SPECIAL FEATURE | AI token costs emerge as new challenge for enterprise CIOs
The report estimates that the surveyed enterprises spent about $2.5 billion on AI tokens last year. Without intervention, that figure could approach $3.6 billion within 24 months.
Enterprise artificial intelligence is creating a new cost category that many organizations are still struggling to see, forecast and connect to business value, as rising AI usage drives token consumption across employees, applications and increasingly autonomous workflows.
A new Accenture Research report, “The CIO’s guide to AI tokenomics,” found that enterprises are facing a widening gap between the amount they spend on AI and their ability to explain what that spending produces. The study surveyed 750 senior executives at companies with annual revenues exceeding $1 billion across 17 countries in July 2026, supplemented by 15 interviews with Fortune 500 technology and finance leaders.
The report estimates that the surveyed enterprises spent about $2.5 billion on AI tokens last year. Without intervention, that figure could approach $3.6 billion within 24 months. Token costs currently rank third among AI cost drivers, behind infrastructure and software development and maintenance.
The issue is not simply that AI is becoming more expensive. Token prices are falling while consumption is rising, creating a counterintuitive spending pattern.
Companies surveyed expect token consumption to increase 78% over the next 24 months while the price per token falls 19%. Accenture estimates that, without sufficient optimization, token costs could increase 44% over the same period.
The report links the pattern to the Jevons paradox: as AI becomes cheaper to use, organizations tend to consume more of it rather than simply pocketing the savings. Ninety-five percent of respondents said they would increase AI use in some form if token prices fell by 25% or more. Only 5% said they would bank the savings and keep usage broadly flat.
For CIOs, that creates a problem that conventional technology budgeting was not designed to handle.
“One in three organizations has already exhausted its annual token budget before the year ended,” the report said. It cited examples including a major US financial institution that experienced a five- to 10-fold increase in six months and a European bank whose annual token spending rose from near zero to $5.8 million in a year.
The growth is also being driven by agentic AI. Unlike a simple prompt-and-response interaction, an agentic workflow can trigger multiple model calls, expand context windows and repeatedly process information. Accenture estimates that agentic workflows currently account for about 11% of token consumption but are likely to drive the next wave of unbudgeted cost growth.
The report identifies another problem: enterprises frequently use more expensive models than necessary.
Accenture analyzed 9,368 occupational tasks from the O*NET database and found that fewer than 10% genuinely required frontier-model capabilities. The remaining 90% could be handled by capable mid-tier or open-weight models, which can cost substantially less per token.
“10% of people are pounding the table … [saying] ‘Why don’t I have the latest model?’ And then you have basically [the] majority of an organization using this as search,” a vice president of enterprise AI strategy at a major global technology company told the researchers.
The potential savings can be significant. The report cites one enterprise deployment in which changing the model reduced token consumption by roughly 90%, while prompt caching produced additional savings.
But reducing costs is only half of the problem. Companies also need to demonstrate what their AI spending is producing.
Only about 20% of every dollar spent on enterprise tokens can currently be expressed as a quantified financial outcome, according to the study. The remaining 80% may be associated with productivity or other benefits but cannot be translated into a financial figure that a CFO, CEO or board can readily evaluate.
The report found that organizations with stronger financial accountability perform differently. Financially attributed value rises from about 5% under no cost allocation to 13% with informal reporting, 23% with showback and 32% with chargeback.
Showback makes consumption visible to a team, while chargeback assigns the financial cost to the team using the resources. Accenture found that only 7% of organizations currently have some form of chargeback.
The report’s proposed response is what it calls five disciplines for AI tokenomics: workload-level visibility, financial accountability, smarter model routing, proof of value and AI skills.
The first requirement is visibility. Organizations need to know which workload is consuming tokens, which team owns it, which model is being used, how much it costs and what output it produces. That information then becomes the foundation for routing workloads to appropriate models and assigning financial responsibility.
Accenture estimates that organizations currently optimize about 23% of token consumption. To keep spending flat as usage grows by nearly 80%, they would need to reach roughly 31% optimization. At that level, the report estimates that every $10 million in annual token spending could avoid about $4.4 million in cost growth over 24 months while making an additional $1.15 million in value financially provable.
The report argues that the objective is not simply to restrict AI usage. Instead, enterprises need to make cost and value visible at the point where AI is deployed.
That means defining the expected business outcome before a workload reaches production, establishing who owns it, selecting the least expensive model capable of performing the task and measuring the result in financial terms.
The broader shift is from treating AI as another software expense to treating AI consumption as a production input that requires continuous management.
As AI models become cheaper and more capable, the report suggests that consumption will continue to expand. The financial question for CIOs, therefore, is increasingly not how much AI costs in isolation, but whether organizations can determine what each unit of AI consumption actually accomplishes.
“The economic question is not … how many [tokens] were consumed, but what did that token actually do?” a director of AI strategy at a Fortune 500 financial services institution told the researchers.
For enterprises moving AI from experimentation into production, that question could become as important as the technology itself.
Full disclosure: All news articles published on the TechSabado website are written by human journalists, unless otherwise specified. Final text editing is also performed by human editors, with artificial intelligence (AI) used only to assist with additional grammar and style guide corrections..
————————————————————————-
TEN YEARS OF TECHSABADO!
PLEASE LIKE our FACEBOOK PAGE and SUBSCRIBE to OUR YOUTUBE & SPOTIFY channels.
WATCH TECHSABADO ON OUR YOUTUBE CHANNEL:
WE ARE ALSO ON SPOTIFY
WATCH OUR OTHER YOUTUBE CHANNELS:
PLEASE LIKE our FACEBOOK PAGE and SUBSCRIBE to OUR YOUTUBE CHANNEL.
