Firms tripled what they spend on AI. Most still can't prove it worked.

Enterprises spent about US$37 billion on generative AI in 2025, roughly 3.2 times the US$11.5 billion spent in 2024 (Source: Menlo Ventures, 2025: The State of Generative AI in the Enterprise). That is one of the fastest climbs of any software category.

The costs are real, the proof is not#Link to this section

The returns tell a quieter story. In McKinsey's State of AI in 2026, about 37% of respondents said AI had contributed positively to their organisation's EBIT, essentially flat from a year earlier even as more companies scaled the technology (Source: McKinsey, The state of AI in 2026). Among the ones who report any impact at all, most put it below 5%, and only about 6% qualify as high performers (Source: McKinsey, The state of AI in 2025). Separately, MIT's Project NANDA found that 95% of the generative AI pilots it studied produced no measurable profit and loss impact (Source: MIT Project NANDA, The GenAI Divide: State of AI in Business 2025). And only about 45% of organisations can quantify their AI ROI at all (Source: NVIDIA, State of AI).

The gap sits between the seat that signs and the seat that proves#Link to this section

The Open Future Forum analysis put a sharp number on it: 72% of chief executive and founder respondents expect payback inside six months, against 45% of finance-seat respondents. That is a 27-point gap between the seat that signs the cheque and the seat that has to prove the return (Source: Open Future Forum, Why Most Companies Still Can't Prove AI ROI in 2026).

The same pattern shows up elsewhere. Infosys published a report called The AI ROI Gap, describing the growing disconnect between what leaders expect from AI and what they can actually demonstrate (Source: Infosys, The AI ROI Gap). In banking, US banks now expect AI to cut expenses by about 4.5 to 5% while lifting revenue only 2.5 to 3% (Source: D.A. Davidson survey of US banks). And Deloitte found most organisations reach satisfactory ROI on a typical AI use case in two to four years, far longer than the seven-to-twelve-month payback most technology has historically delivered (Source: Deloitte).

Why the old ROI maths misses it#Link to this section

Traditional ROI frameworks were built for a purchase with a fixed cost and a fixed output. AI costs don't behave that way. They are driven by usage rather than seats, spread across many tools, and often owned by no single function. Value then leaks at the handoffs almost nobody tracks. Saved hours only become cash if the cost is removed, the capacity is redeployed, or output rises against a steady cost (Open Future Forum). A general-purpose copilot rolled out to thousands of people with no defined success metric produces no number at all, because nobody owns one.

What to measure instead#Link to this section

Here is where a leader can actually move the number.

Measure adoption depth, not licences bought. Track weekly active use per tool against who has access. Good looks like sustained use by most of the intended group after the novelty period, not a spike in month one that fades by month three.

Measure saved time that converts to cash. Ask the people doing the work how much time a tool actually saves, then track whether that capacity was redeployed to higher-value work or simply absorbed. Time saved that changes nothing is not a return.

Measure waste and shadow AI. Every tool paid for but rarely used, and every tool used but never approved, is a line you can act on. Good looks like a shrinking gap between what you pay for and what people rely on.

Measure ownership. Assign one named owner to every AI investment and the metric that proves it. When ownership thins, proof disappears with it.

Measure the expectation gap directly. Ask the seat that signs and the seat that proves what payback they expect, and by when. Closing that 27-point gap starts with knowing how wide it is.

None of these studies says what the organisations seeing returns do differently. What they do show is that many organisations can't yet put a figure on the return at all: in NVIDIA's survey, only about 45% could. We built The GAiGE to help you put a figure on yours, using short pulses (anonymous by default, identifiable when you choose) that tell you what your people actually use, what it saves, and what you are wasting.