Divergent Paths of SME Digitalization: A Latent Class Approach to Regional Modernization in the European Union

Abstract

Small and medium-sized enterprises (SMEs) constitute the backbone of the EU economy, yet their uneven digital transformation raises challenges for competitiveness and territorial cohesion. This article examines the organizational and spatial aspects of SME digitalization across the European Union using Flash Eurobarometer 486 data and latent class analysis (LCA) combined with Bayesian multilevel multinomial regression. The results reveal four SME digitalization profiles—Digitally Conservative Backbone; Partially Digital and Upgrading; Digitally Advanced and Diversified; and Focused Digital Integrators—reflecting diverse adoption patterns of key technologies such as AI, big data and cloud computing. Digitalization is shaped by organizational factors (firm size, value chain integration, digital barriers) and territorial factors (urbanity, border proximity, national digital infrastructure as measured by the Digital Economy and Society Index, DESI). Contrary to linear modernization assumptions, digital adoption follows geographically embedded trajectories, with sectoral uptake occurring even in low-DESI or non-urban regions. These results challenge core–periphery models and highlight the significance of place-based innovation networks. The study contributes to modernization theory and regional innovation systems by showing that digital inequalities exist not only between countries but also within regions and among adoption profiles, emphasizing the need for nuanced, multi-level digital policy approaches across Europe.

1. Introduction

Since the 2010s, digitalization has emerged as a critical process for the economic growth, resilience and competitiveness of the European Union (EU). The backbone of the EU economy, small and medium-sized enterprises (SMEs) are at the center of this transformation. Despite focused policies and colossal investments, the level of SME digitalization diverges significantly between EU member states as well as within states. This poses significant questions regarding the spatial dimension of the digital transformation and the structural determinants supporting or excluding it.

SMEs’ digital transformation is a top concern in EU policy and research, as it is regarded as a key driver of innovation, competitiveness and territorial cohesion [1,2]. However, more and more empirical evidence supports the argument that this is an uneven spatial process. Instead of alleviating regional disparities, digital transformation is reproducing or even increasing them [3,4].

Under these conditions, the classic core–periphery distinction [5,6] regains analytical significance. Digitally developed regions in Northern and Western Europe have concentrated digital infrastructure, top-class workforces and innovation-based industries. Peripheral and semi-peripheral regions in Southern and Eastern Europe are lagging due to their insufficient human capital, weak institutional capacity and infrastructural deficits [7,8].

This aspatial development manifests itself as a spatial digital divide [9] beyond mere internet availability, including second-and third-order inequalities such as differences in digital competencies, use routines and the ability to make socio-economic advancements through technology [4,10].

2. Theoretical Framework and Literature Review

2.1. Theoretical Framework

Underlying such an issue is the concept of territorial capacity and regional innovation systems (RIS). As Camagni [7] defines it, this refers to the sum of physical infrastructures, human and social capital, institutional quality and local networks which enable the possibility for regions to exploit processes of innovation. Without sufficient territorial capacity, even resource-rich digital projects can fail [11]. Evolutionary economic geography explanations [12] and new regionalist explanations [13] both stress the significance of local innovation systems and path dependency in regional modernization. Modernization is not merely the external diffusion of models but a locally rooted process based in historical pathways, social capital and networks of entrepreneurs [14]. The differentiated modernization concept by Heidenreich and colleagues [15,16] builds on this perspective by arguing that technological change and digitalization are inherent in region-specific modernization regimes. Modernization, they argue, is structurally conditioned by combinations of structural opportunities—such as innovation potential, institutional structures and strategic capabilities—that vary geographically. This explains why some regions and firms develop dynamically out of their digital potential and others remain passive.

Within this theoretical context, greater weight has been given to the actions of local actors as purposeful forces of digital change. Some research [11,14,17] emphasizes that local institutional entrepreneurs, public administrators, business leaders and intermediary organizations play a significant role with regards to mobilizing assets, forging innovation networks and translating digital policies into tangible action. In all this, digital transformation is more than a technological diffusion problem; it is an institutional coordination and local agency phenomenon. This argument is reinforced by the smart specialization strategy of the EU, which invites regional policy based on local strengths, innovation capacity and institutional configuration [18,19]. However, the success of this strategy depends considerably on the administrative and strategic capacities of individual regions.

Taken together, these perspectives highlight that digital transformation is non-linear: it does not follow a single staircase of adoption but comprises multiple, territorially rooted pathways shaped by firm capabilities and regional systems. This theoretical expectation—that SMEs’ digitalization unfolds through distinct profiles rather than uniform stages—informs the empirical strategy adopted here.

2.2. Literature Review

Even if more people are using digital technology (DT), the research reveals that they are still not evenly spread out among different types of businesses, industries and regions. European large and medium-sized businesses have caught up with their worldwide counterparts while small businesses still lag behind, especially when it comes to more advanced applications like artificial intelligence or advanced analytics [20]. This gap is not limited to company size but also concerns resources and the ability to absorb new information. Smaller companies have a harder time getting the money, management and people they need to support digitalization. Sectoral diversity makes things even more complicated. Research indicates that process industries, often perceived as conservative, in fact exhibit a dual structure. Traditional industries like wood and paper are slower to go digital, whereas the food, beverage and pharmaceutical industries are significantly faster [21]. This is in line with the finding that new and creative start-ups can change the direction of technology. Their flexibility lets them try out digital products sooner and share them with more people, making them diffusion agents along with established companies [22]. In this way, entrepreneurship has become a structural force behind the use of DT in Europe.

