Ask a chief executive what their firm spent on AI last year and you may not get an answer.
A Gartner survey published on 1 September, covering 1,303 respondents at organisations with at least $50 million in annual revenue, found that 11 per cent of organisations could not say what their function spent on AI in 2025. Roughly one in nine. Not "spent too much" or "spent unwisely", but genuinely could not produce the number.
That would matter less if spending were flat. It is not. The same survey found 85 per cent of functional leaders plan to increase AI spending in 2026, on top of an average 12 per cent of functional budgets already committed to AI in 2025. Tina Nunno, a Gartner Fellow, put it plainly: the lack of financial visibility heightens the risk as spending accelerates.
The scaling number is worse than the pilot number
For two years the headline statistic has been about pilots. MIT's Project NANDA study last year found that 95 per cent of enterprise generative AI pilots produced no measurable profit-and-loss impact. That number is still quoted in most board packs.
The Gartner data moves the problem one step along. It found that only 22 per cent of organisations have successfully scaled AI across multiple business units or adopted an AI-first approach. The other 78 per cent are not stuck at the pilot gate. Many are past it, running AI in one function, spending real money, and unable to repeat the result anywhere else.
A pilot that fails is cheap and instructive. A deployment that works in one place, cannot be extended, and is not being measured is the expensive version of the same problem. It looks like success on a slide.
What the 22 per cent do differently
The Gartner analysis separates high performers from the rest, and the distinction is not about model choice or vendor. High performers reported positive returns on 81 per cent of their AI initiatives. Low performers could not even state the return on 29 per cent of theirs.
The high performers share three habits. They track the return on AI initiatives continuously, not once at business-case sign-off. They treat AI as a portfolio of investments rather than a set of standalone projects. And they reallocate or shut down the initiatives that are underperforming, rather than letting them run because someone senior sponsored them.
None of that is novel. It is how a competent finance function has always treated capital projects. What is striking is how rarely it is applied to AI, where the combination of vendor pressure, executive enthusiasm, and genuine technical uncertainty tends to suspend the normal rules.
AI spend has become its own discipline
The wider market has noticed. The FinOps Foundation's State of FinOps 2026 survey found that 98 per cent of practitioner teams now manage AI spend in some form, and that managing the cost of AI is the field's top forward-looking priority. A cost category that barely registered two years ago is now the thing cloud financial teams say they most need to get right.
The reason is that AI spend behaves differently from a typical software line item. Inference costs scale with usage, not with seats. A successful rollout can double its own bill. Costs sit across model providers, cloud infrastructure, data pipelines, integration work, and the people who keep the thing running, so no single invoice tells the whole story. Without deliberate tracking, the number is genuinely hard to assemble, which is how you end up as one of the 11 per cent.
What to put in place
This is not a call for another governance committee. It is a short list of things that need an owner.
One consolidated view of AI cost
Pull model, infrastructure, data, integration, and run-cost into a single figure per initiative, refreshed monthly. If nobody owns this, the answer to the chief executive's question will stay "we're not sure".
A return figure that is revisited
Every AI initiative should carry a benefit number that someone checks against reality on a set cadence. Not the business-case projection from sign-off, the actual measured effect. Where the effect cannot be measured, that is itself the finding.
A portfolio review with the power to stop things
A quarterly session that looks at every AI initiative together, ranks them by return, and moves money away from the laggards. The review only works if it is allowed to cancel a sponsored project. If it cannot, it is a reporting exercise.
A scaling test before the second deployment
Before extending an AI capability into a second function, confirm that the first one is delivering a measured return and that the integration, data, and operating model will carry over. Scaling a result you have not verified just multiplies the cost.
The bottom line
The firms getting a return on AI are not the ones with the best models. They are the ones treating AI spend the way they treat any other material investment: visible, measured, and subject to being stopped.
Spending is going up in almost every organisation this year. The question the Gartner data raises is not whether to invest more. It is whether you can currently say what the last round bought you. If the answer is no, that is the first thing to fix, and it costs very little to fix it.