Measuring AI ROI in finance: only 35% say they can

Adoption ran ahead of arithmetic

The Protiviti survey, covering 902 finance leaders, captured a familiar pattern in an unfamiliar timeframe. AI use in financial forecasting rose from 58% to 76% in a single year. That is a step change in a discipline where forecast methodology usually evolves over budget cycles, not quarters. The measurement infrastructure did not move with it.

The strategy numbers are more revealing than the ROI number. Only 14% of finance organizations reported deploying AI under a detailed strategy, and just 7% said their organization prioritizes AI governance over speed of adoption. In other words, 86% of finance teams are putting models into forecasting, close and reporting workflows without a documented plan for what those models are supposed to deliver or how anyone would know.

That combination produces a specific failure mode. Renewals arrive without a baseline. Capitalization questions arrive without a defensible useful-life argument. And board decks arrive with adoption statistics standing in for benefit statistics, because adoption is the only thing anyone measured.

When the seller writes the scorecard

On July 18, CFO Dive reported that OpenAI CFO Sarah Friar had floated a framing for AI returns: useful intelligence per dollar. It is a genuinely interesting construct and it is also, unavoidably, a metric denominated in the seller's units. Palantir's CEO went the other direction, arguing publicly that enterprises have grown tired of token economics and that rising token consumption does not automatically convert into return.

The disagreement is less important than what it signals. Two of the biggest AI vendors are competing to define the buyer's ROI method at precisely the moment the buyer admits it has none. Whoever wins that argument sets the denominator for renewal conversations across the market.

The problem with vendor-supplied arithmetic is not dishonesty. It is unit choice. Tokens, seats, model calls and inference volume are all real and all measurable, and none of them is a financial outcome. A finance function that accepts them as proxies has outsourced the one judgment it exists to make: whether a given dollar of spend produced margin, cycle time, working capital or headcount avoided.

The spend curve raises the cost of guessing

Gartner projects worldwide AI spending will rise 47% in 2026 to $2.6 trillion, up from $1.76 trillion in 2025, with a further climb to $5.62 trillion by 2030. At that trajectory, a measurement gap that was tolerable when AI was a pilot line item becomes a material control weakness when it is a multi-year commitment appearing in capex, opex and vendor concentration disclosures.

Risk is compounding alongside spend. Data security and privacy ranked as finance leaders' top priority for the third consecutive year in the Protiviti survey. Separately, a study by MIT and the University of Queensland canvassing 272 experts named finance among the sectors most exposed to AI-weaponized cyberattacks. Finance is therefore absorbing more AI risk while being least able to price the offsetting benefit.

A finance-owned AI benefit register

The practical remedy is unglamorous and familiar: treat AI like any other capital or subscription commitment. Build a benefit register that finance owns, not IT and not the vendor. Each entry needs five fields. A baseline metric measured before deployment. A named business owner who is accountable for the delta. A measurement window with a defined end date. The financial unit the benefit will be expressed in, whether that is margin points, days in the close, or full-time equivalents avoided. And a kill date if the number has not moved by the end of the window.

Two disciplines make the register credible. First, baselines must be locked before go-live, because retrospective baselines are indistinguishable from narrative. Second, benefits should be netted against total cost including inference, integration, retraining and the human review layer that most finance deployments still require.

Useful intelligence per dollar is not a bad idea. It is an incomplete one until the numerator is defined by the buyer. CFOs who can articulate what useful means in their own operating metrics will negotiate from a position their peers do not have. The 65% who cannot will be renewing on someone else's math.