The relationship between digitalization and innovation performance has been extensively debated yet remains complex. A great deal of evidence indicates a positive correlation between the extent and profundity of DT utilization and the creation of novel items and services [23,24,25]. Specifically, logistics and process-oriented digital innovations have a robust correlation with product-level outcomes, emphasizing the significance of organizational routines in moderating adoption effects. However, researchers’ findings also warn against the assumption that the effects will be the same everywhere. Studies that look at more than one country show only weak associations or none at all, meaning that the institutional and market environments affect the extent to which digitalization leads to measurable innovation [26]. This vagueness highlights the necessity for more sophisticated, comparable methodologies.

Territorial conditions also play a big role in how people accept new things. Recent studies show that infrastructure, human resources and R&D capabilities aren’t evenly distributed throughout Europe. This affects companies’ ability to move beyond basic tools [27]. Small and medium-sized enterprises (SMEs) in rural areas face ongoing problems such as poor connectivity, no secure payment mechanisms and a lack of skilled workers [28,29]. Nonetheless, research from outside Europe indicates that rural areas can also experience substantial benefits from the expansion of digital infrastructure, as evidenced by the employment gains reported in China [30]. These findings bolster the assertion that rural regions need not be regarded solely as backward but as prospective arenas for digital opportunity, given the appropriate conditions to facilitate this.

Policy frameworks are very important in this area. The OECD’s D4SME survey and 2024 Digital Economy Outlook show both the progress that has been made and the problems that still need to be solved. Core tools like e-commerce and cloud computing are becoming more commonplace, but the shift to applications using more data and AI are taking longer and has not been as smooth. Capability gaps—digital skills, cybersecurity and management capacity—continue to deepen the cleavage between SMEs and larger enterprises, while regional discrepancies in infrastructure worsen these inequities [31,32].

Our central argument is that digital transformation among European SMEs is not uniform or linear but comprises a number of territorially embedded trajectories, conditioned by institutional environments, sectoral rationalities and spatial variables. Four empirically derived digital classes, ranging from conservative to advanced and sectoral adopters, demonstrate the multi-dimensionality of digital inequality that cuts across national, regional and organizational lines. Upon this basis, we offer responses to three principal research questions:

• What are the hidden digital adoption profiles of EU SMEs, and how do they vary in terms of technological scope and intensity?
• How do firm-level and country-level digital readiness levels (e.g., DESI) predict class membership, and what are the determinants of these factors?
• To what extent do spatial and institutional contexts, i.e., urbanization and border proximity, shape the digital transformation paths of SMEs across regions?

Building on these research questions and the theoretical assumptions of territorial capacity [7] and differentiated modernization [15], the following hypotheses (H) guide our empirical analysis:

H1. 

The digital transformation of SMEs is developing unevenly across regions due to specific territorial and institutional contexts.

H2. 

The degree of digital adaptation among SMEs depends on both firm-level resources and local interaction networks.

H3. 

Regional variations in digital transformation reflect processes of differentiated modernization, with peripheral areas developing adaptive innovation models.

H4. 

Digital maturity in SMEs results not only from technological investments but also from an organizational culture open to change and learning.

3. Data and Analytical Strategy

3.1. Data Sources

The Flash Eurobarometer 486: “SMEs, Start-ups, Scale-ups and Entrepreneurship” [33] was conducted from February to May 2020 across 34 countries, encompassing all EU member states, accession candidates and specific adjacent economies. The European Commission commissioned the survey, and it was carried out by TNS Political & Social. Its results are available to the public via the GESIS repository. The sample design is based on sizes and types of businesses, with about 500 SMEs surveyed in each country. Neither micro-enterprises with less than one employee or major businesses with more than 250 employees were included. The people who answered were high-level decision-makers, such CEOs, general managers or legal representatives. They were interviewed over the phone using computer-assisted telephone interviewing (CATI). There are more than 15,000 SME-level observations in the collection. These include specific information about the organization’s characteristics, innovative activities, digitalization practices, financing conditions and ownership arrangements, as well as how people there think about the business climate. This dataset is the source of all the firm-level variables used in our study.

Observations with missing or ambiguous responses (e.g., “Don’t Know” or “Not Applicable”) on key variables were excluded to ensure analytical validity. We excluded all non-EU countries. The cleaned dataset includes 10,027 firms, with proportional representation according to region and size (1–249 employees).

We also included the Digital Economy and Society Index (DESI) for 2020 as a country-level measure of a given country’s readiness for digital technology. The European Commission publishes this composite index once a year. It offers a snapshot of connectivity, human capital, DT integration and digital public services.

Flash Eurobarometer 486 provided us with all the firm-level variables, while DESI gave us the sole extra country-level variable.

3.2. Methodology

Our research began with the identification of latent subgroups of SMEs by their adoption of advanced digital technologies (DT). An initial latent class analysis (LCA) was performed to classify heterogeneous digitalization profiles, using indicators at the company level, i.e., artificial intelligence, cloud computing, robotics, smart devices, big data, high-speed internet and blockchain. LCA assumes the conditional independence of observed indicators within each latent class. Models were estimated using maximum likelihood with robust standard errors. Model fit was compared using the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), log-likelihood, entropy and classification certainty (>0.80 threshold). Based on these criteria, a four-class solution was retained as the most parsimonious and interpretable. LCA is applied because is especially good at revealing unobserved heterogeneity within populations. It allowed us to categorize the SMEs without relying on theoretically or administratively defined paths to digitalization, which has an exploratory element. This helped us find different digitalization profiles across SMEs instead of relying on overall averages that would obscure crucial variances in their plans and abilities.

