There is a number buried in McKinsey’s latest State of AI report that deserves more attention than it has been getting.

Eighty percent of survey respondents say AI has improved their individual productivity. Half say it helps them make better decisions. These are significant numbers, and they align with what most people who use AI tools seriously report from their own experience.

Then there is this: only thirty-seven percent say AI has contributed positively to their organisation’s EBIT. And that number is essentially unchanged from the previous year.

The gap between those two figures is the most interesting thing in the report.

This is not a new problem

In 1987, the economist Robert Solow made an observation that became known as the productivity paradox. “You can see the computer age everywhere,” he wrote, “except in the productivity statistics.”

Computers had been a serious business investment since the 1960s. By the mid-1980s, organisations had spent heavily on hardware, software, and the training required to use it. Individual workers were more productive in a narrow sense: faster at certain tasks, better at organising information, capable of things that would have been impractical without a machine. And yet aggregate productivity statistics barely moved.

This seemed paradoxical. The technology was real. The individual gains were real. The economic impact, measured at the level of the organisation or the economy, was stubbornly absent.

What eventually changed

The paradox was eventually resolved, though the resolution took longer than most people remember. Productivity statistics did start moving in the 1990s, and computing investment was eventually credited with a meaningful share of the gains.

But what changed was not the technology. By the time the numbers moved, computers were not fundamentally more powerful than they had been in the decade when the productivity gains were absent. What changed was how organisations were built around them.

The businesses that captured value from computing in the 1990s did not just equip their people with better machines. They redesigned how work was coordinated. Hierarchies flattened. Processes were reengineered around what computers made possible rather than what organisations had always done. Management layers whose primary function was information relay were removed, because computers could carry information more reliably and cheaply than middle management. Supply chains were restructured. Decision rights were redistributed.

The technology was necessary. It was not sufficient. The organisational change was the unlock.

What the McKinsey gap is telling us

The 80/37 split is the same story, forty years later.

Individual productivity is up because AI does what AI does well: it generates faster, synthesises broader, and operates at a scale and consistency no individual can match. The people using these tools are genuinely getting more done.

But individual productivity and organisational value are not the same thing. A person can produce twice as much output and still deliver no more value if the bottleneck is not in their individual production. And in most mature organisations, that is precisely the situation.

The implicit assumption behind almost every AI business case is that the constraint is in the individual: their time, their capacity, their knowledge. Remove that constraint and value flows. The McKinsey data, year on year, is suggesting that the constraint is elsewhere.

The coordination layer

There is a layer of organisational life that sits between individual effort and organisational outcome. It is not glamorous. It does not feature heavily in most AI investment conversations. But it is where most business value is actually created or destroyed.

It is the layer that determines how strategy translates into day-to-day priorities. How information moves from the people who have it to the people who need it. How decisions get made, escalated, stalled, or quietly avoided. How one team’s output becomes another team’s useful input, rather than sitting in an inbox or a document nobody reads. How a piece of work crosses a handoff without losing half its meaning in the process.

This is the coordination layer. It is made up of meeting structures, approval processes, information architecture, escalation paths, decision rights, and the informal patterns that emerge when none of those things are working well enough. It is also where most organisational waste actually lives: not in individual inefficiency, but in the friction between individuals.

AI does not touch this layer. A person who is twice as productive at generating analysis still operates in an organisation where that analysis has to get to the right person, be understood, prompt a decision, and feed back into execution. AI improves what happens at the nodes. It does not improve what happens in between.

The visibility problem

There is something else worth noting. AI may not fix the coordination layer, but it does make its weaknesses more visible.

When individual output was the bottleneck, organisations could tell a coherent story about where the friction was. The work was slow because people were slow. The quality was inconsistent because people were inconsistent. The solution was to hire better people, or train the ones you had.

As AI removes those individual constraints, the coordination failures that were always there become harder to ignore. Work moves faster to the point where it hits a handoff, and then stops. Analysis is generated at volume, but nobody changes what they decide based on it. Output accumulates at the edges of the organisation while the centre is still running on the same information it always had.

The bottleneck has not moved. It was always there. It is just no longer possible to blame the individual.

What this means for how you invest

None of this is an argument against AI investment. The individual productivity gains are real, and in competitive markets, organisations that are not equipping their people will fall behind those that are.

But the organisations that will capture the full economic value of AI are likely to be the ones that treat the McKinsey gap as a diagnostic rather than a disappointment. Not “why is AI not delivering more ROI” but “what does a forty-three-percentage-point gap between individual productivity and organisational profit tell us about where the real constraints are?”

The answer is almost never more AI. It is usually something about how the organisation actually works: how decisions travel, how information flows, how work coordinates at the seams between teams.

The lesson from computing took fifteen years to learn. The organisations that learned it earliest captured disproportionate value in the decade that followed. The ones that kept buying more computers and waiting for the numbers to move eventually caught up, but they spent a long time wondering why the investment was not working.

The technology was never the problem. It rarely is.