The six per cent pulling away from everyone else.
McKinsey asked two thousand executives whether AI had moved their margins. Six per cent could say yes. The gap between them and everyone else is the story.
In the summer of 2025, McKinsey and its data-science arm QuantumBlack put one survey in front of 1,993 senior executives across nineteen industries and forty-seven countries. It was the firm's annual State of AI study, the largest running survey of its kind. By 2025 the question was no longer whether companies used AI — almost all of them did. The question was what separated the firms getting financial impact from the firms getting slide decks.
The survey was published in November 2025. Inside it, McKinsey marked out a subset of firms by two conditions. First, the firm attributed five per cent or more of its EBIT in the previous year directly to AI. Second, its leadership rated the value captured as significant — the top of a five-point scale.
Six per cent of the respondents met both. The other ninety-four did not.
Two cautions sit on that number, and both make it sharper. The criterion is self-reported: a firm decides for itself whether AI drove five points of margin, and grades its own success at the top of the scale. Self-assessment runs optimistic. And the respondents are large enterprises willing to answer a survey about their AI maturity — the population most likely to have something to report. Both biases push the six upward. The real concentration is steeper than six per cent.
What separates the six per cent sits several layers beneath the technology, somewhere the budget and the model size do not reach. Spending did not put them there. The largest budgets, the biggest models and the deepest benches of data scientists do not track with the margin line.
For a few years now, the gulf between AI's winners and everyone else has reached boards as a run of frightening headlines — displacement numbers, exposure numbers, the macro-warnings a board is told to monitor and defend against. Read as separate problems, they generate separate defensive programmes: a workstream each for the labour market, adoption, exposure, output, execution, with a risk owner and a line in the board pack. Laid over one another, they turn out to be five cameras filming the same event from different angles.
The World Economic Forum counts 78 million net new roles created globally by the shift to automated work. Read as a labour-market problem, it is a redeployment exercise. Read structurally, those are high-margin roles orchestrating algorithmic work, and the six per cent are positioning to absorb them.
Gartner projects that only 20 per cent of organisations will deploy generative AI at scale across core operations. The standard response treats twenty per cent as a target to hit with harder project management. The six per cent are not aiming at the twenty. They are the twenty.
The IMF estimates 40 per cent of jobs globally are exposed to AI. Goldman Sachs puts 300 million full-time jobs in the path of automation, and separately estimates the productivity released could lift global output by around seven per cent — on the order of seven trillion dollars — over a decade. And McKinsey's own canonical metric, the corporate transformation success rate, has sat unmoved at roughly 30 per cent for three decades.
In 1981 the economist Sherwin Rosen described how a small edge in talent or reach becomes an enormous gap in reward once a technology lets one performer serve a whole market at once. The microphone let a single singer fill a hall that once needed a dozen. The million-token context window is the microphone. The released value does not spread evenly across the market; it pools where the structure is ready to hold it.
So the five figures are one ledger, not five. The roles created already contain the roles displaced — a churn of more than a fifth of the world's formal jobs. None of the numbers asks whether the churn happens. Each asks who ends up on which side of it.
The defensive reading treats each figure as a separate weather system. One for the labour market, one for adoption, one for exposure, one for output, one for execution. Each gets a workstream, a risk owner, a line in the board pack.
Laid together, they stop looking like weather and start looking like a tide. The market is not fracturing into five operational difficulties. It is concentrating along every axis at once, and the same small group keeps surfacing on the gaining side of each.
Which leaves the one question the five figures do not answer between them. The six per cent are not the biggest spenders or the largest models. So what do they hold that the other ninety-four cannot acquire by writing a larger cheque for the same thing?
What the six per cent have is not a tool. It is a different relationship to the work.
The six per cent stopped transforming and started authoring the work.
The ninety-four are running transformations — taking AI and fitting it onto the existing operating model, automating the tasks the old process already contained, typing faster at the same desks. The six have done something the survey can see in the margin line but not in the spend. They have used the tool to redesign what the work is, rather than to speed up what it was.
That distinction compounds, and it compounds in a way capital cannot shortcut. A firm that rebuilds the workflow around the tool generates proprietary data, a new architecture, a changed division of labour — advantages that accrue to whoever crossed early and cannot be bought back later by spending more on the same unredesigned process. The firm behind is not one budget cycle back. It is on a curve that is bending away from it.
This is the superstar pattern inside a single economy rather than a concert hall. Rosen's singer did not out-sing a thousand local performers by a thousand to one. A small structural edge, amplified by a technology that reached the whole market at once, produced a reward gap that looked absurd against the underlying difference in talent. The context window reaches the whole market at once. The structural edge is small. The reward gap will not be.
The reason this stays half-hidden for a while is that the data surfaces it slowly. The front-runners of an earlier era — the firms that won by editing the founding assumption of their industries rather than out-competing on efficiency — were misread for decades, because their advantage was built into physical supply chains that took years to show. The six per cent will be misread for a shorter time. The advantage compounds quarterly, the surveys land annually, and analysts have begun, for the first time, to read the technology and the margin together.
For the executive, the figure reframes the question on the table. How much to spend on AI, and how fast to deploy it, are the ninety-four per cent's questions — and the ninety-four per cent's number has not moved. The live question is whether the spend buys a faster version of the existing work, or the redesign of the work itself. The first keeps a firm inside the band the survey measures. The second is what the six did.
The six per cent stopped transforming and started authoring the work. The margin line is the first place it shows.
The final section — the reframe — is reserved for members. Enter your email to continue reading.