Chief financial officers are the people paid to ask what a company got for its money, and their verdict on artificial intelligence is becoming uncomfortable reading. In recent survey evidence, 83% of CFOs say most of their organisation’s AI projects have not produced measurable results. The spending, meanwhile, continues to climb — on licences, cloud computing, consultants and data projects — creating a widening gap between the AI story told in boardrooms and the numbers arriving in finance departments.
The finding does not mean the technology does nothing. It means that in most companies, nobody can demonstrate, in the language finance understands, what it did. Pilots multiply, demos impress, and usage statistics get reported as success, while the profit-and-loss account stays unmoved. Coverage of enterprise AI spending by Reuters has tracked the same tension: investment racing ahead of verified returns.
Where the Returns Get Lost
The first leak is measurement itself. Many AI projects begin without a baseline — nobody recorded how long the process took, what it cost, or how often it failed before the model arrived — so afterwards there is nothing honest to compare against. The second leak is scope: a tool that drafts emails or summarises meetings saves minutes in ways that are real but almost impossible to aggregate into money, especially if headcount and workload stay exactly the same.
The third leak is integration. A model that works in a demonstration often sits awkwardly inside real workflows, waiting on data that lives in five systems, approvals that were never redesigned, and staff who were consulted late or never. Finance leaders describe the pattern as pilot purgatory: dozens of experiments, each technically promising, none scaled far enough to move a number that matters. Our business technology reporting sees the same pattern across sectors, from retail forecasting to customer-service automation.
What the Successful Minority Do Differently
The companies reporting genuine returns tend to share habits rather than tools. They pick a small number of expensive, measurable problems — fraud losses, inventory waste, churn — and attach the AI project to an existing budget line with an owner who is accountable for the outcome. They measure before they build. And they redesign the process around the model instead of bolting the model onto the process and hoping.
They are also more honest about timelines. Research from firms such as McKinsey consistently finds that value concentrates in a minority of adopters who invest as much in workflow change, training and data quality as in the models themselves. The model, in other words, is the cheap part; the organisation is the expensive part.
The CFO’s New Role in the AI Boom
None of this suggests CFOs are turning against the technology. Most expect AI to matter enormously; they are simply refusing to fund it on faith indefinitely. Expect tighter stage gates in 2026 budgets: pilots that must name their metric in advance, scale-up funding released only against measured savings or revenue, and quiet cancellations of projects that cannot explain themselves.
That discipline may be exactly what the technology needs. The 83% figure is less an indictment of AI than of how it has been bought — as a symbol of modernity first and an investment second. Companies that reverse that order, problem first and tool second, are finding that the returns were there all along. They were just waiting for someone to measure them.
A Note of Caution on the Survey Itself
One caveat belongs with the headline figure. “No measurable results” is partly a statement about the projects and partly a statement about the measuring. CFOs can only certify returns that someone tracked, and in fast-moving pilots tracking is usually the first casualty. Some of the 83% therefore represents value that may exist but was never evidenced — which, from a finance chair, amounts to much the same problem. A benefit nobody can demonstrate cannot be budgeted, defended or repeated, and teaching project teams to evidence their wins may be the cheapest AI investment most companies have left to make.
