Investigating the state-of-the-art and challenges of artificial intelligence for production management in Italian manufacturing SMEs

Interesting because – Manufacturing SMEs are under increasing pressure to adopt AI in production, yet their decision-making remains fragmented and poorly understood. While prior research often treats AI as a generic technology, this study shows that SMEs differentiate strongly across applications: they prioritise resource optimisation, energy efficiency, maintenance, and simulation, but remain sceptical about scheduling, root-cause analysis, and predictive quality. This reveals a gap between technological potential and operational reality, especially in traditional shop-floor contexts where data quality, legacy systems, and human expertise are critical.

Theoretical value – The study advances operations management and digital transformation literature by unpacking how SMEs prioritise specific AI applications rather than adopting AI as a monolithic concept. It highlights the role of technological, organisational, and environmental constraints in shaping adoption decisions, contributing to a more nuanced, context-aware understanding of AI readiness and decision-making in resource-constrained firms.

Practical value – The research provides a structured roadmap for SMEs and their managers by identifying high-impact AI applications and key barriers to adoption. It shows that successful implementation depends on improving data quality, developing internal skills, building trust in AI, and aligning solutions with existing processes. Managers can use these insights to prioritise investments and design realistic, stepwise AI adoption strategies.

Artificial intelligence (AI) is gaining increasing attention in academic and managerial fields and is now a key scientific theme for manufacturing companies (Kovič et al., 2024). Small and medium-sized enterprises (SMEs) are increasingly adopting AI in production due to its potential to enhance productivity, reduce operational costs, and improve quality assurance, key drivers that align with the strategic goals of manufacturing firms (Oldemeyer et al., 2025). However, research on AI in production remains less developed than in areas such as product development or marketing (McKinsey, 2024).

Focusing on production management in B2B manufacturing SMEs reveals an important research gap. SMEs face barriers to AI adoption stemming from limited resources, legacy systems, and organisational constraints, even though they may be more agile than larger firms (Kumar et al., 2024; Nagy et al., 2023; Kahveci, 2025). Moreover, applying AI in manufacturing differs from digital-native contexts because it must be integrated into physical shop-floor processes that were not originally designed for data-intensive tools (Min, 2022). A key theoretical gap, therefore, concerns how resource-constrained SMEs prioritise AI applications, since much of the literature still focuses on larger firms or generic adoption patterns (Schwaeke et al., 2025; Oldemeyer et al., 2025).

This research contributes to the debate on AI in manufacturing, focusing on production management and SME adoption challenges. It also outlines future implementation trajectories. To address literature gaps and offer insights for academics and practitioners, we studied Italian manufacturing SMEs in B2B markets, aiming to answer:

We chose Italy, Europe's second-largest manufacturing economy, because 92% of its manufacturing businesses are SMEs (Italian Ministry of Foreign Affairs and International Cooperation, 2023). With 2.3 million SMEs in Europe's manufacturing sector, representing 16.25% of all business sales (Statista, 2024), our sample is broadly representative. The questions move beyond descriptive analysis by examining the mechanisms behind SME AI prioritisation and by supporting the development of a conceptual framework for AI readiness and decision-making. In particular, RQ1 addresses the tendency in the prior literature to treat AI as a generic solution rather than as a set of distinct applications with varying relevance to SMEs. In addition to answering these research questions, this study develops an Actionable Conceptual Framework (ACF), a central contribution of the paper.

We used a mixed-methods approach. A literature review based on the Technology-Organisation-Environment (TOE) framework identified four key areas. Nine theoretical themes were refined by expert panels and tested through a survey of 229 Italian manufacturing SME respondents. Their notes and suggestions were also analysed to gain deeper insights into organisational challenges.

To adopt a more conceptual structure, this review is organised using the TOE framework (Tornatzky and Fleischer, 1990), which helps integrate diverse contributions while also highlighting the specific challenges faced by SMEs.

