Nearly a third of corporate leaders are tripping over the basics: KPMG says 29% struggle to understand and control operating costs as they scale enterprise AI deployments. That sounds like an internal finance problem, but it is actually a C-suite capability gap, because AI spending increasingly does not behave like traditional software budgets.

The pressure is coming from the pricing model change itself. In recent months, Anthropic, OpenAI, and GitHub have shifted some services away from flat-rate subscriptions toward usage-based billing, and that is where forecasting breaks. KPMG’s point is straightforward: when pricing tracks usage, many organizations are still building the capabilities required to forecast, monitor, and manage AI spending effectively.

Zoom out and you can see why this is baffling boards. Under subscriptions, “cost” is mostly a calendar exercise. Under usage-based pricing, cost depends on how AI gets used, by whom, and how often outputs get generated. That means spend can rise when teams get excited, pilots expand, or agent-like workflows run longer than expected. KPMG also found that a third of senior corporate leaders identified limited understanding of AI costs and economics as a challenge to deploying AI agents, which is a useful distinction: agents can quietly increase volume and iteration, even if no one changes the business ask.

This is not just a theory of runaway costs. The survey reports nearly half of organizations have rephased AI deployments when costs have outweighed expected value. Read that again. In a market that has been selling “move fast” as the operating principle, nearly 50% are putting brakes on timing because the unit economics did not cooperate with expectations.

Meanwhile, the strategy response is getting more selective. KPMG says lower-cost, high-fidelity models are the fastest-growing influence on AI strategy, up 7 percentage points from Q1. And the report pushes back on a common boardroom fear: these actions do not signal reduced confidence in AI. Instead, they suggest a growing willingness to evaluate where AI creates meaningful value and where it does not, concentrating investment where expected returns are strongest.

That selectivity is showing up in capital spending too, because the pricing shift is only one side of the equation. Capacity is another. The source notes that Amazon plans capital expenditure of around $200 billion this year, largely to provide capacity for AI in its AWS datacenters, an increase of 50% on a year earlier. Microsoft’s total capex is expected to reach $190 billion, up 61% from the previous year. Both companies are investing in forward-deployed engineering to help customers develop AI applications that will generate demand for the capacity being built.

The source gives two concrete examples of how vendors are trying to turn infrastructure into an adoption engine. Amazon announced a $1 billion investment in an AWS Forward Deployed Engineering organization to help customers adopt AI agents and reduce deployment timelines. Microsoft is providing $2.5 billion in funding for a new operating entity called Microsoft Frontier Company, described as “enabling customers to amplify their IQ with AI while refining their differentiated value in the markets that they serve.” Translation for a CFO: the spend is meant to reduce customer time-to-value and, indirectly, stabilize the economics of adoption by accelerating the path from experiment to production.

But even if you can model your AI usage and forecast your costs, governance becomes the other expensive failure mode. KPMG flags challenges around AI governance: the question of who takes responsibility for decisions made by statistical models prone to erroneous outputs, or “hallucinate,” as tech vendors would prefer. KPMG says executive accountability matters, but governance “ultimately succeeds or fails through day-to-day operating practices.” The report emphasizes that organizations need clear rules for when employees can intervene, who owns AI-related costs, how AI outputs are reviewed, and what happens when systems fail. It also notes that while most organizations report having at least some governance mechanisms in place, relatively few describe these practices as fully embedded.

And governance is not just a policy document. It touches the internal control environment, audit readiness, and how quickly teams can scale without creating a compliance or reputational mess. When pricing turns variable, governance also turns into cost governance, because every override, review step, and operational exception affects throughput and expense.

One more wrinkle: the source mentions an integrity issue around a separate KPMG AI report. Last month, GPTZero claimed a forensic review of KPMG’s October 2025 report, “Total Experience: Redefining Excellence in the Age of Agentic AI,” found only five of its 45 citations pointed accurately to the cited source, with errors ranging from misleading or invented details to references too vague to verify. The report was removed from some websites and KPMG issued a statement saying it takes accuracy and integrity seriously, that the report has been removed, and that it is reviewing circumstances, expecting all people to follow guidelines including human oversight to validate content and verify independent sources.

This is not about whether AI governance is hard. It is already obviously hard. It is about how rapidly the “AI business” is forcing everyone to tighten fundamentals at the same time: cost forecasting, pricing model literacy, deployment phasing discipline, and day-to-day control systems. For C-suite executives, the stake is simple: if your company cannot predict and govern AI spend under usage-based pricing, you do not just risk margin. You risk slowing the adoption curve while competitors who mastered the operating model move faster.