New Commission figures show AI use in the EU's ten strategic sectors rising from 7 to 18 per cent. In the economy those sectors were carved out of, it went from 8 to 20. The gap did not close. It widened.
The European Commission's Joint Research Centre published new figures on 14 October showing that the share of firms using AI across the EU's ten strategic sectors rose from 7 per cent in 2021 to 18 per cent in 2025. More than doubling in four years, reported as the good news it plainly is.
The same analysis contains a second number that has gone almost unremarked. Across the EU economy excluding those strategic sectors, AI use rose from 8 per cent to 20 per cent over the same period.
The sectors Europe designated as strategically important, and has been building an industrial policy around, started one point behind the rest of the economy and finished two points behind it. Whatever the designation has done so far, it has not made those sectors faster at adopting the technology the policy is organised around.
The aggregate hides an enormous range. Energy went from 18 per cent to 34 per cent and electronic communications from 12 per cent to 30 per cent, both far ahead of the economy-wide trend. Media sits highest of all. Meanwhile mobility, tourism and agri-food lag, and where those sectors do use AI, the JRC finds it confined to marketing, sales and other simple tasks rather than reaching production.
So the designation is not predicting anything. Two of the fastest adopters in Europe and several of the slowest are sitting in the same policy category, which means the category is not the variable. Something else is, and the JRC's own results say what it is: organisational change correlates with AI use in production, and capital-intensive sectors are growing faster than more balanced ones.
Milgrom and Roberts's 1990 account of modern manufacturing is the framework that fits this, and it fits it almost uncomfortably well. Their argument is that the elements of a production system are complements: the returns to improving any one of them rise when the others are improved too. Flexible equipment, shorter runs, lower inventories, faster product change, different marketing and a differently skilled workforce are not a menu to choose from. Adopt one in isolation and the return can be zero or negative. Adopt the cluster and the return is large.
Two consequences follow, and both are visible in the JRC data.
The first is that adoption is lumpy rather than gradual. A firm does not creep towards a new system one component at a time, because the intermediate states are worse than where it started. It waits, and then it moves in a coordinated jump. Sectors whose firms have a reason to reorganise anyway, because they are replacing capital equipment on a cycle, get that jump cheaply. Sectors whose capital is long-lived and whose processes are stable do not, and they sit at the starting point looking like laggards when they are in fact rationally waiting.
The second is that the measured variable, "uses AI", is close to meaningless on its own. A firm that bought a licence and a firm that rebuilt its scheduling around a model both answer yes. The JRC is careful about this, noting that growth is driven largely by general-purpose models that are cheaper and more accessible. Cheap and accessible is exactly how you get the first kind of yes.
Robotics is running the identical argument with different nouns. Hesai, the Chinese lidar manufacturer, has been pushing into what it calls spatial intelligence and physical AI, arguing that the constraint on deploying robots in productive environments is not the model but the supply of high-quality three-dimensional data about real spaces, and pairing that with work on actuators, the components that turn electrical energy into movement.
That is a complementarities claim made by a company with a balance sheet behind it. The reasoning is that a better policy model earns nothing without the sensing that feeds it and the actuation that executes it, so the returns to improving any one of the three depend on the other two, and whoever supplies the scarce complement captures the value. It is the same structure as the argument that behind-the-meter solar in Kenya is a credit and approvals business rather than a panel business, and it recurs because it is how general-purpose technologies actually diffuse.
The series has already argued that the capacity to use external knowledge depends on prior investment in understanding it, which is Cohen and Levinthal's absorptive capacity, and that technological capability accumulates slowly and firm by firm, which is Lall's argument. Complementarities is a different claim and the distinction matters for what you would do about it.
Absorptive capacity is about the firm's stock of prior knowledge and what it can therefore recognise and use. Complementarity is about the structure of returns across simultaneous changes, and it holds even for a firm that understands the technology perfectly well. A sophisticated agri-food processor can read the literature, grasp exactly what a forecasting model would do, and still rationally decline, because the model only pays if procurement, storage and supplier contracts change with it, and those are owned by three people who do not report to each other.
That distinction has a policy consequence. Absorptive capacity says fund learning. Complementarity says fund the coordination, or fund the changeover trough, or reduce the number of things that must move at once. They are not the same instrument and Europe is mostly buying the first.
The JRC uses the Eurostat ICT enterprise survey, which is the right instrument for measuring diffusion and the wrong one for measuring transformation. Three things would close that gap and none of them require a new survey wave.
Split the yes. Ask adopting firms whether the model sits in a production process or in a support function, and whether any process was redesigned when it arrived. The JRC already distinguishes production use in part of its analysis; making that the headline split would change what the number means.
Date the adoption against the capital cycle. If complementarities are driving this, adoption should cluster around equipment replacement and site reconfiguration. That is testable from existing investment data and it is the cleanest way to separate this explanation from a skills-shortage explanation.
Interview the firms that looked and declined. Non-adopters are treated as an absence in survey data. They are the group that can say which complement was missing, and they are reachable, because in the lagging sectors they are most of the sector. Our interview guide covers how to ask without leading.
For policymakers, the conclusion is that designating a sector does not change the returns inside its firms, and the two sectors racing ahead are doing so for reasons the designation did not create. For consortia, it is that an adoption percentage is a weak outcome indicator and a redesigned process is a strong one. For firms in the lagging sectors, it is that waiting until the next capital cycle may be the correct decision rather than the timid one, and that the thing worth preparing now is the data and the coordination, not the licence.
If this reads close to a decision you are weighing, we would like to hear about it. A short note is enough to start.
The Lab is an independent research team, not an advocacy organisation. We do not campaign, endorse products, or accept commissions where a finding is expected in advance. Where a partner's own theory of change is what needs testing, that is a legitimate engagement, and the Lab tests it honestly.