The Capture Gap: 80% of Your People Got Faster and 6% of Companies Found It in EBIT

Eighty percent of people say AI has made them personally more productive. Thirty-seven percent of organisations can point to any EBIT impact from it at all. Six percent can attribute 5% or more of EBIT to it. Those three numbers come from the same survey of the same 1,719 respondents, and the distance between the first and the last is the most important operating number in enterprise AI right now. Call it the capture gap.

The source is McKinsey's State of AI 2026, published 25 August 2026 — an online survey fielded 4 May to 8 June 2026, 1,719 participants across 97 nations, weighted by each nation's contribution to global GDP. Thirty-six percent of respondents work at organisations above $1 billion in revenue, which means the majority do not, and that matters for what follows.

The headline finding is that adoption keeps climbing and financial impact does not. Forty-four percent of respondents now report AI scaling across the enterprise, up from 38% a year ago. Fifty-six percent use it in three or more business functions, up from 51%. And the share reporting EBIT impact sat still at 37%, while the share of high performers — those attributing at least 5% of EBIT to AI and describing the impact as significant — stayed flat at 6%.

Price the gap on a real company

Abstractions do not change budgets. Run it on a mid-market business.

Take a 220-person company doing $28 million in revenue, with fully loaded payroll around $19.8 million — call it $90,000 per head, all-in. Apply McKinsey's finding directly: 80% of those people, 176 of them, report that AI has made them more productive.

Now assign the smallest defensible number to "more productive." Not the vendor claim. Three hours a week — thirty-six minutes a day. Across 176 people and 46 working weeks, that is 24,288 hours. At roughly 1,900 productive hours per full-time year, it is 12.8 FTE-years of capacity. Priced at the same $90,000 fully loaded, it is about $1.15 million a year.

Then apply the second finding. Sixty-three percent of organisations report no EBIT contribution from AI whatsoever. In that majority case, all $1.15 million of created capacity is real — the people are not lying about being faster — and none of it appears in the P&L.

This is a directional calculation and we will say so plainly. It combines McKinsey's self-reported individual productivity finding with an assumed hours-saved figure and a standard fully-loaded cost. Self-reported productivity is generous by nature. The point is not the decimal place. The point is that the number is measured in seven figures for a mid-market company, and that for most companies it goes somewhere other than earnings.

Where the capacity actually leaks

McKinsey's own comparison of high performers against everyone else identifies the leaks with unusual precision. There are three, and each has a specific owner.

Leak one: the unclaimed hour. Time saved with no decision attached to it gets reabsorbed. Meetings expand, response times shorten by amounts nobody bills for, and the thirty-six minutes a day disappear into the general texture of the working week. Nothing in this is anyone's fault. Capacity does not convert itself; someone has to decide whether the recovered hours become more output, fewer contractors, faster cycle time, or a role you no longer need to backfill. Absent that decision, the default is reabsorption.

Leak two: the unredesigned workflow. This is the largest one and McKinsey quantifies it. Nearly three-quarters of high performers report fundamentally redesigning workflows because of their AI use, up from 55% last year. Just one-quarter of everyone else does. A three-to-one gap. Inserting AI into an existing process makes individuals faster at the steps of a process that was designed around human throughput. It does not remove the steps, the handoffs, or the queues — and the handoffs are where the time was. Personal speed rises. Cycle time does not.

Leak three: the unowned number. High performers are twice as likely as others to say their senior leaders demonstrate commitment to AI initiatives, and twice as likely to have defined processes to measure the impact of those initiatives. Those two variables travel together for an obvious reason: measurement without authority produces a report, and authority without measurement produces enthusiasm. Together they produce a number somebody has to answer for.

Individual productivity is created by the tool. Enterprise productivity is created by the decision about what the saved time becomes.

The gap is widening for smaller companies, not closing

Here is the finding that should concern anyone running a company below $1 billion in revenue, and it is buried in the exhibits rather than the headline.

Among large organisations — above $1 billion in annual revenue — the share scaling AI agents in one or more functions rose from 27% to 40% in a single year. Among smaller organisations, it was essentially flat at 22%.

