In two years, the share of European workers using AI at work has risen from 26% to 52%. The measured productivity gain averages 4%, but this average masks a divide: in 2024, 41% of large companies (250 people or more) used AI, compared to 11% of companies with 10 to 49 people and 21% of those with 50 to 249 people; all companies with 10 or more employees reached 13.5%; companies with fewer than 10 people are not covered by available statistics. AI adoption increased sharply between 2023 and 2024. Adoption remains very unevenly distributed by company size.
The essentials
- AI use at work has doubled in two years across Europe, rising from 26% to 52% according to the ECB.
- The average 4% gain is mainly driven by medium-sized and large companies.
- No primary data source consulted confirms 60% for micro-enterprises with fewer than ten employees; SMEs with 10 to 249 people stand at 19% in 2025.
- The gap reveals that spreading a technology is not always enough to distribute its benefits equitably.
- The window for action before inequality becomes entrenched is short: available data shows adoption gaps and suggests that support for AI diffusion among SMEs could boost productivity.
90% against 60%: two adoptions that do not look alike
Available data show a marked difference. The EIB estimates that 45% of European companies with more than 250 employees deploy AI; this figure does not measure the share of employee users. No primary data source consulted establishes a share of 60% of employee users in companies with fewer than ten employees.
Adopting AI and drawing a productivity gain from it are two distinct operations. The first is a purchasing decision. The second requires training, organizational change, management time. A large company has human resources departments, training budgets, sometimes dedicated teams for digital transformation. A seven-person bakery or a four-person accounting firm has a subscription to a tool and a twenty-minute meeting to explain how to use it.
The result: the 4% gain measured in Europe by the EIB Working Paper 2026/02 is an average. It combines companies that have restructured their processes and companies that have installed a tool without changing much about how they work. The average sounds good. It does not say who benefits from the gain.
This mechanism—a technology accessible to all, a benefit captured by those with complementary resources—is not unique to AI. The computerization of the 1990s produced the same gradient. Adoption was rapid, while the risk of gaps becoming entrenched remains a forward-looking hypothesis.
The market diffuses without distributing
Economist Dani Rodrik has formalized an idea that markets diffuse technologies but do not necessarily distribute the complementary gains: European data offer an illustration of this. His thesis on conditional industrial policy rests on this observation. According to Rodrik, the benefits of an innovation tend to concentrate in organizations that already have complementary resources.
Access and provision by the employer count among the obstacles, along with training, task relevance, and reliability concerns. AI tools are inexpensive, often accessible through monthly subscription. The problem is downstream: the capacity to train employees, reorganize tasks, measure gains, and adjust. Larger companies more systematically have these capacities.
A competing interpretation exists. Liberal economists—Philippe Aghion on Schumpeterian growth, or Tyler Cowen on innovation diffusion—recall that adoption gaps are transitory in long-term history: technologies eventually diffuse along with their associated gains, as training costs fall and practices become standardized. AI tutors, for example, are beginning to make tool learning less dependent on internal training infrastructure. The argument deserves to be taken seriously.
This thesis assumes that time works in favor of spontaneous diffusion. Yet ECB data indicate that the doubling of adoption occurred in two years, at a rapid pace that raises questions about rebalancing mechanisms. If large firms consolidate a productivity advantage while micro-enterprises adopt without investing in training, the gap risks becoming entrenched. Companies that capture the gain reinvest it in new competitive advantages. The others remain 4% behind.
The tension between these two interpretations—spontaneous diffusion or irreversible stratification—is precisely what the data do not yet settle. They identify the gap. They do not say whether it closes on its own.
The situation of workers on the ground
A survey by IG Metall published in July 2026 among German industrial sector employees sheds light that aggregated statistics do not contain. Workers at large factories describe AI integrated into formalized processes: initial training, support, internal points of contact. Employees at smaller subcontractors describe a different reality: the tool is there, sometimes imposed by a customer, rarely accompanied. Some use it through trial and error. Others work around functionalities they do not understand.
A poorly understood tool produces errors that the user cannot detect. In logistics or accounting, a misinterpreted AI recommendation can cost more than it saves. Productivity gains depend on the quality of interaction between the tool and the user, and that quality depends on training.
The ECB blog of August 26, 2026 indicates that time gains and adoption are unequal depending on tasks and occupations, and that employer training and support are important. AI boosts the productivity of those who already know how to organize an adoption process. It leaves others with a subscription and a solitary learning curve.
The debate on automation focuses above all on the redistribution of jobs, not their net destruction. Robotization and its effects on employment shows that technology redistributes more than it destroys. This redistribution assumes that workers have the skills required to occupy the jobs toward which they are directed. Without training, redistribution becomes exclusion.