From a technological perspective, simulation is a well-established tool in production management, enabling firms to analyse system variants and predict the outcomes of structural or behavioural changes. Studies by Kandemir et al. (2025), Zhang et al. (2022), Nagy et al. (2023), and Carl May et al. (2024) showed how digital twins and virtualisation enhance decision-making in material flow, assembly, and shop-floor optimisation. These technologies are now integral to AI-enabled production systems (Alfaro-Viquez et al., 2025; Sajadieh and Noh, 2025).

AI is widely applied to predictive tasks across manufacturing, scheduling, maintenance, and energy management (Javaid et al., 2020; Azeem et al., 2021). Predictive maintenance, in particular, has evolved through sensor-based analytics and multisource data frameworks supported by industrial case studies (Sahal et al., 2020; Drakaki et al., 2021; Nordal and El-Thalji, 2021) and multiple reviews (Bousdekis et al., 2019; Sajid et al., 2021; Wen et al., 2022). Recent works highlight a broader shift toward data-driven maintenance and energy optimisation practices (Polenghi et al., 2021; Skoumpopoulou et al., 2025).

AI in production scheduling has evolved significantly. Early works (Dong, 2021; Parente et al., 2022) highlighted AI's potential for adaptability and optimisation. Del Gallo et al. (2023) reviewed AI-driven scheduling and demonstrated strong performance in dynamic contexts with shifting demand, machine availability, and workforce constraints. Neural networks and reinforcement learning enable predictive and adaptive scheduling that reduces downtime and improves throughput. Kriouich and Sarir (2024) provided a structured overview of AI for job-shop and flow-shop scheduling, noting its suitability for complex systems. Liu et al. (2024) introduced a deep reinforcement learning model with attention mechanisms that optimises scheduling in smart factories by learning from historical data, supporting more autonomous planning. Overall, the literature points to a paradigm shift: firms that adopt AI-driven scheduling gain responsiveness, lower costs, and better decision-making.

Predictive quality control and Root-Cause Analysis (RCA) are also advancing. Chen et al. (2023) and Wehner et al. (2024) developed models that trace quality deviations across production stages, improving prediction and interpretability. Muaz et al. (2025) warned that RCA models can mislead if not rigorously validated. Machine learning-based RCA approaches (Rotter et al., 2023; Pietsch et al., 2024) and predictive quality frameworks (Omisola et al., 2020; Wehner et al., 2024) show how the goal of “zero-defect” production is accelerating AI integration with quality assurance. Collectively, these studies indicate a convergence of predictive analytics, scheduling, and quality management.

Studies highlight the challenges SMEs face in implementing AI across technological, organisational, and strategic dimensions, often due to limited resources and expertise. Key barriers include skills and infrastructure deficits: Sánchez et al. (2025) noted shortages of technical personnel and IT systems, while Omrani et al. (2024) and Mittal et al. (2018) emphasised upgrading infrastructure and digital skills. Strategic issues involve unclear roadmaps, cultural resistance, and weak alignment, though phased adoption can help (Rasdi and Baki, 2025). Trust and interpretability remain critical: Hickman (2025) stressed explainability, as black-box models face resistance. Accordingly, RCA research prioritises explainable or hybrid models (Wehner et al., 2024; Pietsch et al., 2024), and integration with tools like SPC (Omisola et al., 2020) helps bridge trust gaps. Similar concerns arise in scheduling, where heuristic or hybrid methods prevail. Consequently, Schwaeke et al. (2025) found SMEs adopt AI in finance, marketing, and HR but rarely in production management, leaving a gap.

The environmental context encompasses industry trends, regulations, and market pressures shaping AI adoption. Industry 4.0 frameworks provide the foundation, emphasising connectivity and cyber-physical integration (Azeem et al., 2021; Parente et al., 2022; Amarasinghe et al., 2023). Competitive and technological pressures further encourage learning-based approaches in production, maintenance, and quality.

