Key Findings

  • AI spending is becoming increasingly difficult to predict and manage, with unexpected costs forcing some organizations to delay, freeze, or abandon related initiatives—according to a recent report by AI cost management company Mavvrik.
  • 62% of surveyed organizations said that over the past year, a singleunexpected AI costmaterially changed a business decision. Among these organizations, 40% required board-level involvement, 33% implemented emergency spending freezes, and 25% delayed or canceled AI projects. The data comes from the "2026 State of AI Cost Governance Report" released last month by Mavvrik in partnership with software-as-a-service performance metrics company Benchmarkit.
  • "Enterprises are experiencing significant AI cost overruns," said Sundeep Goel, CEO of Mavvrik, in an interview.

Deeper Analysis

These findings come as AI costs are increasingly tied to the consumption of "tokens"—units of data processed by AI models. As workloads grow, token consumption can quickly accumulate.

Mavvrik found that 43% of enterprises cited token costs as one of the top sources of unexpected AI spending.

In a July report, Gartner noted thattoken consumption is becominga "significant component" of enterprise operating costs, yet its connection to business outcomes often remains unclear.

The research firm stated that CFOs need clearer visibility into how tokens are used, the value they generate, and whether consumption is economically justified. Gartner recommends that enterprises shift from merely tracking AI costs to "AI unit economics," linking token consumption to measurable outputs and outcomes.

The goal is not necessarily to minimize token spending, but to "right-size" it based on the value of the outcomes delivered.

Another Gartner reportpredictsthat by 2028, AI coding costs will exceed the average salary of developers, driven primarily by rising large language model token consumption and a shift toward consumption-based licensing models.

Mavvrik's research shows that volatility in AI pricing makes cost estimation more difficult. Only 11% of organizations said they can keep AI spending forecast errors within plus or minus 10%, down from 15% in 2025.

Complicating matters further, "shadow IT"—where employees adopt technology without the company's knowledge or approval—is also in the mix.

The Mavvrik report found that 98% of engineering organizations use AI coding assistants, but only 42% include developer AI tool spending in their AI cost reporting, meaning many organizations are not fully accounting for these costs.

Goel said these findings reflect a lack of discipline in AI spending, a problem that is becoming more pronounced as costs continue to rise. "It's a bit of a perfect storm of wasted money," he said.

Goel draws parallels to the early days of cloud computing—when development teams could sign up for services with corporate credit cards without centralized oversight. "What we're seeing now is a similar situation," he said.

This can lead to fragmented, decentralized AI spending that is harder to track and control.

"If there's no control at all over who can spend on what, it's impossible for a company to execute a budget," Goel said.