Everybody got faster. The ledger didn't notice.

McKinsey published its 2026 State of AI survey this week — 1,719 business leaders and professionals worldwide. Three numbers from it sit next to each other awkwardly, and the awkwardness is the whole story.

80% of the people using AI say their personal productivity improved. Among companies above a billion in revenue, the share scaling AI agents jumped from 27% to 40% in a year. And the share of organizations attributing any EBIT impact to AI held flat at 37%. The genuinely high performers — the ones who can trace more than a slice of earnings to it — stayed at about 6%. Same as last year.

So the tools got better, more people used them, more agents shipped, and the earnings line stayed where it was. That is not a story about disappointing models. The models did their job. Something between the desk and the ledger ate the difference.

Where the saved minutes go

I have watched this happen up close. Somebody gets an hour back on Tuesday. Nothing downstream is waiting for that hour, so it becomes a slightly calmer Tuesday. The next handoff still takes the same two days, because the handoff was never the thing that got faster.

An individual gain only reaches the P&L if something is standing at the other end to collect it. In most operations nothing is. The saved minutes scatter across twenty desks, each one real, none of them measurable, and at the end of the quarter the CFO asks what changed and the honest answer is: everyone feels better and the number is the same.

That is the seam problem again, wearing a new outfit. Work does not fail inside roles. It fails between them — in the handoff nobody owns, in the second system somebody retypes into, in the wait that everyone treats as weather. Point a very good tool at the inside of a role and you get a faster role inside an unchanged operation.

What the 6% appear to be doing

Nothing exotic. They redesign the flow of work rather than accelerating its current shape. They pick a seam, write down what it actually costs today, change the shape of it, and measure the same thing again afterward. The intelligence is the smallest part of that sequence.

The unglamorous prerequisite is the baseline. If you cannot say what a process costs you now — in hours, in reaction time, in escalations landing on one desk — then you cannot prove anything you do next, and you will end up in the 37% who deployed plenty and can attribute nothing. In our own work the numbers that survived scrutiny were the ones we had measured before we touched anything: inbound reaction time going from three days to roughly sixty seconds, escalations falling from seven to ten a day down to three to five over six weeks. Both are ordinary findings. They are only defensible because somebody wrote down the "before".

And the redesign has to change who decides what, not just how fast the typing happens. If every judgment call still funnels to the same person, you have widened the pipe leading into the same valve.

The uncomfortable read

Another year of record spending bought a flat result for 94% of the field. Next year's survey will look the same unless the work changes shape, because better models cannot redesign a workflow — that is a management act, and it has no vendor.

The good news buried in the same data: the gap is not technical. Nobody in the 6% has a model you cannot get. They did the mapping. That is available to any operation willing to spend a fortnight being honest about how the work really moves.

Personal productivity is easy to feel and impossible to bank. Until a saved minute has somewhere to go, you are buying speed and shipping nothing.

Sources: McKinsey — The State of AI: Global Survey 2026 · The Register — reporting on the same survey, 25 August 2026