Training as missing infrastructure
Economists studying labor markets—Paul Osterman on job quality, Bruno Palier on social protection in transition—converge on the same diagnosis: continuous vocational training is the most fragile link in European systems. Not employment, not wages, not protection. Training.
In large companies, AI training integrates into existing frameworks: skills development plans, partnerships with certified organizations, dedicated time during work hours. In micro-enterprises, training remains less systematized than in large companies.
This asymmetry suggests that productivity gains are unevenly distributed by company size. The aggregate hides the fork in the road. A productivity gap can widen if better-resourced companies consolidate their lead.
Carl Benedikt Frey has documented the mechanisms by which progress stops when a technology benefits a few without diffusing across the productive fabric as a whole. The history of industrial automation shows that winners consolidate their lead and that laggards struggle to catch up on a gap that formed quickly. The window for virtuous diffusion is short.
Rodrik calls this approach conditionality: public aid focuses not on adoption, which happens on its own as the two-year doubling proves, but on complementary investments. Train before equipping. Support use, not just access.
Before 2030: two trajectories for the European productive fabric
No primary data source consulted confirms 60% for micro-enterprises with fewer than ten employees; SMEs with 10 to 249 people stand at 19% in 2025. The question concerns the benefit they derive from it: a gain comparable to that of large firms, or adoption without return on cost.
Two trajectories are plausible by 2030, and current data do not allow us to choose between them.
In the first, targeted public investments allow companies with fewer than fifty employees to access complementary support—training, consulting, organizational adaptation. The 4% gain today concentrated in large firms gradually redistributes across the fabric of SMEs and micro-enterprises. Adoption at 60% becomes a springboard rather than window dressing. Regional programs for digital transformation consulting, funded by a portion of the fiscal dividend generated by large beneficiary firms, constitute the most identifiable lever in this scenario.
In the second, micro-enterprises continue to adopt without training. Large firms can reinvest their productivity gains in other advantages: process improvement, recruiting skilled talent, or cost compression. The EIB observes adoption and gain gaps favoring medium-sized and large companies. A divide could emerge between companies that exploit the technology for progress and those that adopt it without gaining from it. This scenario would produce a structural divergence whose later correction would be more costly than preventive investments today.
The signals that will allow us to distinguish the two trajectories are identifiable. The first: the evolution of the productivity gap between large firms and micro-enterprises, measured annually on a European scale. The EIB and ECB have the instruments to do this, provided they publish this monitoring systematically and disaggregated. The second: the share of micro-enterprises that have adopted AI and have an associated internal training program. A high adoption rate coupled with a low training rate is the precursor signal of the fork in the road.
China faces a comparable version of this problem in manufacturing: rapid robotization without skilled human capital to manage it, with structural shortages that slow the promised gain.
European action in the face of market failures
European industrial policy has a lever that markets do not: conditionality. Structural funds, programs supporting SME digital transformation, co-financed national programs can all be reformatted to require, in exchange for support for technology adoption, proof of an associated training plan.
Several regional programs in Germany and the Netherlands have tested digital transformation consulting windows for very small businesses, publicly funded and delivered by local organizations. Preliminary results show sustained engagement from beneficiary companies, even if measurable productivity gains take time to appear.
The problem with these initiatives is their scale: they reach thousands of businesses in pilot regions, while Europe’s micro-enterprise fabric numbers in the tens of millions of units. Scaling up requires structural funding and European coordination that dispersed national initiatives cannot produce alone.
Rodrik’s thesis on conditional industrial policy requires a state capacity, or, in the European case, a supranational capacity, to set conditions, verify them, and draw consequences. The European Union has cohesion funds and the ESF+ as well as recommendations and cooperation frameworks on vocational training. Their application to AI conditionality varies by region and program.
The doubling shows rapid diffusion of stated use, despite persistent adoption obstacles. Is the same pace possible for diffusion of benefits? Not spontaneously. But Europe has the instruments to organize it, provided it uses them before the gap between large firms and small structures becomes the new normal.
Sources
- EIB Working Paper 2026/02, European Investment Bank
- ECB Blog, August 26, 2026, AI adoption and productivity in European firms
- IG Metall, German industrial sector employee survey, July 2026
- Dani Rodrik, works on conditional industrial policy and the trilemma of globalization/democracy/national sovereignty (Harvard Kennedy School)
- Carl Benedikt Frey, How Progress Ends (Oxford University Press)
- Paul Osterman, works on job quality and vocational training (MIT Sloan)
