Most firms still cannot prove their AI is paying off

Nearly every organisation now uses AI, and only about a third can point to any effect on earnings, a share that has barely moved in a year (Source: McKinsey, The State of AI 2026).

Why the proof keeps going missing#Link to this section

Start with the headline number. McKinsey's State of AI 2026 survey of 1,719 respondents found that only 37% attribute any EBIT impact to AI, essentially unchanged from a year earlier, even as more organisations scaled the technology across the enterprise (McKinsey). The share of genuine high performers, those attributing at least 5% of EBIT to AI and describing the impact as significant, stayed flat at about 6% (McKinsey).

Individual experience is not the problem. 80% of respondents said AI improved their own productivity (McKinsey). The money just isn't showing up at the top of the house. The CEO view is starker. PwC's 29th Global CEO Survey of 4,454 chief executives across 95 countries found that 56% saw neither higher revenue nor lower costs from AI over the past year, while only 12% reported both (Source: PwC, 29th Global CEO Survey, 2026). MIT's Project NANDA put the same pattern more bluntly: 95% of the generative AI pilots it studied showed no measurable profit and loss impact (Source: MIT Project NANDA, The GenAI Divide: State of AI in Business 2025).

That 95% is narrower than the headline suggests. It means those pilots produced nothing that reached the P&L within about six months. The uncomfortable part is that most firms can't tell which side of the line they are on, because they never set a baseline to measure against.

The three failure modes to check for#Link to this section

Across these studies the same pattern repeats, and it has three parts.

The first is no baseline. If you didn't record how long a task took, what it cost, or how good the output was before AI arrived, you have nothing to compare against afterwards. "It feels faster" is not a number a CFO can defend.

The second is measuring activity instead of outcomes. Seat counts, licence counts, logins, prompts sent and messages generated all tell you people are busy with AI. Whether the business is better off is the question around them. Adoption is near universal already, so counting it proves very little.

The third is bolt-on deployment. AI gets added on top of existing work without changing the work itself, so individual productivity gains stall before they reach the income statement, because the workflows, metrics and governance were never redesigned to capture them.

How to measure AI return you can defend#Link to this section

Measure the outcome, not the activity. Pick the specific tasks and workflows your AI tools are meant to improve, then track the thing that actually moves money: hours per task, cost per output, error and rework rates, cycle time, and the revenue the work supports. Tie each AI tool to a line someone in finance already trusts.

Set a baseline before you judge anything. Capture the current state for those same tasks now, even roughly. A defensible before and after beats a confident guess every time.

Separate the people signal from the cost signal. Ask the people doing the work two plain questions: is this tool actually saving you time, and would you miss it if it were gone? Short, frequent pulses, anonymous by default, get answers dashboards can't, and they surface tools that are paid for but quietly unused. One in five organisations now say AI operating costs are starting to constrain their use, so finding the waste is money straight back (McKinsey).

Read the result as a ratio, not a story. Good looks like a clear link between what a tool costs and what it returns, a baseline you can point to, and a majority of users who say the tool is worth keeping. If you can't produce those three things for a given tool, you don't yet have a measured return.

McKinsey doesn't say what the top 6% do differently. The studies agree on one thing: a baseline set before rollout is what makes a return provable. The GAiGE runs short pulses, anonymous by default or identifiable when you choose, that turn what your people tell you about AI into numbers you can defend.