Debate persists over AI's contribution to sustainability. Some studies report improved resource and energy efficiency (Song et al., 2017; Bevilacqua et al., 2017; Liu et al., 2018; Andrei et al., 2022; Shuford, 2024) and reduced emissions (Beier et al., 2022), while others find no direct effects, indicating context-dependent outcomes (Dubey et al., 2019). Recent work extends this discussion through AI-based environmental forecasting (Thombre et al., 2024). Market and regulatory pressures also drive predictive quality and scheduling tools (Del Gallo et al., 2023). RCA and predictive quality target waste reduction and downtime, supporting zero-defect standards (Omisola et al., 2020). Regulations and external factors further mediate AI adoption (Anser et al., 2025).

The literature review reveals several important gaps that justify this study. First, although existing research highlights the benefits of AI in manufacturing and other sectors, it rarely analyses in depth the specific challenges that SMEs face when integrating AI into established production processes, which are often characterised by legacy systems and limited digital maturity (Kumar et al., 2024; Nagy et al., 2023; Schwaeke et al., 2025). Second, the current body of knowledge is mainly focused on large organisations or generic industry settings, leaving SMEs underexplored despite their unique constraints in terms of resources, skills, and organisational structures (Mittal et al., 2018; Omrani et al., 2024). Third, while barriers to AI adoption are frequently mentioned, they are often treated superficially, without a systematic analysis of how technological, organisational, and environmental factors interact to shape adoption decisions (Rasdi and Baki, 2025; Sánchez et al., 2025). Finally, the literature presents contradictory evidence regarding the environmental impact of AI. Some studies report improvements in energy efficiency and sustainability performance (Beier et al., 2022; Andrei et al., 2022), while others find limited or context-dependent effects (Dubey et al., 2019), indicating the need for a more nuanced understanding of these conditions.

Within this fragmented landscape, prior studies converge on several key domains of AI application in production management. These include simulation and optimisation, which leverage digital twins and virtual models to improve decision-making (Kandemir et al., 2025; Alfaro-Viquez et al., 2025); predictive processes, such as maintenance and quality forecasting based on data-driven analytics (Bousdekis et al., 2019; Wen et al., 2022); environmental management, focused on monitoring and reducing emissions (Thombre et al., 2024); and energy management, aimed at improving efficiency and reducing consumption (Skoumpopoulou et al., 2025). These domains provide the conceptual foundation for the empirical investigation conducted in this study.

Our study combines qualitative and quantitative research methods. Mixed-methods research enables the exploration of diverse perspectives and helps uncover relationships that may not be evident through a single method (DeJonckheere et al., 2024). To ensure the consistency and complementarity of our mixed-methods approach, we employed a methodological triangulation strategy: the qualitative insights from a Delphi panel were used to calibrate the survey instruments, while the qualitative comments provided by the 229 SME respondents were systematically reviewed and matched with the quantitative Likert-scale data to either support statistical trends or provide a contextual explanation for any observed discrepancies.

In the first stage, we applied the Delphi method through semi-structured interviews with a panel of 12 experts. The interviews were guided by four literature-based questions, enabling participants to speak freely within their respective domains. For example, to explore simulation and optimisation, we asked:

1. What do you think are/will be the impacts of AI on simulation and optimisation?

Additional question included:

1. What do you think will be the general impacts of AI on production management?

The experts were informed that the discussion focused on shop-floor processes in manufacturing SMEs, excluding logistics, and was grounded in practical experience.

The Delphi method was chosen for its suitability in emerging topics requiring expert consensus. Our panel included AI system integrators, manufacturing consultants, and SME production managers, ensuring diverse perspectives across technological domains. The coding process, spanning from open questions to final themes, was based on Thematic Content Analysis (TCA) (DeJonckheere et al., 2024) and is included in  Appendix. Three rounds of Delphi were carried out.

Round 1: involved 45-minute interviews, transcribed and analysed using thematic content analysis, resulting in 29 initial codes (A_x).

