Two years into the corporate AI spending boom, the people who sign the cheques are delivering a blunt verdict. Survey findings circulating this week suggest that 83% of chief financial officers say most of their organisations’ AI projects have not produced measurable results. For readers of our business coverage, the number lands like a bucket of cold water: billions spent on pilots, licences and consultants, yet the finance chief — the executive most allergic to unquantified enthusiasm — cannot find the return.
Pilot purgatory is real
The pattern repeats across industries. Companies launch dozens of proofs of concept, each demoed to applause, and almost none graduate into production workflows where value is actually created. Data is scattered across systems that do not talk to each other; processes were never redesigned, so AI is bolted onto the same slow approvals it was meant to replace; and success metrics were never agreed before the money went out. When the CFO’s team asks what changed — cost per invoice, claims handled per hour, churn avoided — the honest answer is often a shrug and a user-satisfaction score.
It is a management problem, not a model problem
Consultancies that study adoption, including McKinsey & Company and Deloitte, keep reaching the same conclusion: the technology works best where leadership treats it as an operating-model change rather than a software purchase. The successful minority pick a small number of high-volume, measurable processes, assign a business owner who is accountable for the number, rebuild the workflow around the tool, and track baseline versus outcome from day one. They also invest in unglamorous foundations — clean, accessible data and staff training — that rarely feature in vendor keynotes.
What CFOs are doing about it
Finance chiefs are responding with tougher gates: funding released in stages tied to measured outcomes, AI inventories that kill zombie pilots, and a preference for use cases with hard savings — invoice processing, customer-service deflection, code assistance with tracked velocity — over open-ended experimentation. Some are discovering returns in unexpected places, such as working-capital improvements from better demand forecasting, which only appear in the numbers quarters later. The 83% figure may therefore soften as measurement catches up with deployment.
The measurement gap behind the headline
Part of the problem is that AI’s benefits often arrive in forms finance finds awkward to book. Minutes saved across thousands of employees are real but do not automatically become cash unless roles, overtime or vendor spend actually change. Quality gains — fewer errors, faster responses, better forecasts — show up in other departments’ numbers, leaving the AI budget line holding only costs. Sophisticated adopters are fixing the accounting by agreeing in advance how a benefit will be recognised: which cost line falls, which revenue line grows, and over what period. They also separate experimentation budgets from transformation budgets, so early exploration is not judged by production standards. As those disciplines spread, expect CFO sentiment to improve — not because the technology suddenly changes, but because the scoreboard finally reflects the game being played.
The bottom line
The CFO verdict is not that AI is useless; it is that unmeasured AI is unfundable. Companies that define the number they intend to move before they buy the tool are finding real returns, while those that bought first and justified later are supplying the statistics. Expect the next year of enterprise AI to be quieter, narrower — and considerably more productive.