The Thirty-Eight Point Gap

In 2024, the difference between large and small EU enterprise AI adoption was 30 percentage points. In 2025, it is 38. The tools got cheaper. The interfaces got simpler. The gap got bigger.

That is not a forecast. It is a measurement. The Eurostat release of 11 December 2025 reported that 55% of large enterprises (250+ employees) used AI technologies in 2025, against 17% of small enterprises (10-49). One year earlier, the same survey showed 41% and 11%. The gap widened in the year the tools were supposed to democratise.

This is not an awareness problem. It is an architecture problem.

What the Numbers Say

The Eurostat Community Survey on ICT Usage in Enterprises is the canonical EU baseline. Sample size: roughly 157,000 enterprises across all 27 member states, with at least 10 employees. The 2025 wave was released on 11 December 2025. The headline — 20% of EU enterprises use AI — was repeated everywhere. The size-class breakdown was not.

Here is the size-class data, year over year:

Small enterprises moved from 11% to 17% — six percentage points of growth. Large enterprises moved from 41% to 55% — fourteen. Medium enterprises (50-249) moved from 21% to 30%. Every size class grew. The largest grew fastest.

The Ipsos report Making AI Work for Europe, commissioned by Google and published in March 2026 by Reece Decastro and Nathan Bransden, names the pattern in one sentence: “Large businesses are considerably more likely to adopt AI than small and medium-sized enterprises (SMEs), with this gap widening from 30 percentage points in 2024 to 38 percentage points in 2025.” The authors drew on approximately 70 studies and 15 expert interviews across 13 EU member states. They reached the same conclusion the Eurostat microdata reached, then named it.

The EIB Investment Survey 2025 — using a different methodology, with brand-name AI tools rather than technical categories — reports 28% of SMEs and 44% of large businesses using AI. Different absolute numbers. Same shape. SMEs lag, and the gap is not closing.

The geography of the gap is as sharp as the size class. Eurostat’s 2025 country-level numbers run from 42% in Denmark, 37% in Finland, 35% in Sweden — to 5% in Romania, 8% in Poland, 8% in Bulgaria, 9% in Greece. The variance is not random. It tracks digital infrastructure, not GDP. The same is true by sector: 63% of information and communication enterprises use AI; 40% of professional, scientific and technical services; 17% of manufacturers; 11% of transportation and storage; 11% of construction. The OECD’s December 2025 paper for the G7 Montreal meeting notes the same pattern across G7 economies: 40% of firms with 250 or more employees used AI in 2024, against 20.4% of firms between 50 and 249 and 11.9% of firms between 10 and 49. The country names change. The structural shape does not.

Why the Gap Widens When Tools Get Easier

The intuition is that easier tools narrow gaps. A chatbot interface is a chatbot interface whether the user is at Siemens or at a 30-person logistics firm in Tarragona. The licence fees are similar. The cloud APIs are identical.

The intuition is wrong, and the reason is structural.

Large enterprises have three things small enterprises do not. The first is dedicated digital infrastructure. A 500-person manufacturer has an IT department that evaluates tools, manages integrations, handles security review, and absorbs the cost into an existing budget line. A 30-person logistics company has one person who manages the phone system, the CRM, the printers, and is now expected to evaluate AI tools as an interruption to their actual job.

The second is data. The OECD discussion paper AI adoption by small and medium-sized enterprises, published in December 2025 by Flavio Calvino and colleagues for the G7 Industry, Digital and Technology Ministerial Meeting in Montreal, lists four enablers of AI adoption: connectivity, AI-enabling inputs, skills, and finance. AI-enabling inputs means data — structured, accessible, sufficient in volume, governed well enough to feed a model. Large enterprises crossed that threshold a decade ago, when they implemented ERP systems and cloud data warehouses. Small enterprises mostly have not. An AI tool that needs structured input cannot work against a whiteboard.

The third is the ability to define a use case. BPI France’s 2024 IA Révolution report found that 72% of SME leaders in France do not yet have a practical application of AI for their business. Not “have not deployed” — do not have. The use case itself does not exist in their head. They have heard of AI. They cannot point to the task it would replace.