Do the arithmetic. The gap between large and small organisations on agent scaling went from 5 percentage points to 18 percentage points in one year — it widened by a factor of roughly 3.6. On enterprise-wide AI scaling generally, 54% of large organisations report it against roughly a third of smaller ones.

And note what "smaller" means in this cut: everything under $1 billion in revenue. That single bucket contains the $900 million company with a data team and the $25 million company without one. The true position of an SMB or lower-mid-market business is almost certainly worse than the 22% figure, because the 22% is being propped up by companies ten to forty times its size. We made the same denominator argument about headline adoption statistics in 18% or 78%? — the number you read depends entirely on who is being counted.

The convenient explanation is budget. The data does not support it. Twenty-eight percent of all respondents now put more than 10% of their entire ICT budget into AI, and 60% expect to increase AI investment over the next year. Meanwhile one in five organisations reports that AI operating costs — including token costs — are already constraining their use. Money is flowing in and hitting a ceiling on the way. Spending is not the differentiator, and a company that responds to a flat agent-adoption number by increasing its AI budget is buying the leak, not fixing it.

The subtraction that keeps not happening

One more finding, because it corrects a specific and expensive error in how leaders are modelling this.

In last year's survey, 32% of respondents expected AI to reduce their organisation's total headcount over the following year. In this year's survey, 14% report that it actually did — less than half. Two-thirds report little or no AI-related change in total employment. McKinsey checked this against the 552 respondents who completed both surveys and the pattern held.

And yet the expectation has gone up: 39% now expect AI-driven headcount declines in the coming year, against 43% expecting no change.

Leaders have now been wrong about this by a factor of more than two, once, on the record — and have responded by forecasting a larger reduction. If you are building an operating plan in which AI pays for itself through payroll subtraction, the best available evidence says you are making the same forecast that just missed by half. The return in the data does not arrive as reduction. It arrives as redesign, and only for the quarter of companies that did the redesigning.

There is a second-order finding worth noting here: among mid-level managers and individual contributors, 47% report experiencing AI-related strain, against 31% of executives and senior managers. The layer being asked to absorb the change is reporting the friction, and it is the same layer we described disappearing in The Missing Middle Rung. Redesigning workflows is work that lands on middle management. Companies have spent a decade removing middle management.

What the 6% did differently

McKinsey is explicit that the high performers are not distinguished by technology. They are distinguished by four things, all of them organisational:

They set a growth objective, not only an efficiency one. About 80% of both groups pursue efficiency. Most high performers also pursue growth or innovation. Efficiency-only mandates cap the ceiling at cost savings, which is exactly the category that leaks back.

They redesigned the work. Nearly three-quarters, against one-quarter of everyone else.

They put a named senior leader behind it. Twice as likely to report demonstrated senior-leader commitment, and twice as likely to have a defined measurement process.

They intend transformation, not augmentation. 3.3 times more likely to say they intend to fundamentally transform the business with AI within three years.

Every one of those is a leadership variable. None of them is a procurement variable. Which is why the honest answer to "why isn't our AI spend showing up" is usually not a model, a vendor or a platform — it is that no one person owns the number, and the workflow was never opened up. That is the same conclusion we reached from a different dataset in Five Questions That Decide Whether Your AI Spend Becomes Revenue, and it is the structural reason 88% of AI pilots never reach production.

The capture gap is not evidence that AI does not work. Eighty percent of people saying they are more productive is a large, consistent, and repeatedly replicated finding. The gap is evidence that most organisations have built no mechanism to convert individual productivity into enterprise performance — and that in the absence of such a mechanism, the productivity is real, unbanked, and quietly reabsorbed.

For a company under $1 billion in revenue, the honest read of this data is uncomfortable. Your larger competitors moved 13 points on agent scaling this year. You moved zero. The difference is not their budget, because one in five organisations of every size is already hitting cost ceilings. The difference is that they have someone whose job is the redesign — and the mid-market's version of that person is usually not a hire it can justify at $350,000 to $450,000 in first-year comp. It is a fractional Chief AI Officer with a defined mandate and a number to answer for, or a scoped diagnostic that establishes the baseline before anything else is bought.

Either way, the first move is the same, and it is not technical. Decide what the recovered hour becomes. Until somebody writes that down, the $1.15 million stays exactly where it is: real, distributed across 176 people, and invisible to the P&L.