To explain class membership, we estimated a Bayesian multilevel multinomial logistic regression with firms (level 1) nested within countries (level 2). Firm-level predictors included number of employees, innovation activities, financing, barriers to digitalization and participation in global value chains (GVCs). At the country level, the Digital Economy and Society Index (DESI) was included as a predictor of digital readiness. Random intercepts were specified to account for intra-country variance. Priors were set as weakly informative Normal (0.10) for fixed effects and Half-Cauchy (0.5) for variance parameters. Posterior estimation was conducted via Markov Chain Monte Carlo (MCMC) sampling with four chains of 4000 iterations each (2000 burn-in). Convergence was assessed using trace plots and Gelman–Rubin R-hat statistics (<1.01). Model fit and explanatory power were evaluated using posterior predictive checks and intraclass correlation coefficients (ICCs). Adding LCA to a multilevel Bayesian regression framework solved the problem of the data’s “nested structure” (firms within countries) and allowed us to look at how several levels of influence act at the same time. Firm-level predictors could explain differences within countries, while country-level covariates, such as DESI, could show how wider structural factors affect adoption trajectories. The Bayesian formulation additionally facilitates an adaptable modeling of uncertainty, the incorporation of weakly informative priors and the estimate of random effects, hence enhancing the robustness and interpretability of our findings.

This is achieved via random effects at a higher level (e.g., country level) or the inclusion of higher-level predictors (e.g., digital infrastructure at the country level, policy structures). This technique addresses the dependency of observations within the same higher-level unit and enables the examination of how context variables affect class membership and predictor impact [34,35].

This combination offers statistical robustness—via enhanced model fit, precise classification and the capacity to identify hierarchical dependencies—as well as theoretical validation for modernization and regional innovation system frameworks, which stress that digitalization manifests in varied and context-specific manners. In this way, our method puts into action the premise that decisions about adopting new technologies at the business level are influenced by larger institutional, infrastructural and cultural factors. They cannot be boiled down to a single linear path of modernization.

This (LCA) technique is particularly designed for modernization and regional innovation system theory–oriented research [11,16,36,37,38,39,40,41,42,43,44,45]. According to these theories, wherein the structural modernization processes take place non-linearly and should result in identifiable “types” of actors (latent classes) whose innovative adoption depends on micro-level resources in addition to macro-level institutional conditions. For example, in the case of European SMEs, the method can separate qualitatively different digitalization profiles and measure the degree to which firm-level features and national-level stages of modernization co-operate to promote advanced digital take-up.

3.3. Variables

In the following section, we present the LCA model’s predictors. The model identifies classes of firms with the same adoption behavior for these technologies. For example, one class may adopt cloud and infrastructure but not the others (basic adopters); another will adopt AI, big data and robotics (advanced integrators) and so on, with limited or selective adoption. These classes were not a priori determined but inductively derived from the data, based on patterns of co-occurrence in technology use:

• Artificial Intelligence (AI), the use of machine learning or pattern recognition systems;
• Cloud computing, the adoption of remote data storage and processing capabilities;
• Robotics, the integration of automated physical systems in production or logistics;
• Smart devices, the use of connected sensors and IoT (Internet of Things)-enabled technologies;
• Big data analytics, the application of data mining and predictive analytics;
• High-speed infrastructure, the availability of fast, reliable internet connectivity;
• Blockchain, the use of distributed ledger technologies for secure digital transactions.

3.4. Covariates

The digital barrier index considers various barriers to digital adoption noted in the question “Which of the following, if any, is a barrier to digitalization in your enterprise?”: lack of financial means, digital skills, IT infrastructure, regulative obstacles, cybersecurity, uncertainty regarding standards and internal resistance. This index determines the structural constraints that impede digital transformation, particularly among SMEs with limited absorptive capacity. Heidenreich [15] argues that these constraints—based on both firm-specific ability and institutional frameworks—lead to structurally divergent modernization pathways. The existence and intensity of barriers have been established as well-fitting explanatory variables in digital divides [46,47].

Firm size was measured by the reported number of employees, which is a standard proxy for organizational complexity and resource endowment. Large SMEs are likely to be in a stronger position to invest in costlier DT, as argued by resource-based theories of inequality in technology adoption [48]. In the context of modernization, size encompasses differences in capacity between firms and is an important control variable in stratified development models.

Firm location was introduced as an ordinal variable to distinguish between rural areas, small towns and large cities to capture spatial differences in digitalization. This aligns with Heidenreich’s concept of uneven modernization, which focuses on the territorial context to influence access to infrastructure, skills and innovation ecosystems. Ordinal structure captures a digital opportunity gradient, where cities tend to possess more institutional and technological resources than peripheries [1,3]. Including this variable allows our model to specify whether SMEs in more urbanized settings are systematically more likely to belong to digitally advanced classes, while rural and small-town firms are disproportionately at risk of digital disengagement.

Industry cluster or SME support organization affiliation indicates a company’s institutional embeddedness as well as access to external resources. Such networks can facilitate technology diffusion, innovation support and knowledge transfer and are thus theoretically significant predictors of digital adoption. By incorporating this variable, the model can take into account whether support-embedded firms are likely to be members of digitally advanced latent classes, in line with modernization and diffusion of innovation theories [36,49].

Global value chain (GVC) membership is a sign of a firm’s integration within transnational value-added production, distribution or service networks. GVC-affiliated firms are facing competitive pressure to utilize DT to attain international standards, improve coordination and increase efficiency in their processes. Adding this variable enables the model to examine whether globally connected businesses are likely to belong to digitally advanced classes, as predicted by modernization theory and empirical observations connecting GVC integration with innovation and digital upgrading [36,50].