When tools get simpler, the people who already had infrastructure, data, and a use case move faster. The people who had none of those three are still where they were. Cheaper tools do not generate use cases. Simpler interfaces do not produce structured data. A frictionless API does not give a 30-person company the slack to design a workflow.

The gap widens because the prerequisites widen first.

The Skills Story Is Half the Story

Half of every EU policy discussion about the adoption gap defaults to skills. The Ipsos report cites an OECD survey covering four G7 countries: 50% of SMEs report that their employees lack the skills to use generative AI. The OECD’s own D4SME survey, sampling nearly 1,000 respondents across six G7 countries, found over 50% citing lack of knowledge about how to use generative AI as a barrier — with wide national variation, from 80% in Japan to 40% in the United Kingdom and Germany.

The number is real. It is also incomplete.

A Public First poll of European business leaders, also cited in the Ipsos report, found only 14% citing “we don’t have the expertise to introduce AI” and 12% citing “we don’t have the skills to make use of AI” as main barriers — well below cybersecurity concerns (26%), inaccuracy (24%), and cost (22%). Asked one way, skills are the dominant barrier. Asked another way, they are one barrier among several.

What the data actually shows, when you read it across surveys instead of within them, is this: skills are correlated with adoption, but skills are downstream of structure. An academic interviewed for the Ipsos report put it plainly: “It’s not size…it’s really related to digital maturity.” A Belgian think tank representative in the same study: “SMEs are already slower in basic digitalisation, so building AI on top of that is even more difficult.”

The OECD survey adds the operational fact that determines everything: under 30% of SMEs that already use generative AI report that their employees participate in AI-related training. The range runs from 11.3% in Japan to 29.4% in Canada. The training problem is not that the training does not exist. The training problem is that the time to attend it does not exist — because SMEs cannot release staff from revenue-generating activities to learn. The skills gap is, in part, a time gap that masquerades as an aptitude gap.

You do not close a structural gap with awareness campaigns. You close it with architecture.

What Architecture Actually Means

In an SME context, “architecture” is not a slide with five boxes and arrows. It is four operational decisions, taken in order.

One: data readiness. Before any AI tool, the company must know which data it has, where it lives, how it flows, and what is missing. Most SMEs cannot answer those four questions on Monday morning. The first month of an AI deployment is not a model. It is a data audit: invoices, customer records, supplier information, operational logs. Where is it. How is it stored. Who owns it. What is broken. Until that is mapped, every AI tool sits on top of a foundation nobody has inspected.

Two: a defined use case, narrow enough to ship. Not “improve customer service with AI.” A specific task, a specific person, a specific frequency: classify inbound customer emails by urgency, route to the right team, do it within sixty seconds, measure escalation rate weekly. The OECD’s taxonomy in the same December 2025 paper distinguishes AI Novices, AI Explorers, AI Optimisers, and AI Champions. The leap from Novice to Optimiser does not happen because tools improve. It happens because a single concrete use case gets shipped and then a second one. Use cases compound. Strategies do not.

Three: role-specific deployment. AI tools are not deployed to a company. They are deployed to a role. The person who classifies invoices needs a different tool, a different interface, and different training than the person who writes customer outreach. A single “AI rollout” addressed to “the team” produces a single rate of abandonment: high. Role-specific deployment — one role at a time, with the workflow examined, the tool configured for that workflow, and the user trained to use it inside their actual job — produces adoption. This is not a methodology preference. It is what the corpus of SME case studies shows works.

Four: adoption measurement from day one. Not “did we deploy it.” Did the people use it last week, in their actual work, for the task it was meant for. Daily active use by role. Weekly outcome metric — fewer escalations, faster turnaround, more accurate forecasts. The Ipsos report ties this back to organisational capacity: “AI value creation depends on organisational capacity.” A tool nobody used last week did not create value last week, regardless of how much was paid for it.

These four are not a transformation programme. They are the minimum architecture for a single shipped AI tool. The companies in the 17% closed the loop on those four. The companies in the 83% did not — and the reasons they did not are structural, not motivational.

