Every finance chief in the world is currently being asked the same question by their board: we spent the money on AI — where is the return? A new survey suggests the honest answer, at most companies, is uncomfortable. Some 83% of chief financial officers say that most of their organisation’s AI projects have not produced measurable results. The technology works in demonstrations; in the ledger, it is still mostly a cost line.
Why Enterprise AI Stalls
The failure pattern is remarkably consistent across industries. Projects begin as technology initiatives rather than business problems: a model is procured, a pilot is launched in a tolerant department, and success is declared on engagement metrics rather than financial ones. Meanwhile the unglamorous prerequisites — clean data, redesigned workflows, staff whose incentives change with the tool — lag behind. CFOs, whose professional instinct is to ask “compared to what, and over what period,” are simply the first executives forced to quantify the gap. Our AI and data analysis has tracked this pattern from exuberance to audit.
What the Successful Minority Does Differently
The projects that do produce measurable results share a recognisable anatomy. They start from a costed pain point — invoice processing time, forecast error, churn — with a baseline measured before any model is switched on. They are owned by the business unit that keeps the savings, not by an innovation lab. And they are scoped tightly enough to reach production in a quarter, generating evidence while larger ambitions are still in design. Industry research from firms such as Gartner echoes the point: AI value correlates less with model sophistication than with deployment discipline.
The Measurement Problem Is Real Too
Fairness requires one caveat: some AI returns are genuinely hard to isolate. A support tool that prevents escalations, or a forecasting model that avoids one bad quarter, produces value as absence — invisible in the accounts. CFOs themselves acknowledge that measurement frameworks built for machinery fit software intelligence poorly. The mature response is not to stop measuring but to measure earlier and smaller: instrument the workflow, not just the outcome.
A CFO’s Checklist for the Next AI Proposal
Finance chiefs who have lived through this cycle offer a consistent interrogation for any new AI proposal. What is the current cost of the process, measured, not estimated? Which specific line — headcount hours, error rates, cycle time, revenue — will move, and by how much would the project need to move it to pay back? Who owns that number after the vendor leaves? What data does the model need, and is that data actually clean today? And what is the exit cost if it disappoints?
Proposals that answer all five crisply tend to be the ones that later appear in the successful minority of surveys like this one. Proposals that answer with adjectives tend to join the 83%. The checklist is not hostility to AI — several of its loudest advocates in the C-suite are CFOs whose early, disciplined projects paid for the company’s entire programme. It is simply the recognition that in a business, an unmeasured benefit and an absent one are indistinguishable, and budgets follow the distinction.
What Happens Next
The survey is best read as a sorting mechanism. Budgets will migrate from scattered experiments toward the minority of programmes with baselines, owners and evidence — and vendors will be held to contractual outcomes rather than demonstrations. AI in the enterprise is not failing; it is being audited, which is what technologies undergo right before they become infrastructure. The 83% figure will fall when the discipline rises. More business technology coverage is on our homepage.
