions for who can operate them and when they can run. Lastly, organizations should change physical workflows to reflect the operational advantages of AI.
“If AI spending is volatile and exceeds expectations, the models might not actually be the problem,” Sommer said. “The issue may be a misalignment between how humans think about the problem and how AI actually solves those problems.”
Executives who do not understand the business use cases and how they’re going to measure AI success should reel in their investments until that’s clarified, Le said. They also need a realistic picture of their employee’s abilities.
“How much do people really understand what is being used and deployed accordingly?” Le said. “That’s really hard if you’re [coming] from a CIO’s perspective, wrapping your arms around both an education and re-education of people at all levels.”
Every organization — even those with early AI adopters or tech-natives — will need to increase the AI knowledge of their workforce to see ROI, Gownder said.
Organizations shouldn’t rely on online, on-demand training, Gownder said. He recommended companies devote time to learning how to solve problems using AI through sandbox hackathons for non-technical employees. It’s important that organizations don’t push employees to learn AI skills on their own time, he said.
It’s clear the days of racing toward a cliff’s edge with an “AI-or-bust” mentality are gone, Le said. The technology is too pricey, too risky, too far from seamless integration at this moment to continue being the magic touch many companies felt they were promised.
But it is certainly not on its way out. Between the choice to accelerate or hit the brakes on their AI operations, Le said companies will continue to hit the gas.
“If you brake, you will be penalized. With the accelerator, you may be in a stronger position if you can get enough oversight, accountability and compliance to manage it,” Le said. “But it is a risky point. Hopefully, if you hit the edge of the cliff, you have built enough wings that you can fly.”
