McKinsey has run a global survey on AI adoption for several years. The 2025 edition was in the field from 25 June to 29 July 2025 and drew 1,993 responses across 105 countries.

The two findings that belong together

Adoption is close to universal. Eighty-eight percent of respondents said their organisation uses AI, up ten percentage points on the previous year.

Attribution is not. Most respondents who reported any earnings impact from AI put it below 5 percent of earnings before interest and taxes. Of 25 organisational attributes tested, the one with the largest effect on whether a firm saw earnings impact from generative AI was the redesign of workflows.

Put together: nearly everyone has the technology, few can point to the money, and the thing that separates them is whether the work itself was rebuilt around it.

Why this is the predictable result

Adding a capable tool to an unchanged process speeds up the steps that were already there. If the process contains a step that should not exist, the tool performs that step faster, and the saving is consumed inside the same process rather than appearing at the bottom of it.

Redesign is harder for reasons that have nothing to do with technology. It means deciding which steps stop, which roles change, and who owns the judgment that remains. Those are decisions with people attached, and a tool purchase is a way of appearing to act without taking them.

What we take from it

This is the clearest external support we have found for running an allocation rule rather than an adoption programme. The question worth answering is not which tools to deploy. It is which work in your commercial system is production, which is judgment, and where the boundary between them sits.

The survey also found that tracking well defined indicators for generative AI work had a strong effect on bottom line impact, which is the same finding from the other side. A firm that cannot say which number a system was supposed to move cannot tell whether it moved it, and will keep buying.

The limitation worth stating is that this is a self-reported survey of mostly large organisations, and attribution of earnings impact to any single cause is the weakest kind of measurement there is. We read it as strong evidence about where the difficulty lies and weak evidence about the size of the prize.