CIO Dive reported on August 6 the findings of the 2026 State of AI Cost Governance Report, in which enterprise cost management software provider Mavvrik surveyed 396 organizations across multiple industries in April and May. Nearly half of the organizations said unexpected AI spending issues had escalated to the board level, and a quarter of respondents said they had delayed or canceled AI projects due to unforeseen costs. Two-thirds reported that unexpected AI costs had materially affected at least one business decision, and a third said they had implemented emergency spending freezes as bills ballooned.
Costs aren't confined to model fees alone
Spending isn't limited to model costs—it's spread widely across developer tools, data platforms, infrastructure, and agentic workloads. The report's diagnosis is that this fragmentation, combined with a lack of oversight, has made it difficult for CIOs and finance leaders to grasp the full cost of AI initiatives and accurately measure return on investment. Mavvrik CEO Sundeep Goel said in the report that AI is fundamentally changing how infrastructure is consumed and how costs accumulate, and that IT spending, once predictable, has become fluid and distributed, making it increasingly hard to attribute.
Rita Sallam, research lead for data and analytics at Gartner, told CIO Dive in an email that this problem becomes more pronounced as organizations move beyond simple generative AI deployments toward autonomous agents. She explained that while the unit cost of models—such as per-token pricing—appears to be falling, the actual cost per completed task is rising as agentic workflows grow more complex and demand more advanced reasoning. As billing shifts to a usage-based model, she added, the cost of running complex autonomous agent workflows often ends up exceeding the very savings those workflows were originally meant to generate.
ROI is arithmetic
Ray Rike, CEO of Benchmarkit, noted in the report that everyone talks about AI's ROI, but ROI is arithmetic—and without knowing your costs, you can't accurately calculate it. Hidden costs are nothing new. Data released last year by Wasabi Technologies showed that nearly two-thirds of organizations had exceeded their cloud storage budgets due to accumulating unexpected usage and egress fees. In response to growing scrutiny, cloud providers like Google Cloud, AWS, and Microsoft have since lowered certain storage-related fees, and FinOps practices originally developed for managing cloud costs are now being applied to protect enterprises from ballooning AI bills.
If you're considering adoption
Sallam recommended that CIOs treat AI cost optimization as an architectural requirement rather than an after-the-fact financial reconciliation exercise, building cost optimization directly into agentic workflows. Her core message: match the model to the task, using the phrase 'avoid using a Ferrari for a job a Kia could do.' She also suggested routing simple queries to cheaper models and using smaller, domain-specific language models. She added that investments in governance, security, and data foundations need to keep pace with AI deployment to sustain long-term ROI, and that literacy training is needed to teach users the financial consequences of AI token usage—no one should be allowed to use the most expensive reasoning model just to check the weather.
For organizations here moving beyond pilots to embed agents into actual operations, the sequence should start with measuring cost per completed transaction, not the unit price written in a contract. The fact that a third of organizations resorted to emergency spending freezes also means that just as many only checked the numbers after the bill arrived. There's one thing worth checking this week: can you break down and attribute last month's AI-related bills by department and by task? If you can't, then neither your ROI calculations nor your effort to match models to tasks has actually begun.
Source: Surprise AI costs threaten enterprise implementations