1 in 4 AI Dollars Wasted: The Ugly Truth Behind Fragmented Model Pricing

For AI engineers and data science teams, cost often feels like someone else’s problem—until the CFO freezes the GPU budget. This report lays bare the technical mess: multi-model pricing, opaque SaaS add-ons, and zero ownership of the total bill. Embedding cost observability into the MLOps pipeline isn’t a nice-to-have; it’s the only way to sustain the velocity of AI experimentation without bleeding cash.

A new Harness report has delivered a stark warning to enterprises pouring money into artificial intelligence: one in every four dollars spent on AI goes to waste. The finding, drawn from a broad survey of IT and finance leaders, underscores a crisis of cost management that threatens to undercut the business case for AI adoption just as budgets balloon. The scale of the waste is not a niche problem; Harness found that the gaps in visibility, ownership, and forecasting are consistent across companies of all sizes and geographies, from those spending $300,000 a month to those burning through $3 million.

“

Instead, costs are often an afterthought, owned by no one, and forecasting is shockingly primitive: more than 50% of respondents admitted to forecasting AI spend through guesswork rather than data, and over 40% still rely on spreadsheets.

The root of the waste is a fundamental mismatch between how AI costs are incurred and how organizations track them. Traditional IT and cloud spending revolves around servers, storage, and network egress—categories that are relatively easy to tag, allocate, and forecast. AI spending splinters across a far more complex spectrum: compute infrastructure for training and inference, consumption-based foundation model APIs, SaaS subscriptions for productivity tools like AI copilots and coding assistants, and managed AI services. Most organizations now rely on three or more major AI providers, each with its own opaque pricing tiers, discounts, and billing cadences. This multi-provider sprawl creates blind spots that ordinary cloud cost management tools were never designed to address.

Productivity software has emerged as a particularly sneaky drain. Tools like Microsoft Copilot or GitHub Copilot are procured as seat-based licenses, often expensed through departmental budgets rather than centralized IT. They look like ordinary software, not infrastructure, yet they can collectively rival compute costs. Without a clear mapping of which teams are using what, and whether that usage translates to actual productivity gains, enormous sums leak into underutilized or redundant licenses.

The lack of clear ownership compounds the problem. More than half of surveyed businesses have no dedicated person or team responsible for AI costs. In the cloud era, the FinOps discipline eventually embedded cost accountability within engineering and finance teams. AI has not yet reached that maturity. Instead, costs are often an afterthought, owned by no one, and forecasting is shockingly primitive: more than 50% of respondents admitted to forecasting AI spend through guesswork rather than data, and over 40% still rely on spreadsheets. With AI budgets now material enough to attract CFO scrutiny, this ad hoc approach leaves organizations vulnerable to sudden overruns and missed innovation opportunities.