Two years into the corporate artificial intelligence boom, the people who sign the cheques are delivering a sobering verdict. Survey evidence from finance chiefs suggests that around 83 per cent of CFOs believe most of their organisation’s AI projects have yet to produce measurable results. The spending, by any measure, has been enormous. The returns, by the finance department’s own accounting, largely have not arrived.
That gap between investment and evidence is now the central question in boardrooms, and it runs through much of our business coverage. Licences, consultants, data platforms and pilot programmes all cleared approval on the promise of productivity gains. What CFOs are reporting is not that AI does nothing, but that what it does is remarkably hard to pin to a number anyone can audit.
Why the Results Are So Hard to Measure
Part of the problem is definitional. A chatbot that drafts emails saves minutes that never appear in any ledger. A forecasting model may improve accuracy by two points, a gain that is real but entangled with everything else the market did that quarter. Finance teams are trained to ask what changed because money was spent, and AI projects are often approved without a baseline against which that question can be answered.
Deployment patterns make it worse. Many companies scattered small pilots across departments — a copilot here, a document summariser there — with no owner responsible for the overall outcome. Each pilot produced anecdotes, and anecdotes do not survive a budget review. Research from McKinsey on the state of AI has repeatedly found that the minority of firms capturing real value are those that redesign workflows around the technology rather than sprinkling it over existing ones.
What the Successful Minority Do Differently
The companies CFOs describe as successes share habits that have little to do with model choice. They pick a small number of use cases tied to a cost line or a revenue line that already exists. They measure before they deploy. They assign a business owner, not just a technical sponsor, and they are willing to stop projects that miss their own targets — a discipline that sounds obvious and is apparently rare.
They also invest in the unglamorous substrate: clean data, integrated systems and staff training. An AI assistant connected to fragmented, contradictory data produces fluent, contradictory answers. Several finance chiefs now describe data remediation, not model licences, as the true cost of AI — a line item few original business cases included.
A Reckoning, Not a Retreat
None of this suggests the spending is about to stop. Most CFOs in these surveys still expect AI budgets to grow, and fear of falling behind competitors remains a powerful motivator. What is changing is the burden of proof. Projects that once won funding with a demonstration and a promise are now being asked for a metric, a baseline and a date.
That shift is healthy, and it mirrors other technology cycles our readers can follow on our homepage. Cloud computing and earlier automation waves both passed through a phase where enthusiasm outran accounting before the real productivity arrived. The difference this time is speed: the spending reached extraordinary scale before the measurement culture caught up.
The Question for Next Year’s Budget
For boards, the practical test is now straightforward. Can the team name the number the project was supposed to move, show the number before, and show it after? Where the answer is yes, AI spending tends to survive scrutiny and grow. Where the answer is a shrug and a slide deck, CFOs are learning to say a word that was missing from the first wave of the boom: no — or at least, not yet. Industry trackers such as Gartner’s technology research expect exactly this sorting process to define the next phase, separating durable value from expensive experimentation.