Another variable captures significant DT demand differences between goods-supplying and service-supplying companies. Product-based firms, particularly in manufacturing or logistics, are likely to benefit from and invest in production-related technologies such as robotics, Internet of Things (IoT) and smart devices. The inclusion of this variable allows the model to examine whether SMEs with a production orientation are more likely to embark on niche or specialized forms of digital take-up (e.g., Class 4), depicting sector-led trajectories of modernization [15].

The distance measure, whether or not a firm is located near a border of an EU member country, is utilized here as a proxy for geographic openness and cross-border proximity. Border regions can benefit from greater institutional assistance, exposure to trade or infrastructure investment due to EU integration policy. Introducing this variable allows the model to explore whether border proximity reduces digital exclusion probability, particularly among firms in peripheral or rural areas, thus linking territorial embeddedness with outcomes of modernization.

The 2020 Digital Economy and Society Index (DESI) was used as a second-level (country-level) covariate in this analysis to keep variations in digital development across EU member states constant. The DESI is an annually revised composite indicator by the European Commission that condenses performance in four fields: connectivity, human capital, DT integration and digital public services. Higher DESI scores indicate more developed digital economies and societies. At the secondary level, DESI enabled us to take into account the broader national digital context that influences firm-level technology adoption and digitalization bottlenecks [51].

4. Results: Latent Class Construction and Regional Distribution of Digital Modernization Patterns

4.1. Latent Classes Construction

We began with calculating the optimum number of classes (

Table 1. Latent Class Model Fit Comparison.

Model	AIC	BIC	Log-Like.	Entropy	Certainty	Non-0 Classes
3-Class	75,696	75,960	–37,813	0.489	0.533	3
4-Class	75,243	75,611	–37,572	0.534	0.523	4
5-Class	76,473	76,947	–38,173	0.238	0.792	3

Notes: For AIC/BIC, lower values indicate better model fit with penalties for complexity. For entropy, higher values suggest clearer separation between latent classes. Certainty refers to the proportion of cases classified with >80% posterior probability. The 4-class model provided the best balance of statistical fit, class clarity and substantive interpretability.

). To determine the optimal fit number of latent classes, we estimated models with 3, 4 and 5 classes and compared them based on conventional fit statistics (Table 1). The 4-class model had the lowest scores for AIC and BIC, representing better statistical fit with a penalty for added complexity. It also had higher entropy (0.5343), indicating more clear-cut separation between classes, as well as satisfactory classification certainty (52.37% cases with >80% confidence assigned). Although the 5-class solution would be highly accurate for classification, it revealed lower entropy and included an additional class of zero or near-zero size, which indicates overfitting.

Overall, the 4-class solution offered the best balance of fit, transparency and explainability and was used in the final analysis. These classes include basic digital infrastructure (cloud computing, high-speed internet), operational technologies (robotics, smart devices) and more advanced or emerging tools (AI, big data analytics, blockchain). The model clusters firms according to co-occurrence patterns in their adoption behavior, obtaining four internally consistent classes that vary in both the extent and sophistication of digitalization.

Before going into the full class descriptions, it would be helpful to explain how to read the findings of the latent class analysis. Each class comprises SMEs that use DT in comparable ways. The percentages in

Table 2. Latent Class Structure of SME Digital Adoption.

Variable	Class 1	Class 2	Class 3	Class 4
AI	0.0109	0.1483	0.4989	0.0560
Cloud	0.2273	0.5877	0.9457	0.7586
Robotics	0.0163	0.5183	0.3742	0.0268
Smart Devices	0.0999	0.6465	0.7355	0.3417
Big Data	0.0138	0.2155	0.7099	0.1863
High Speed Infra.	0.1160	0.4182	0.8236	0.5394
Blockchain	0.0008	0.0328	0.2400	0.0336
Employee Count (β)	—	1.46 [1.29, 1.63]	1.09 [0.97, 1.21]	0.41 [0.31, 0.52]
Digital Barrier Index (β)	—	0.31 [0.22, 0.39]	0.28 [0.20, 0.35]	0.30 [0.26, 0.34]
GVC (β)	—	1.28 [0.92, 1.64]	1.77 [1.50, 2.04]	1.13 [0.90, 1.36]
Mainly Goods (β)	—	1.35 [1.01, 1.69]	−0.03 [−0.26, 0.21]	−0.37 [−0.53, −0.20]
Location (Ordinal) (β)	—	−0.28 [−0.49, −0.07]	0.78 [0.58, 0.99]	0.59 [0.47, 0.72]
Border Proximity (β)	—	0.10 [−0.36, 0.56]	0.46 [0.11, 0.81]	0.47 [0.23, 0.71]
DESI Index		0.04 [0.03, 0.06]	0.06 [0.03, 0.09]	0.04 [0.02, 0.06]
Average Posterior Prob.	0.9873	0.9245	0.8566	0.9760
ICC		0.155	0.105	0.111
Country-Level Variance		0.190	0.017	0.039
Estimated Pop. Share (%)	50.78%	8.67%	7.12%	33.43%

Notes: Authors’ calculations. AI, Cloud, Robotics, Smart Devices, Big Data, High-Speed Infrastructure and Blockchain report the probability (0–1 scale) that an SME in a given class has adopted the technology (multiply by 100 for %). Estimated Population Share is the fraction of SMEs in each class. Rows labeled (β) are posterior means (log-odds) from a Bayesian multinomial multilevel model (brms) with 95% credible intervals; Class 1 is the reference. Average posterior probability is the mean class-assignment confidence. Country-Level Variance and ICC refer to the random intercept at the country level for each non-reference logit; ICC was computed as σ2/(σ2 + π2/3).

illustrate how likely it is that companies in a certain class will utilize a certain technology. A greater percentage means that the technology is more widespread for that category. The predicted population share shows how many SMEs are in each class. Regression coefficients (β) show which factors of a firm or country make it more or less likely that someone will belong to a certain class relative to the reference group. These results give us a picture of the digitalization trends in each class and an explanation of the structural and environmental reasons that set them apart.

