Enterprises spent roughly $37 billion on generative AI in 2025 — about three times the prior year, by Menlo Ventures' count. In the same window, MIT's NANDA project reported that 95% of enterprise generative-AI initiatives had produced no measurable profit-and-loss impact, and McKinsey found that fewer than four in ten organizations could attribute any EBIT impact to AI — most of them under 5%. The spending is real. The proof is not.
The comfortable reading is that the technology underdelivered. The evidence points somewhere less flattering to management: most companies are generating value they have not instrumented, and asking the wrong seat to certify it on the wrong clock. AI ROI is failing as a measurement problem before it fails as a technology one.
AI ROI is the financial return a company can defend — incremental revenue, margin, cost or risk reduction, net of the full cost of the system — attributable to an AI deployment. The defensible part is where it breaks. Two-thirds of the reason is a chain most organizations only measure at the ends. The other third is that nobody owns the middle.
Usage is not financial value, and the gap between them has four links
Ask most teams whether their AI is working and you will hear one of two answers: adoption is high, or people are saving hours. Both are true and neither is ROI. Between “people are using it” and “the business is measurably better off” sit four handoffs, and value can leak at every one.
- 01Usage — are people using the tool? Almost always yes: in Open Future Forum's largest finance room, 71% already run an AI tool.
- 02Workflow effect — does the work get faster or better?
- 03Operating effect — did a metric the business runs on actually move: cycle time, error rate, conversion, throughput?
- 04Financial effect — did revenue, gross margin, operating expense or cash change?
- 05Attribution — how much of that change is the AI, versus price, seasonality or a reorganization?
Enterprises are fluent at the first link and assert the last. They rarely instrument the middle. That is why a company can be genuinely more productive and still fail an ROI review: it measured usage and time saved, skipped operating and financial effect, and had no attribution model when finance asked what changed.
Where the value leaks: saved hours that never become cash
The most expensive leak is the handoff from workflow to operating effect. Time saved is not money saved until the organization does something with the freed capacity. If an analyst reclaims five hours a week but headcount, throughput and revenue are unchanged and the hours are absorbed as slack, the financial value realized is close to zero. Saved capacity converts to ROI in only three ways: the cost comes out, the capacity is redeployed to work with measurable value, or output rises against steady cost. Absent one of those, “productivity” is a real experience for the employee and an unbooked number for the CFO.
This is the distinction that separates the deployments that get renewed from the ones quietly switched off. It also explains the abandonment data: S&P Global Market Intelligence found the share of firms scrapping most of their AI initiatives rose to 42% in 2025 from 17% a year earlier, with the average organization scrapping 46% of its proofs of concept before production. When Gartner predicted in 2024 that 30% of generative-AI projects would be abandoned after proof of concept, it named unclear business value as a lead cause. A pilot that never measured the operating link has nothing to defend when the budget review arrives.
The proof problem is organizational: the signer and the prover are 27 points apart
Here the picture stops being a spreadsheet problem and becomes an org-chart one. In Open Future Forum's August 2026 finance data, 72% of CEO and founder respondents expect a measurable return on an AI investment inside six months. Among finance-seat respondents, the figure is 45% — a 27-point optimism gap between the seat that most often signs for AI and the seat that has to prove it. Finance answers “not sure” about four times as often as the CEO seat. The optimism sits with the buyer; the doubt, and the burden of proof, sits with the measurer.
That gap has consequences a generic ROI framework cannot address, because it is not a modelling error — it is two functions holding different beliefs about the same purchase. The signer commits capital against a six-month expectation. The prover inherits a number that, on the evidence, usually takes longer and is harder to isolate. When the six-month mark arrives and the operating link was never instrumented, the result reads as failure to the person who approved it and as “we never set this up to be measured” to the person now asked about it.
Nobody owns the number — and the clock is stretching
The gap widens because ownership is thinning. In the same data, business-unit sign-off on AI purchases fell from 18% to 8% inside a single month, while “no single owner yet” doubled to 14%. Tellingly, not one CEO or founder respondent reported an unowned AI purchase at their company; one level down, 29% of technology respondents did. Seen from the top, someone owns it. Seen from where the work happens, increasingly no one does. A number owned by nobody does not get produced.
The clock is moving too. Across editions of the CFO AI Leverage Report, the expectation of a return inside six months eased from 60% to 54%, and “not sure” rose from 11% to 19%. Proving ROI has remained the single most-named blocker throughout, at 51%. The market is not rejecting AI — deployment is running ahead of the budget line. It is struggling to convert deployment into a defensible number inside the window the signer assumed.
What a defensible AI ROI case actually contains
The companies keeping their AI budgets are not the ones with better models. They are the ones that instrumented the two links everyone else skips and assigned the proof to one seat.
A CFO-grade case does four things. It separates the kinds of value — revenue, margin, cost reduction, cost avoidance, working-capital and risk reduction are not one number and should not share a formula. It counts the full cost — licenses, inference, integration, data preparation, security, governance, retraining, workflow redesign, monitoring and the pilots that failed — not just the subscription. It captures a baseline before deployment — cycle time, error rate, cost per unit, throughput — so the operating link can be measured rather than asserted. And it states an attribution stance — a control, a holdout, a pre/post comparison, or an explicit assumption — so the financial change can be defended when challenged.
On timing, the honest answer to “how long should an AI investment get” is: long enough to show movement in the operating metric, not the financial one. Cycle time and error rate move first; margin and cash follow. A program with no operating movement at six months is a genuine warning. A program with clear operating movement and no booked financial result yet is usually a conversion problem — capacity not removed or redeployed — and that is a management decision, not a technology verdict.
The conclusion executives should draw
The “AI doesn't pay” narrative and the “here's a five-metric dashboard” narrative are both incomplete. The first mistakes an unmeasured return for an absent one. The second hands a measurement template to an organization whose real problem is that the optimistic seat and the accountable seat never agreed on what would count, or when.
So the work in 2026 is smaller and harder than another framework. Name one seat that owns the AI ROI number — in most companies the CFO's office, which is where the sign-off is already migrating. Instrument the operating link before deployment, not after. Set the return clock to the metric that actually moves first. And close the 27-point gap in the room, before the purchase, by making the signer and the prover agree on the evidence that will settle it. Companies that do this will not suddenly get more from AI. They will finally be able to prove what they were already getting — which, on the current evidence, is most of the problem.
Last updated: August 19, 2026
Frequently Asked Questions
Compare your AI ROI model with your peers
Open Future Forum convenes finance leaders in the CFO Executive Forum and publishes the CFO AI Leverage Report on how enterprises actually budget for, sign off on and measure AI. Global, founded in Silicon Valley.