McKinsey’s latest global AI survey reads like a success story until it reaches the profit line. Eighty percent of respondents say AI has improved their individual productivity, and half say it helps them make better decisions (McKinsey Global Survey on the state of AI, Aug 2026). Adoption is still climbing: 44% of organisations now report AI scaling across the enterprise, up from 38% a year earlier. Then the success story stops. The share of respondents attributing any EBIT impact to AI is 37% — essentially unchanged from the year before.
One number moved and the other did not. That is the finding worth sitting with. More organisations are deploying more AI in more functions, and the share of respondents attributing it to operating profit is essentially unchanged. McKinsey’s own summary is blunt: organisations’ conviction in AI is growing faster than the immediate financial returns they can attribute to it.
The gap is not an adoption problem
The stalled 37% would be easy to read as a technology shortfall. The survey suggests something narrower. Individual gains are real but diffuse. Minutes saved across a thousand desks do not roll up into a P&L line by themselves; between the personal gain and the profit claim sit a redesigned workflow, a measured outcome, and someone accountable for the number. Most organisations have not built that middle.
The small group that has built it looks different in specific, checkable ways. McKinsey defines AI high performers as respondents who attribute at least 5% of EBIT to AI and describe its impact as significant — about 6% of the sample, a share that has not moved either. Nearly three quarters of that group report fundamentally redesigning workflows because of their AI use, against roughly a quarter of everyone else, and they are about twice as likely to report that their organisations have defined processes to measure the impact of their AI initiatives (QuantumBlack, AI by McKinsey, Aug 2026). Those two differences are practices, not purchases.
A survey answer is not a measurement
There is a second reading of the 37%, and it cuts the other way. The survey measures what respondents say, not what their ledgers show. An organisation can believe AI added to EBIT without holding a record that would survive a finance review — and it can have real impact it never measured and so does not claim. Both failure modes point at the same missing artifact: a record that ties a specific deployment to a specific measured outcome.
This is why the flat line matters more than its level. If attribution were mostly a measurement backlog being worked through, the share attributing impact would grow year over year as records accumulated. It is essentially flat. The evidence infrastructure that would let an organisation answer the question is not being built at anything like the pace the tools are being deployed.
The decision to make now
Pick one workflow where AI is already claimed to help and run the conversion the high performers run. Redesign the workflow around the capability rather than layering the tool onto the old process. Record the baseline before the change. Measure the delta where finance can see it — cycle time into cost, error rate into rework, hours into either capacity or cash, each kept in its own category. Name the owner who will defend the number.
Then hold the categories apart. A personal productivity report is testimony. A measured workflow delta is evidence. An EBIT attribution is a claim built on that evidence, and it should carry its method with it. Organisations that blur these three can report a feeling. Organisations that keep them separate can defend a number — to a board, a CFO, or an auditor who asks how they know.
What would count as proof?
- One named workflow with a dated description of how it ran before AI and how it runs now, specific enough that a reviewer could see what changed.
- A baseline measured before the redesign — cycle time, unit cost, error rate, or throughput — with the measurement method stated.
- The post-change measurement of the same quantity, over a stated period, with the same method, and the delta computed from the two.
- A conversion step that states how the operating delta becomes a financial figure, and which category it lands in — cash, capacity, or structural value — without mixing them.
- A named owner who signed the figure and would restate it under questioning.
- For any enterprise-level claim, the roll-up showing which workflow records sum into it — and which reported gains were left out because they had no record.
The test is the one a survey cannot administer: could the organisation show a stranger how it knows? Answering it takes a record, not a recollection.
What remains unclaimed?
The survey establishes what respondents report, across a sample of 1,719 participants in 97 nations weighted by each nation’s contribution to global GDP. It does not establish that AI added no profit at the other 63% of organisations — only that those respondents do not attribute any. It does not establish that redesign and measurement cause the high performers’ results; the practices and the outcomes are observed together, and the direction is not proven. And an attribution given to a surveyor is not a measured result: nothing in the survey shows how many of the 37% could produce the record behind their answer.
Nothing here says the individual gains are illusory. Eight in ten people reporting that a tool improves their work is a strong signal about the tool. It is simply not a financial statement — and treating it as one is how an AI programme ends up defending a feeling with a dashboard.
Build the middle of the chain
Adoption is no longer the differentiator; the survey shows everyone adopting. The differentiator is the unglamorous middle of the chain — the redesigned workflow, the measured outcome, the named owner — that turns a personal gain into a profit line an organisation can stand behind.