Class 1, Digitally Conservative Backbone (50.78% of SMEs), is the most numerous cluster, representing over half of total SMEs but having the lowest uptake of advanced DT. Uptake of AI (1.09%), robotics (1.63%), big data (1.38%) and blockchain (0.08%) for this group is approximately zero. Even relatively straightforward tools like cloud computing (22.73%), smart devices (9.99%) and high-speed infrastructure (11.60%) are only modest in scope. In the Bayesian latent class regression model, this group is the reference category; thus, its profile is indirectly defined through comparisons with the other groups. Its characteristics suggest small firm size, weak digital fundamentals and a general lack of external or infrastructural enablers. It suggests that most SMEs in the EU are still not very tech-savvy and only use basic technologies. They are at risk of being left behind as the world goes digital.

Class 2, Partially Digital and Upgrading (8.67% of SMEs), is a small, transitory cluster with patchy adoption, mostly involving higher rates of robotics (51.83%), smart devices (64.65%) and cloud computing (58.77%). However, AI (14.83%), big data analytics (21.55%) and blockchain (3.28%) adoption remains low. High-speed internet is available to 41.82% of firms in this group. Regression results indicate that these firms are significantly larger than those in Class 1 (β = 1.46), with moderate exposure to digital barriers (β = 0.31) and integration in GVCs (β = 1.28). They are also more likely to be specialized in the production of goods (β = 1.35) and mildly penalized by peripheral locations (β = −0.28). No border proximity difference was found to be significant. This class characterizes SMEs that are digitally active but still experiencing structural and geographic constraints. These findings show that fewer than one in ten SMEs are in a transitional stage, meaning they only currently use some digital tools but still have problems and restrictions (e.g., where they can use them). This group is developing at different rates, which makes it hard for these firms to catch up to others who have been using DT for a while.

Class 3, Digitally Advanced and Diversified (7.12% of SMEs), is the smallest yet the most technologically advanced group. It has extremely high uptake of cloud computing (94.57%), smart devices (73.55%), big data (70.99%) and high-speed infrastructure (82.36%). Nearly half use AI (49.89%)—a big advantage over the rest—and over a third use robotics (37.42%). Even blockchain, while still niche, is more widespread (24.00%). Firms in this class are considerably larger than those in Class 1 (β = 1.09), strongly integrated into GVCs (β = 1.77) and facing moderately lower digital barriers (β = 0.28). They are geographically closer to urban economic centers (β = 0.78) and are more likely to be close to EU borders (β = 0.46). The results suggest that just about 7% of SMEs have attained a completely diverse and advanced level of digitalization, utilizing a wide range of technologies like AI and big data. This shows that the digital frontrunners are a small group, yet they are setting essential competitive standards.

Class 4, Focused Digital Integrators (33.43% of SMEs), is a big class exhibiting a specialized digitalization pattern. Cloud computing (75.86%), high-speed infrastructure (53.94%) and smart devices (34.17%) are utilized by the vast majority of these firms. While use of AI (5.60%) and blockchain (3.36%) remain low, the group is distinct for its robotics use (2.68%), reflecting targeted automation in particular industries. These firms are relatively larger than those in Class 1 (β = 0.41]) and well-integrated into GVCs (β = 1.13). They are less likely to be specialized in manufacturing (β = −0.37, CI: [−0.53, −0.20]) but have good spatial attributes (β = 0.59) and are more likely to be located near EU borders (β = 0.47). Overall, this group is a mid-level innovator, strategically taking on primary technologies without reaching full-spectrum adoption. Our findings show that around a third of SMEs use this selective digital approach, focusing on key technologies like cloud services and high-speed internet and staying away from more experimental tools. This strategy is a practical proposal for digitalization, taking into account both the needs of the sector and the resources available.

4.2. To What Extent Do Spatial and Institutional Contexts, I.E., Urbanization and Border Proximity, Shape the Digital Change Paths of SMEs in Regions?

Figure 1 considers the effect of location in the national spatial hierarchy (rural to large cities) on class membership probabilities. Once again, Class 1—Digitally Conservative Backbone—dominates at the rural end of the spectrum with over 70% but declines dramatically with increasing urbanization to fall into virtual equality with Class 4 in urban settings. Class 4—Focused Digital Integrators—on the other hand, has a unique rising gradient, becoming increasingly likely in large cities, where its predicted probability comes close to or exceeds that of Class 1. The shift suggests that urban areas offer a more favorable infrastructure or institutional setting for SMEs adopting focused digital integration strategies.

Class 3—Digitally Advanced and Diversified—also experiences a modest increase, especially in urban and town locations, in accordance with its technologically progressive nature and need for established infrastructure and networks. Meanwhile, Class 2—Partially Digital and Upgrading—continues to show minimal change, forming only a small portion of SMEs regardless of location type. Its relative isolation from spatial advantages may be the result of internal constraints instead of external environments.

Figure 2 examines the effects of EU border closeness on the likelihood of an SME belonging to each of the digitalization classes. Firms in Class 1—Digitally Conservative Backbone—exhibit a declining predicted probability of occurrence as the proximity to EU borders increases. Although they are dominant in non-border firms (with over 60% predicted probability), their share drops progressively and dips below 50% in regions close to EU borders.