Round 2: grouped these codes into broader themes; disagreements were explicitly discussed, and divergent views were documented to preserve the richness of the data. To guarantee the independence and objectivity of experts, reduce consensus or opinion-leader influence during the second round, all interactions were managed through an anonymous, asynchronous feedback loop, where experts reviewed and synthesised codes independently before a final moderated discussion was used only to document, rather than force, consensus on the emerging themes.

Round 3: validated and refined the themes, yielding nine conceptual categories (T_X) that informed the quantitative phase. This structure preserved exploratory depth and methodological rigour.

Importantly, we addressed a common critique of Delphi studies by incorporating mechanisms to handle expert disagreement. Rather than forcing consensus prematurely, we allowed for minority viewpoints to be retained and flagged for further exploration in the survey instrument. This approach enhances methodological rigour while preserving the exploratory nature of the study.

In the second stage, the nine themes were transformed into hypotheses and tested via a survey. Qualitative outputs from the Delphi rounds do not simply provide themes but also reveal stability signals (Jorm, 2025), areas where expert judgments converge early and consistently across iterations. These stability signals were used in our study to formulate preliminary hypothesis directions for the quantitative stage. Rather than treating expert statements as findings to be confirmed, we used their degree of convergence as an indicator of which AI applications experts implicitly considered high- or low-priority. The survey then served as a generalisation check, allowing us to test whether these expert-level convergence patterns were reflected across a much broader population of SMEs. A questionnaire was sent to 19,706 Italian manufacturing SMEs, selected based on the following criteria to reduce bias:

1. 0–250 employees, with production units in Italy;

2. Intermittent-flow production systems (job-shops, machinery, assembly lines/cells); continuous-flow processes excluded;

3. Not classified under EU major-accident hazard regulations, which require strict environmental systems and compliance.

The sample was selected through a purposive, criterion-based approach because the aim was not to generalise to all SMEs, but to analyse AI implementation in a specific production context. To reduce bias, firms were diversified across industrial sub-sectors and Italian geographical areas, while respondents were selected among production and operations managers with direct knowledge of shop-floor applications.

The questionnaire included questions on sector and AI knowledge (measured on a five-point Likert scale). Of the 295 responses, 59 were excluded because they did not meet the criteria, and 7 were excluded due to missing data, leaving 229 valid responses. Company sizes were: 11 (4.8%) with 1–20 employees, 78 (34.1%) with 21–50, and 140 (61.1%) with 51–250.

Respondents rated the following nine AI applications for production management (1 = not important; 5 = very important), with comment fields for each:

1. Production routing simulation (SIMUL);

2. Machine parameters and yield optimisation (PARAM);

3. Production scheduling optimisation (SCHED);

4. Predictive and preventive maintenance (MAINT);

5. Predictive quality control (QUALI);

6. Root-cause analysis and identification (CAUSE);

7. Resource consumption optimisation (RESOU);

8. Reduction of environmental impacts (ENVIR);

9. Energy efficiency (ENERG).

These nine applications (SIMUL, PARAM, SCHED, MAINT, QUALI, CAUSE, RESOU, ENVIR, ENERG) were operationalised as variables for testing, with open comment fields providing additional organisational insights. These comments marked a key stage, revealing additional themes related to the challenges hindering AI implementation.

The first test addressed common method bias, which may arise when respondents aim for consistency in their answers (Podsakoff et al., 2003, p. 882). We tested this via factor analysis (see Table 1). The eigenvalue in column two shows how much variance each factor explains. Crucially, the first factor accounts for 39.478% of total variance—below the 50% threshold—indicating no significant common method bias.

Next, we applied a one-sample Wilcoxon signed-rank test, a non-parametric alternative to the t-test, suitable when the data is not normally distributed, and the variance is not homogeneous. This test assesses whether the population median equals a hypothesised value (ranging from 1 to 5), and it is known for its robustness with moderate sample sizes (Happ et al., 2019). The null hypothesis (H₀) for each of the nine variables, SIMUL, PARAM, SCHED, MAINT, QUALI, CAUSE, RESOU, ENVIR, and ENERG, is that the median difference from the hypothesised value is zero. The alternative hypothesis posits a non-zero difference.