A Worked Example

A specific shape clarifies the abstraction. A 40-person specialty manufacturer in northern Portugal — real type, anonymised — wanted to “use AI” for customer service. The first conversation surfaced the actual problem: a single operations manager spent four to six hours per week sorting customer emails by urgency and routing them to the right shop floor team. The volume was rising. The manager was the bottleneck.

That is the use case. Not “AI for customer service.” Sort inbound customer emails by urgency, classify by product family, route to the right team, measure response time. One role. One task. Measurable in a week.

The data audit took five working days. The emails existed in a shared mailbox. The product taxonomy existed in the ERP. The routing rules existed in the manager’s head. Two of those three were not in any structured form — the routing rules had to be extracted in a half-day session with the manager, written down, and turned into a classification schema. The schema is the actual deliverable of week one. The model is the deliverable of week two.

The tool, when it shipped, was small. It read inbound emails, applied the schema, assigned an urgency level and a product family, and pushed the message into the right team’s queue. The manager kept the override button. Daily active use began the day the tool shipped. Weekly outcome metric: average response time dropped from forty-one hours to nine. The manager got four hours per week back. The team stopped missing urgent messages from the largest accounts.

The architecture was the work. The model was twelve lines of API calls. The cost of building it was not the model — it was the schema, the audit, the role definition, and the agreement on what success looked like before anything was deployed. That is what closing the gap looks like at the unit level.

What Bluewaves Sees

The first conversation with every Bluewaves prospect goes through the four points above, in order. About a third of conversations stop at point one. The company does not have structured data, does not have a system of record, does not know where its operational information lives. The work they need is not AI deployment. It is the digital foundation that AI requires. We say so. We do not take the engagement until the foundation is in place — or until the company decides that putting the foundation in place is the engagement.

About a third stop at point two. The company has data, has interest, has budget, but does not have a use case narrow enough to ship in three weeks. The first deliverable for these companies is not a model. It is a use-case workshop with the people who would actually use the tool, designed to leave the room with one specific, narrow, measurable task. Then we build.

The last third arrives with all four points already in place — data, a defined use case, the role identified, the willingness to measure. These deployments ship in three weeks and stay in use because the prerequisites were met before the code was written.

This is not a sales methodology. It is the operational reality of why the Eurostat gap widens. The companies in the 17% have crossed thresholds that the companies in the 83% have not. The crossing is not bought with software licences.

The Architectural Position

The gap is widening because the conditions for closing it are not in the tools. They are in the companies. A cheaper API does not produce data readiness. A simpler interface does not produce a use case. A friendlier chatbot does not give a 30-person company the slack to design a workflow. The structure has to come first, and most SMEs have not been given any help building it — because help, in the EU policy conversation, has been defined as training, webinars, awareness campaigns, and pilot funding. Those are not architecture. They are commentary.

The Ipsos report cites Copenhagen Business School research that names the architecture honestly: “AI value creation depends on organisational capacity.” That is the sentence. Organisational capacity is the bottleneck. Capacity is built by structure, not by content. The report’s own second principle to policymakers is consistent: “Support the development of organisational capacity and readiness.” Not more training. Capacity.

If the 2026 Eurostat wave is published in December and shows a 40-point gap, no one should be surprised. The same conditions are in place. The same prerequisites are missing. The same support apparatus continues to address awareness rather than architecture. The gap is doing exactly what its structural causes predict it will do.

Closing it requires a different intervention. Not more pilots. Not more case studies. Not more webinars. Pre-qualified tools for pre-defined use cases, pre-mapped to the role and the data the company already has. External championship where internal championship does not exist. Adoption measurement that begins on day one, not after the third quarter review.

That is the architectural fix. Everything else is theatre. The gap is structural. The fix is structural. The cheaper tools and the simpler interfaces are real — and they will keep widening the gap until the prerequisites are addressed at the same speed the tools improve.

Thirty-eight points last year. The number will be larger next year unless the support stops being content and starts being structure.