In contrast, Class 4—Targeted Digital Integrators—indicates a markedly increasing trend with proximity to borders. This class gradually overtakes Class 1 in areas near EU borders, suggesting that cross-border processes can enable or spur more advanced or targeted DT adoption.

Class 3—Digitally Advanced and Diversified—and Class 2—Partially Digital and Upgrading—are low and fairly flat across the border proximity continuum. Although they are small in occurrence overall, Class 3 enterprises have a slightly increasing trend, indicating a minor effect from borders.

4.3. How Do Firm-Level and Country-Level Digital Readiness (According to DESI) Predict Class Membership, and What Are the Determinants of These Factors?

Figure 3 shows the impacts of the DESI index—a composite measure of country-level digital preparedness and infrastructure—on the probability of SMEs to belong to each latent class of digitalization. There is a clear and reversed trend for Class 1 (Digitally Conservative Backbone). Within low-DESI countries, this class is highly concentrated with values near 75%, but it declines steadily as national digital readiness increases. Its predicted probability according to the highest DESI also scores drops below 50%, indicating a displacement effect in more digitally advanced environments. Most heartening is the trajectory for Class 4 (Focused Digital Integrators), whose membership is forecast to rise steeply along the DESI axis. This class begins at around 25% among low-DESI countries and rises towards 50% in the most advanced environments, which implies that improved national digital contexts promote more focused and systematic forms of digital uptake. Class 3 (Digitally Advanced and Diversified) has a limited but notable increase, as well, particularly in the top half of the DESI band. This is due to the likelihood that SMEs at the technological frontier require enabling environments, such as infrastructure, policy and support systems. At the same time, Class 2 (Partially Digital and Upgrading) enterprises are largely static and low throughout the DESI index. This indicates that their challenges are most likely internal or industry-specific, rather than based on national-level digital circumstances.

4.4. What Are the Hidden Digital Adoption Profiles of EU SMEs, and How Do They Vary In Terms of Technological Scope and Intensity?

Figure 4 plots the predicted probability of SME class membership for European countries ranked by their national DESI scores. The plot shows that those countries with the highest DESI values, for example, Finland, Denmark, Sweden and the Netherlands, are dominated by Digitally Conservative Backbone SMEs in Class 1. While all these countries have the most advanced level of digital infrastructure overall, the majority of companies there remain behind in terms of how they utilize cutting-edge DT. What this means is that high levels of national digital readiness may or may not translate to high digital sophistication among SMEs. Moving down the DESI continuum, there is a perceivable shift in the digital class composition. Mid-ranking DESI nations, like France, Italy, Slovenia and Portugal, feature more prominently in Class 4 (Focused Digital Integrators). These SMEs possess better-structured yet selective digitalization strategies—reliant on cloud computing, robotics and smart devices but with constraints limiting full-range technological adoption.

In these contexts, Class 4 companies typically have a strategic response to limited resources or sector-specific digital imperatives rather than pan-firm digital transformation. Class 3 (Digitally Advanced and Diversified) seems to have a proportionally small percentage in most countries but increasing representation among less-developed-DESI nations such as Romania, Bulgaria and Greece. This indicates that in structurally disadvantaged settings, there exist small but important numbers of heavily digitalized enterprises. These might be firms included in GVCs, being the beneficiaries of specially targeted policy incentives, or in export-oriented sectors with more stringent digital needs. Class 2 (Partially Digital and Upgrading) is continually a minority group within Europe but has disproportionately higher membership in countries at the lower end of the DESI index like Romania, Cyprus and Poland. These firms constitute an intermediate stage, with digital transformation underway but retarded by infrastructural, financial or organizational limitations.

Figure 5 graphs the result of a Principal Component Analysis (PCA) of EU countries against the organization of their SMEs in four latent digitalization classes. The first principal component (PC1), explaining 71.7% of variance, captures the dominant axis of variation in terms of countries’ digitalization organization. The second component (PC2) accounts for an additional 21.9%, making it feasible to plot it in two dimensions that represent over 93% of the total variation in class distributions. On the left of the PC1 axis, Sweden (SE), Finland (FI), Denmark (DK) and the Netherlands (NL) are characterized by SME dominance in Class 1—Digitally Conservative Backbone—with relatively uniform digital strategies and lower levels of advanced or diversified digital adoption at the SME level. These countries are grouped very close together, indicating a shared national pattern where the majority of enterprises rely on core digitalization irrespective of otherwise solid national infrastructure. By contrast, on the right side of the PC1 axis, Poland (PL), Italy (IT), Romania (RO) and Slovakia (SK) have greater SME class structure diversity and higher values for Class 3—Digitally Advanced and Diversified—and Class 4—Focused Digital Integrators. This shift is an increase in a diversified digital landscape where different firm segments follow distinctive digitalization paths, possibly because of unbalanced structural conditions or sector specializations.

The second dimension (PC2) also sorts countries by finer differences. Cyprus (CY) and Greece (GR), for example, are in the positive direction on the PC2 axis, indicated by a highly differentiated SME structure with higher proportions of either transition-oriented (Class 2) or niche-oriented (Class 4) adopters. Austria (AT), Malta (MT) and Finland (FI) are near the negative direction, aligned with a more aggregated, possibly less fragmented SME digital context.