To describe the quantitative results of the online questionnaire, we first applied cross-tabulation. We immediately noticed that SIMUL, MAINT, RESOU and ENERG recorded a relatively high number of respondents who chose 5 (“very important”). While SCHED and CAUSE had a relatively high number of respondents who chose 1 and 2 (“not important at all” and “of little importance”, respectively). PARAM showed a more uniform distribution of values; QUALI recorded a relatively high number of respondents who chose 3 (“of average importance”); and ENVIR showed 90% of respondents who chose 1, 2, and 3.

We performed the non-parametric one-sample Wilcoxon signed-rank test to compare different medians and identify those (if any) that reject the null hypothesis (p-value < 0.05). Table 2 presents the test results, including the medians that retain the null hypothesis and the corresponding p-values.

The results reported in Table 2 reveal that we could not find a median that retained the null hypothesis for ENVIR, meaning that for this variable, we cannot claim any statistically significant trends.

These types of application showed the highest validated median, almost 5, with 151 fives (65.9%), 55 fours, and only 3 ones. The result reflects rapid market fluctuations for all resources that began with the pandemic and were intensified by the Russia–Ukraine conflict, the Strait of Hormuz crisis and tariff uncertainties. Thirty-one comments highlighted the sharp price increases in recent years affecting manufacturing materials. One respondent illustrated the challenges manufacturing firms have been facing:

Analogous themes emerged from other comments, particularly those referring to real-time data input, procurement agility, and advanced forecasting. But most of the comments clarifies how this kind of AI application is in an early stage of development. From a respondent:

One manager dealing with the same application noted:

Thus, he/she highlighted the challenge of data quality.

Survey respondents and experts consistently considered AI highly suitable for improving energy efficiency, reflected in a validated median of about 4.5. Around 30 comments described existing systems that already collect big data and use AI analytics to schedule machines, adjust production rhythms, detect consumption anomalies, select energy providers, and identify optimisation opportunities. One respondent noted:

The issue of quality data arose again from at least 46 managers, for instance:

Other 39 managers underlined the issue of the lack of internal skills:

Another commented:

In this case, legacy system integration issues emerged.

The 12 experts unanimously highlighted the importance of predictive and preventive maintenance. Survey results supported this view: 132 respondents rated it “5” and 63 rated it “4”, and the Wilcoxon test confirmed a validated median of 4.5, the second highest among all applications. Forty-five respondents reported having already implemented fault-detection tools for machines and assembly stations, typically using embedded smart sensors and AI software. The main challenges concern failures and undesirable events at assembly stations, where conditions are more variable and harder to predict. One respondent noted:

Once more, dozens of managers were worried about a lack of internal expertise:

A relevant issue raised by several respondents is the possibility of dynamically adjusting preventive-maintenance schedules, highlighting a direct link with AI. One respondent explained:

Another respondent refers to this dynamic preventive scheduled maintenance as a combination of predictive and preventive maintenance methods.

AI-based production routing simulation emerged as an important and widely adopted application. The variable SIMUL shows strong support, with 92 respondents rating it “5” and 79 rating it “4”; the Wilcoxon test confirms a validated median of 4. Respondents consistently highlighted the value of simulating, predicting, and optimising production flows to prevent bottlenecks, often referring to these tools as digital twins. However, 15 comments noted a key challenge: developing data-driven systems capable of functioning as dynamic digital twins that adapt to real-time production changes. One respondent observed:

Thus, simulation is key to avoiding unplanned events. Nevertheless, there is a challenge about the possibility of readapting the routing in real-time and the use of historical data.

