KPMG Finds Firms Lack Clarity on AI Costs

According to an as-yet-unreleased KPMG survey, only 26% of companies report having a comprehensive view of their AI usage costs, 50% say they have some visibility, and 22% report little or no visibility, the Wall Street Journal reports. The WSJ quotes Steve Chase, KPMG's global head of AI, saying KPMG is working with clients that have consumed annual token and cloud budgets within months, and that one client saw token usage rise sixfold. The WSJ also reports that AI providers including Anthropic and OpenAI charge enterprise customers, at least in part, by usage measured in tokens. The WSJ coverage includes examples of firms limiting access or building controls, noting Corning restricted employee access to certain AI tools and Life360 is seeking real-time token monitoring, according to Life360 finance chief Russell Burke.

Editorial analysis - technical context: Usage-based pricing in AI is commonly denominated in tokens, a basic unit representing input and output text processed by generative models. Per the KPMG briefing paper, organisations are adopting generative AI at scale while facing a new "total cost of ownership" dimension, including metered compute and cloud charges. The Wall Street Journal frames the move to token-based billing as part of a broader industry shift where vendors combine seat-based licensing with consumption meters, changing how costs map to activity.

The combination of rapid adoption and metered pricing creates operational risk for finance and engineering teams, according to the WSJ reporting and the KPMG briefing. The KPMG document cited in the coverage also reports 83% of respondents expect GenAI investments to increase over the next three years, highlighting pressure to manage variable spend while scaling deployments. The immediate consequence reported by multiple outlets is that several companies are imposing controls, building consumption monitors, or rearchitecting agents to reduce token usage.

Editorial analysis: Companies adopting AI at scale will need visibility into request-level consumption and cost attribution to teams, products, or workflows, an industry pattern observed across recent deployments. Engineering teams integrating models into production commonly encounter unexpected costs from agentic workflows, prompt engineering changes, and background retraining or embeddings updates. Finance teams used to fixed-subscription models face unpredictability when costs are driven by variable token volumes and cloud compute, creating cross-functional coordination needs.

For practitioners: Observers should track three indicators reported in the coverage: adoption of real-time token monitoring and chargeback tooling across organisations, vendor pricing changes that blend seat and usage tiers, and architecture choices that trade model size or context window for lower per-token consumption. Public reporting, such as KPMG and WSJ coverage, will also reveal whether firms move toward stricter governance controls or invest in tooling to optimize prompt/token efficiency.