5. Summary and Discussion

5.1. Summary

This study is based on evidence from Flash Eurobarometer 486 [33], a cross-sectional large-scale comparison survey completed in early 2020 across 34 European countries. It gathered responses from over 15,000 SMEs, tracking firm-level characteristics, the process of innovation and DT adoption. After data cleaning and restricting the sample to EU countries, the analytical dataset under consideration included 10,027 SMEs from all 27 EU countries.

Our analytical approach employed latent class analysis (LCA) to identify concealed digitalization profiles based on uptake of seven innovative technologies: AI, robotics, cloud computing and big data analytics. Such inductive classification led to four latent classes representing various SME digital strategies. Class membership was taken as the dependent variable for a Bayesian multilevel multinomial regression model. The model prescribes firm-level predictors (global integration, location, size and digital barriers) and is embedded in country-level contexts, with the Digital Economy and Society Index (DESI) as the key macro-structural variable.

The findings reveal apparent structural cleavages in European SMEs’ digital transformation trajectories. With the exception of a few, all companies belong to Class 1, Digitally Conservative Backbone, a low-penetration group of innovative technologies based primarily in rural areas and lower-DESI membership countries. The smaller but discernible company group, Class 2 (Partially Digital and Upgrading), is obviously transiting with moderate penetration of some technologies. Class 4, Targeted Digital Integrators, represents a third of the sample and is characterized by industry-specific, goal-oriented adoption of technologies such as cloud computing, robotics and smart devices. The smallest but most technologically developed group, Class 3 (Digitally Advanced and Diversified), represents slightly more than 7% of SMEs in higher-DESI nations and urban areas that tend to be well integrated into GVCs.

These classifications debunk the concept of a linear one-way path to modernization. While urban–rural polarities do enter into it, the location of concentrated integrators (Class 4) in non-core or infrastructurally disadvantaged areas demonstrates that digitalization may be discriminatory and sectorally rooted rather than homogeneously progressive. This spatial diversity renders normal convergence presumptions via the digital transition challenging and emphasizes the importance of regional and path-dependent factors.

Geographical characteristics—urban proximity and closeness to EU external boundaries—are significant predictors of SME digital advancement. The digitally conservative class decreases with greater DESI values, and its membership becomes increasingly more likely to belong to the focused or advanced adopter classes. PCA analysis of country-specific SME class distributions for the EU verifies the occurrence of a structural digitalization gradient, which follows national digital readiness levels.

5.2. Discussion

Our latent class analysis offers a rich digital transformation context that validates and complicates existing research on SMEs in the EU periphery and beyond. However, we would particularly like to discuss our results alongside other studies using Flash Eurobarometer 486 data. The predominance of Class 1, Digitally Conservative Backbone (far more than 50% of the sample), supports empirical findings about ongoing digital exclusion within Romania and other structurally constrained regions [52]. This suggests the need for targeted digital training programs in rural regions, designed to overcome infrastructural and human capital-related barriers. Yet our pan-European model, extrapolating this trend to much of the EU, shows that digitally conservative SME profiles are not manifestations of peripheral economies but are pervasive even in high connectivity environments. This exemplifies what DESI (2020) has called the infrastructure paradox, whereby digital access is not necessarily an indicator of adoption. Simultaneously, the widespread prevalence of Class 4, Focused Digital Integrators, across diverse regional contexts renders deterministic assumptions included in the Spatial Inequality Thesis [53] doubtful because this thesis assumes that digitalization maintains existing territorial inequalities. Contrarily, our evidence shows that sector-specific and targeted adoption tendencies can be cultivated even under challenged or mid-range DESI conditions. This implies that digital transformation could follow non-linear and context-dependent paths, which are shaped not merely by infrastructural readiness but also by sectoral embeddedness, institutional context and firm-level strategy.

Although rural areas face infrastructural and skills deficits [9], this does not necessarily mean that non-urban areas cannot benefit from agglomeration effects [54]. Digital tools are locally appropriated, mostly in path-dependent ways [55], where technology adoption reflects the area’s manufacturing and industrial profile and is shaped by local cultural norms [56] and institutional thickness [57]. Since digital systems remain grounded in physical place, peripherality per se may not be the decisive factor for industrial digitalization. Instead, other place-specific characteristics—such as the availability of social capital and skilled labor, appropriate infrastructure, industrial traditions and supportive local governance—appear to be more influential.

The spatial patterns presented in our analysis above demonstrate how urban concentration fosters digital adoption, particularly among Class 4 firms (Focused Digital Integrators). This finding complements Segarra Blasco, Porres and Terruel’s [58] conclusion that digital skills and infrastructure are significant AI adoption predictors while adding an important nuance: even where basic infrastructure exists (as in many peripheral countries), adoption remains low due to cultural–institutional barriers.

Our findings have three significant theoretical implications. First, they invalidate the Core–Periphery Dynamics thesis [34,53], which posits the presence of “digital deserts” ordered by spatial digital divides. The dominance of Class 1 (Digitally Conservative Backbone) firms in the periphery as well as core areas, rather than a higher concentration in structurally vulnerable areas, suggests that low digital take-up is more a systemic problem than one that is localized to the geographical periphery of Europe. Contrary to assertions of “free-riding” behavior [58] by less developed ecosystems, our results show that Class 2 (Partially Digital and Upgrading) SMEs represent a residual class through and with minimal spatial variation. Second, the presence of Class 2 even in non-urban contexts shows that digital transformation has the ability to penetrate through specialized, context-specific adoption patterns. Third, our results complement Heidenreich’s regional innovation systems (RIS) and non-linear modernization theory by demonstrating that digital divides occur across several dimensions: between countries (e.g., DESI-based gradients), across regions (urban–rural differentials) and between adoption strategies (generalist vs. specialist). These findings support Asheim, Isaksen and Trippl’s [59] contention that regional digitalization should not be measured solely in terms of uptake volume but also in terms of the forms, function and institutional embeddedness of digital adoption. With varying SME digital classes across various spatial contexts comes the need for multi-level, multi-path explanations of digital transformation in Europe.