The experts were almost sure that each machine and station could measure, control and readjust its parameters to increase the yield of the station itself, measured, for instance, through indicators such as the OEE. But when we consider the frequency table of PARAM and the results in Table 2, the difference among the five values of the Likert scale is not very pronounced, and the Wilcoxon test result accepts the hypothesis of a median equal to 3 (“of average importance”). There are 9 comments from respondents who state that machine parameters control applications could be relevant to minimising the machine's cycle time, and the outcomes about quality, and even in relation to resources and energy consumption. But we found 26 other respondents who believe these applications are more in the domain of the equipment manufacturer and do not necessarily have to use AI. One respondent expressed this clearly:

Applications for production scheduling are considered not particularly important for our SME respondents. The variable SCHED has a frequency of 94 ones and 69 twos out of 229 (41% and 30.1%, respectively), and the Wilcoxon test result in Table 2 accepts the hypothesis of a median equal to 2 (“of little importance”). Thus, the respondents from the sample of manufacturing companies do not agree with the 12 experts in this area. Some relevant comments:

According to 63 respondents, the use of simple Lean tools seems to be way out of the instability and scheduling without AI applications.

The experts proposed that RCA and identification applications based on AI are relevant for production management. But the respondents were particularly negative about this prospect. The variable CAUSE has a frequency of 110 ones and 64 twos out of 229 (48% of ones), and the Wilcoxon test result in Table 2 accepts the hypothesis of a median equal to 2 (“of little importance”). The following quote is highly representative of the 38 comments relating to this issue:

Thus, the respondents identified problems with managing structured and unstructured data and with often having to cope with diverse problems and root causes. Moreover, they again argue that people's ability, rather than AI, could solve production problems.

The variable QUALI has a frequency of 104 threes, and the test results in Table 2 accepted the hypothesis of a median equal to 2.7. This suggests that the respondents are neutral regarding AI for quality control. In fact, the suggestions indicate contradictory directions. While we found 48 respondents who are trying to implement AI solutions with digital vision to inspect processes, another 56 have been investing in more traditional applications and software that do not use AI. Furthermore, the challenge of dealing with vast quantities of unstructured data is revealed in comments like this:

pointing to the challenge of capturing tacit knowledge. Another added:

referring to unstructured and badly integrated data.

No median was validated for ENVIR. 24 suggestions indicated that the respondents are investing in more traditional applications and software that do not use AI. From one respondent:

The respondents believe that reducing environmental impacts is more a matter of controlling legal parameters relating to processes and reacting swiftly, rather than analysing and predicting potential incidents.

Despite growing interest in most of the above AI applications across production, many respondents reported general, across-the-board strategic and organisational challenges that hinder adoption. These issues are often less technical and more structural, rooted in company culture, planning, and resource limitations.

A recurring theme among 72 respondents is the lack of strategic planning for AI implementation. AI is often seen as a buzzword rather than a structured investment. As one manager put it:

As previously found, this uncertainty is compounded by a shortage of internal expertise. 50 managers noted a lack of personnel with the skills to evaluate, implement, or maintain AI systems. One respondent noted:

Limited financial resources also play a critical role. Even when managers are convinced of AI's potential, budget constraints often prevent experimentation or pilot projects. These come from an astonishing 79 similar comments. A comment that captures this sentiment:

Another barrier is scepticism toward AI's practical value. We saw how some managers question whether AI can truly outperform human intuition and experience, especially in complex or unpredictable environments.

Resistance to change remains a significant obstacle. Employees may fear job displacement or feel overwhelmed by new technologies, and again, this was expressed by 76 managers. One manager described the situation candidly:

Finally, 41 respondents identified market and customer demands, along with regulatory and standard requirements for quality and environmental performance, as key barriers to AI implementation. One relevant note:

These general issues suggest that successful AI adoption in manufacturing SMEs requires more than just technical solutions; it demands strategic alignment, cultural readiness, and targeted support.

