Organisations estimate a quarter of their AI costs are wasted, and nobody owns the bill

Harness put a number on something a lot of finance leaders have been feeling. In its 2026 State of AI in FinOps report, based on a survey of 700 engineering leaders and practitioners across five countries, the company found that organisations estimate 26% of their AI costs are wasted (Source: Harness, 2026 State of AI in FinOps). For the one in five respondents spending US$1 million a month, that works out to about US$260,000 a month with no measurable return (Harness).

The reason that number matters is not the size of the waste. It is that most organisations can't say where it went. Harness described spending that is moving faster than any organisation's ability to see it, attribute it, or explain it (Harness).

Why the money outran the systems#Link to this section

AI costs don't sit in one tidy line item. They climb across infrastructure, software and models at the same time, across several providers, each with its own billing. So the bill arrives, it's bigger than expected, and no single team can explain the jump.

The gap starts at the point where the money is actually spent. Fewer than half of engineers, 45%, say they understand the cost of the AI features they build, and 56% say forecasting AI costs is based on guesswork rather than data (Harness). The incentives make it worse. Harness reported that 57% of engineers say their organisation encourages maximising AI usage regardless of tangible value (Harness).

Policy hasn't closed the gap either, because a policy nobody can check is only an intention. In the same report, 73% of organisations had AI cost policies in place, yet only 13% had basic visibility into what they spend on AI (Harness).

What good measurement looks like#Link to this section

Harness found that organisations reaching full AI cost maturity follow a consistent order. First they name a single accountable owner for AI costs before adding new tools or providers. Then they build one view of spending across infrastructure, software and models before trying to optimise it. Then they put cost data into engineering workflows at the point where models are chosen and deployed, rather than treating it as a finance-only concern (Harness).

We would add the other half of the equation, because cost is only one side of the ledger. Only 26% of organisations in the report had a robust method for measuring the business value of their AI spending (Harness). Knowing what a tool costs tells you nothing about whether it is working. Here is what we would track:

• A named owner per tool. If you can't say who is accountable for a given AI tool's cost and value, that tool is unmanaged by definition. Start every review by filling that name in.
• Adoption, not seat count. Licences bought is a purchasing fact. Weekly active use by the people meant to use the tool is the real signal. When those two numbers diverge, you have found waste.
• Reported value against reported cost. Ask the people using the tool whether it saves them time, and where. Put that next to the invoice. A tool with high cost and low reported value is where the US$260,000 goes.
• Time to value. How long from purchase to the first measurable outcome. If a tool has been live for two quarters with nothing to show, that's a decision to make.

Good looks like a short list you can go through in a meeting: each tool, its owner, its cost, its adoption, and whether the people using it say it's worth keeping.

The pattern is bigger than one report#Link to this section

This is not only a FinOps story. In July 2025 Commonwealth Bank said 45 customer service roles would go, after introducing an AI voice bot it said had cut calls by 2,000 a week. The Finance Sector Union said call volumes were in fact rising, with staff offered overtime and team leaders answering calls. On 21 August 2025 the bank reversed the decision, saying its initial assessment "did not adequately consider all relevant business considerations", and apologised to the employees concerned (Source: Commonwealth Bank statement, 21 August 2025, reported by Information Age and Bloomberg). The decision went out ahead of the evidence that the bot was working.

That is the through line. Across banking and across the wider enterprise, the money is often committed before anyone builds the measures that would show whether it paid off. The organisations that stop wasting a quarter of their AI budget will be the ones that measure value with the same discipline they measure cost, and that assign an owner before they assign a budget.