The people who sign the cheques have delivered their verdict. A survey of chief financial officers finds that 83 percent say most of their organisations’ AI projects have not produced measurable results, a figure that captures the widening gap between the technology’s promise and its performance on the only scoreboard finance departments respect.
That does not mean the spending stops tomorrow; AI budgets remain enormous, and few CFOs believe their competitors are standing still. It means the terms change. The era of funding experimentation on faith is ending, and the era of demanding baselines, owners and payback periods for every deployment has begun.
Where the Returns Get Lost
Finance chiefs point to familiar failure patterns. Pilots are launched without a baseline, so improvement cannot be proven either way. Projects automate fragments of processes while the surrounding workflow stays manual, stranding the savings. Data quality, unglamorous and expensive to fix, quietly caps what models can achieve. And success is often measured in licences deployed or hours ‘saved’ that never convert into actual cost or revenue. Research and advisory firms such as Gartner’s finance practice have been warning that AI value concentrates in organisations that redesign work rather than sprinkle tools over it.
What the Successful Minority Does Differently
The CFOs reporting real returns describe a recognisably different pattern: narrow use cases with hard metrics, such as invoice processing cycle time or forecast accuracy; a senior owner accountable for the number; and integration into the systems staff already use rather than another dashboard to check. They also start with the process, not the model, mapping where work actually happens before deciding what to automate.
The Spending Implication
Expect a barbell market. Undifferentiated ‘AI transformation’ programmes will struggle for renewal, while projects with measurable unit economics get funded aggressively, a dynamic consultants at firms like Deloitte are already reporting in client portfolios. Our business coverage will follow the earnings season for CFO commentary, which has become the most reliable reality check on the AI economy.
An 83 percent dissatisfaction rate sounds like an indictment of the technology. Read closely, it is an indictment of how it was bought. The CFOs are not saying AI does not work. They are saying they intend to make it prove that it does.
The Vendor Reckoning
Software vendors are about to feel the CFO’s new mood directly. The survey suggests a market split forming: generic AI platforms sold on vision will face longer sales cycles and tougher proof requirements, while point solutions with demonstrable unit economics, dollars saved per invoice processed, hours returned per analyst, will command premium pricing. Expect contracts to change shape accordingly, with more pilots structured as paid proofs of value, success fees tied to measured outcomes, and expansion clauses contingent on audited results. The shift mirrors what happened in cloud computing’s maturation, when the conversation moved from migration narratives to unit economics. Vendors who adapted thrived; those selling transformation as a vibe did not. For enterprise buyers, the new discipline is clarifying rather than chilling: it concentrates spending on the deployments that actually work, which is where the technology’s real economic impact was always going to come from. The CFOs are not ending the AI era. They are pricing it correctly.
There is a final irony in the numbers. The function delivering the harshest verdict on AI is also, in most companies, among its most successful users, applying it to reconciliation, forecasting and audit with real returns. The CFOs know the technology works. Their survey answers say only that they are done pretending it works everywhere, automatically, on announcement.
