Corporate spending on artificial intelligence has reached historic scale, but the people who sign the cheques are delivering a blunt verdict: according to survey evidence, some 83% of chief financial officers say most of their AI projects have not produced measurable results. The finding captures the central tension of the enterprise AI boom — a technology executives believe they cannot afford to skip, whose returns they mostly cannot yet prove.
The gap is not necessarily a sign that AI does not work. It is, first, a measurement problem. Many AI projects were launched as experiments — pilots in customer service, document processing, coding assistance — without baseline metrics, defined success criteria or a finance partner in the room. When the CFO’s office later asks what the company got for its money, the honest answer is often a collection of anecdotes: hours ‘saved’ that were never converted into output, satisfaction scores, demos. As boards tighten scrutiny and AI budgets grow from experimental line items into serious capital, anecdotes stop being enough. Analysis of enterprise technology spending trends is published by firms such as Gartner.
Where the Returns Get Stuck
Finance chiefs and consultants point to recurring failure patterns. The first is pilot purgatory: dozens of small proofs of concept, none scaled, each consuming integration effort the business case never counted. The second is the last-mile problem — a model that performs impressively in a demo meets messy real data, legacy systems, compliance review and staff who were never trained or incentivised to change how they work. The technology spend is visible in the budget; the organisational spend required to make it pay is not, and is routinely underestimated.
The third pattern is the productivity paradox in miniature. AI tools frequently make individuals faster at tasks that were never the bottleneck, so the saved time dissolves into the working day instead of appearing as lower cost or higher output. CFOs, whose job is precisely to distinguish activity from results, are the first to notice. Our business coverage has reported the same scepticism cycle in previous technology waves — cloud migration and big data programmes both went through a phase where belief ran years ahead of accounting.
What the Successful Minority Do Differently
The projects that do show measurable returns share habits that look almost boring next to the hype. They start from a costed business problem, not a technology looking for a use. They measure a baseline before deployment and agree in advance what number will move: cost per ticket, claims processing time, conversion rate, error rates. They redesign the workflow around the tool instead of bolting the tool onto the old workflow, and they count total cost honestly — licences, integration, oversight, the humans reviewing the machine’s output.
They also tend to be fewer. The emerging consensus among finance leaders is that a portfolio of five committed, measured deployments beats fifty exploratory ones, both for returns and for organisational learning. AI, in this view, is becoming a normal capital-allocation discipline — which is another way of saying the hype phase is ending.
Why Spending Will Continue Anyway
None of this means budgets are about to collapse. CFOs who report no measurable results still report fear of being left behind, and competitive pressure is itself a financial argument: if AI does reset cost structures in an industry, the firms that waited for certainty will buy their catch-up at the worst price. The likely outcome is not less AI spending but more demanding AI spending — staged funding, kill criteria for underperforming projects, and benefits that must be signed for by a business owner, not just demoed by a vendor.
That is a healthy maturation, even if it makes uncomfortable reading for the boom’s loudest promoters. Eighty-three percent is not an indictment of the technology; it is a measurement of how far deployment ran ahead of discipline. The companies that close that gap — quietly, with baselines and business cases — are the ones whose AI results will eventually show up exactly where CFOs look for them: in the numbers. Broader research on AI adoption economics is available from McKinsey & Company.
