The board asked what the AI spend returned

In an RGP survey of 200 US finance chiefs, reported by CFO.com in December 2025, 14% said they’ve seen clear, measurable impact from their AI investments. The other 86% are spending on faith. Two-thirds expect the impact to arrive within two years, which is what people say when the spend is committed and the evidence isn’t.

The spend, meanwhile, has stopped being small. In CloudZero’s 2026 report with Benchmarkit, AI spending now exceeds $10M a year at 40% of surveyed companies, and only 22% of finance executives can tie that spend to business outcomes. Boards have noticed the gap between those two numbers. If 2024 was the year of the AI mandate and 2025 the year of the pilot, 2026 is the year the board asks for receipts.

The results are real, and modest, and concentrated

The honest published numbers cluster in an uncomfortable middle. McKinsey’s State of AI survey (November 2025, 1,993 respondents) found 39% of organizations report enterprise-level EBIT impact from AI, and most of those put it under 5% of EBIT. Gartner’s April 2026 survey of 782 infrastructure and operations leaders found 28% of AI use cases fully meet ROI expectations, while 20% fail outright; 57% of the leaders surveyed reported at least one failed attempt to apply AI in their operations, failures Gartner attributes largely to initiatives that were overly ambitious or poorly scoped.

So AI produces returns for a minority of adopters. The interesting question is what that minority does differently, and a large part of the answer is plumbing: they can see their numbers.

ROI is a telemetry problem, not a spreadsheet problem

Return on investment has a numerator and a denominator, and most organizations can produce neither at the level where decisions get made.

The denominator is what the feature costs: model calls, retries, context, the agent loop that fans one task into thirty calls. If spend arrives as one monthly invoice, you cannot say what any individual feature costs, which means you cannot say whether it’s worth what it returns. That attribution work is the subject of FinOps for AI, and it’s the easier half.

The numerator is what the feature returns, and this is where most ROI cases quietly die. A claim like “support resolves tickets faster” needs a baseline: how long resolution took before, measured, for a representative period. If nobody recorded the before, the after is unfalsifiable. And it needs a quality floor. A model that answers twice as fast and wrong more often is a cost increase wearing a productivity costume, which is why an evaluation harness is as much a finance artifact as an engineering one.

One popular shortcut fails the test outright: pointing at reduced headcount. Gartner surveyed 350 executives at organizations piloting or deploying autonomous systems and found workforce reduction rates nearly equal between the ones reporting strong ROI and the ones reporting modest or negative outcomes. As Gartner’s Helen Poitevin put it: “Workforce reductions may create budget room, but they do not create return.”

What the provable version looks like

Before the next feature ships, four artifacts, in order:

1. A baseline: the metric the feature claims to move (resolution time, cycle time, error rate), measured before rollout.
2. Cost attribution: the feature’s spend tagged and metered per feature and per task, so the denominator exists.
3. An eval suite: quality scored against the baseline, so speed gains can’t hide accuracy losses.
4. A unit number: cost per resolved ticket, per processed claim, per generated draft — one number a CFO can trend quarter over quarter.

None of this requires new vendors or a data-science team. It requires deciding, before launch, that the feature will be judged, and building the two weeks of instrumentation that makes judgment possible. Teams that do this stop having the ROI argument, because the number is just there, on a dashboard, moving or not moving.

If your AI spend has crossed the threshold where the board wants receipts and nobody can produce them, that’s the AI cost optimization engagement: per-feature attribution, a measured baseline, and savings quantified before any change ships.