Our study also reveals a spatio-descriptive pattern of place-based innovation. The unusually high prevalence of Class 2 (Digital and upgrading) enterprises in non-urban and less populated areas goes against mainstream theories of innovation geography, which widely associate digital transformation with the effects of urban agglomeration. In fact, such observations point towards the possibility of alternative routes of modernization, in line with research on environmentally embedded innovation mediators [60] and reiterating the need for spatially embedded approaches to digital transitions.

Our study with LCA makes four substantive contributions to our understanding of digital transformation in peripheral EU areas. First, it advances the typological analysis of SME digital uptake. By empirically validating and enhancing the conceptual model presented by Ogrean and Herciu [52] for Romanian SMEs, we have established its generalizability to other peripheral contexts within a four-class latent structure. Second, our results confirm the digital competences–infrastructure–adoption nexus hypothesized by Segarra Blasco, Tomás Porres and Terruel [58], while also uncovering structural barriers that still linger even when very low infrastructural thresholds have been met—most notably in the transition-biased Class 2, Partially Digital and Upgrading.

Third, from a policy perspective, the research outcomes highlight the need for graduated and differential interventions, targeted at the particular digital competences and barriers of each class. For example, basic digital capabilities and empowerment could be the solution regarding Digitally Conservative Backbone (Class 1), whereas facilitation at the ecosystem level and sectoral integration arrangements could be more appropriate to enable scale-up among Focused Digital Integrators (Class 4). Fourth, we have emphasized the place-specificity of digital strategies within various EU regions. On the basis of Veiga’s [60] view of environmental innovation and complementing it with global value chain (GVC) integration perspectives, we argue that surmounting peripheral disadvantages requires the coupling of territorial knowledge bases with transnational linkages. The empirical results largely confirm the four hypotheses formulated in the methodological section. H1 is supported by the evident unevenness of digital transformation across European regions, shaped by distinct institutional and territorial contexts. H2 is validated through the observed importance of local networks, sectoral structures and institutional thickness in facilitating digital adaptation. H3 finds support in the presence of alternative modernization paths, where peripheral regions demonstrate adaptive innovation models rather than linear diffusion. H4 can also be substantiated, as organizational culture and learning orientation emerge as key drivers of digital maturity across all classes. Collectively, these findings confirm that the digital transformation of SMEs is territorially embedded, context-dependent and shaped by the interaction between structural constraints and firm-level agency.

6. Conclusions and Limitations

6.1. Conclusions

Based on non-linear modernization theory, Regional Innovation Systems (RIS) and perspectives from organizational sociology on preparedness and dynamic capabilities, this study examined how European SMEs are going digital. A four-class solution was found using latent class analysis (LCA), which balanced statistical fit with meaningful interpretation. The classes are: Digitally Conservative Backbone (50.8%), Partially Digital & Upgrading (8.7%), Digitally Advanced & Diversified (7.1%) and Focused Digital Integrators (33.4%).

These groups are very diverse in terms of their technological reach, resources and location. Where a city is, its proximity to EU borders and how ready the country is for digital technology (according to DESI) all have a big impact on class membership. But the fact that there are still low-adoption companies in nations with high DESI scores illustrates that infrastructure alone is not enough for advanced uptake. Adoption, on the other hand, is the result of trajectories that are rooted in specific places and driven by firm-level skills, regional systems of innovation and larger structural variables.

Policy should therefore move beyond one-size-fits-all strategies, targeting class-specific needs: basic digital vouchers and cyber hygiene for conservative adopters; financing and workforce upskilling for upgraders; AI-readiness and data-governance support for integrators; and diffusion incentives for advanced SMEs. Future research should employ longitudinal designs to document class shifts, incorporate more granular regional indicators and merge quantitative modeling with qualitative case studies to elucidate the organizational and institutional foundations of SME digitalization.

6.2. Limitations

Several limitations should be highlighted when translating this evidence. Firstly, the conceptual model—integrating the technology readiness and maturity models, organizational trajectories of digital transformation and resource-based/dynamic capability approaches—is focused on firm-level processes and may underestimate macroeconomic, policy and international market shocks affecting digitalization paths as well. Second, the latent class analysis (LCA) approach, while well suited to discovering unobserved patterns of adoption, is sensitive to model specifications (number of classes, choice of indicators) and does not handle dynamic class-to-class transitions through time. Due to the cross-sectional nature of the analysis, causal implications for the link between technological scope-contextual conditions and SMEs’ digital strategy are restricted. Third, Flash Eurobarometer 486 data employed self-reported measures, which are subject to response bias and uneven interpretation across countries and sectors. Fourth, country-level digital preparedness as operationalized by DESI captures institutional and infrastructural capabilities but may not be exhaustive in addressing sectoral or regional differences, especially in cross-border and peripheral environments. Fifth, the model doesn’t account for informal digitalization practices or hybrid models of adoption that may blur the boundaries between defined classes. Finally, the patterns found are in relation to the European policy and institutional environment; thus, caution is advised when generalizing them into non-EU contexts. Subsequent research ought to use longitudinal, multi-wave data—e.g., latent transition/growth-mixture and Bayesian dynamic multilevel models—to trace class development over time and sharpen causal claims, cross-validating self-reports against corresponding administrative/transactional or IoT logs and harmonized skill surveys.