The preceding analysis supports the development of an Actionable Conceptual Framework (ACF), a central outcome of this study. The ACF is designed as an integrated conceptual model that extends the TOE framework by connecting technological, organisational, and environmental conditions with the concrete AI applications that SMEs prioritise in production management. In contrast with prior research that treats AI adoption in general terms (Schwaeke et al., 2025), the ACF shows how different TOE conditions shape the feasibility, attractiveness, and timing of specific AI applications.

More specifically, the framework conceptualises AI adoption as a cause-and-effect process. Technological, organisational, and environmental conditions act as input factors that influence adoption decisions. AI applications represent the process layer through which firms respond to these conditions, addressing the “trust gap” and “black-box” concerns highlighted by Hickman (2025) and Omrani et al. (2024). In this way, Figure 1 moves beyond a descriptive summary of findings and becomes both a theoretical extension of TOE and a practical roadmap for managers seeking to align AI investments with SME-specific constraints.

Using the TOE framework, we discussed our findings, comparing them with the literature reviewed in Section 2. The goal is to highlight how the empirical evidence from manufacturing SMEs aligns with, contradicts, or extends existing theoretical insights.

Respondents broadly support using AI to optimise resource consumption amid volatility in availability and pricing. SMEs are adopting AI to adjust production parameters, schedule processes, and guide procurement, though systems remain in early stages. A key challenge is developing customised software that integrates real-time data, such as production rates and market prices, and uses predictive analytics to automate purchasing, thereby enhancing agility. This aligns with collaborative resource-sharing models (Kessler and Arlinghaus, 2022) and anticipatory production systems (Yang et al., 2024).

Predictive and preventive maintenance is another core AI application, as seen in Polenghi et al. (2021), Abbas (2024) and Tortorella et al. (2024). The emerging challenge is dynamically adjusting schedules based on real-time data, moving beyond static approaches (van Staden et al., 2022). SMEs are also testing AI for predicting failures at assembly stations, an area with limited research. AI-based production routing simulation is gaining traction to optimise processes and prevent bottlenecks.

While production routing simulation is well studied (Javaid et al., 2020), respondents highlighted challenges and new opportunities. Dynamic digital twins (Zhang et al., 2022; Carl May et al., 2024) could enable real-time adjustments beyond historical inputs. Developing AI algorithms for automatic recalibration will enhance flexibility and resilience, an evolution not yet discussed in the literature.

Equipment manufacturers already optimise machine parameters with built-in automation. Respondents suggest AI may not greatly improve these systems, particularly for Industry 4.0 machines. However, advanced AI could still provide refined optimisation insights, though currently seen as less critical given ongoing vendor improvements.

AI for RCA received limited SME interest due to unstructured data and ERP limitations, especially with high variability and complex defects. Similarly, predictive quality control still relies on traditional models. Respondents acknowledged AI's potential, especially in digital vision, but view handling massive unstructured data as a challenge, consistent with Taleb and Serhani (2018). Data quality and legacy system issues, noted by Müller et al. (2018), remain widespread barriers.

Organisational barriers were among the most cited. A lack of strategic planning, expertise, and financial resources aligns with Sánchez et al. (2025), Omrani et al. (2024), and Mittal et al. (2018). AI is often perceived as a buzzword rather than a structured investment, reinforcing Rasdi and Baki's (2025) call for phased strategies.

Contrary to literature supporting advanced production scheduling (Parente et al., 2022), respondents are hesitant to adopt AI-driven systems, preferring Lean methods. This supports findings that Lean can precede and support digital transformation (Tortorella et al., 2021; Naciri et al., 2022; Pirrone et al., 2024). Lean tools, such as Kanban and Just-in-Time, remain attractive due to their simplicity, low cost, and worker empowerment. AI, in contrast, demands high investment and technical expertise, making it less accessible for SMEs according to the studies of Shahin et al. (2024). Limited skills and infrastructure lead to delays, reliance on costly support, and hesitation (Oldemeyer et al., 2025). Consequently, SMEs often favour familiar Lean approaches requiring less specialised knowledge.

Respondents emphasised the irreplaceable value of human intuition, especially for RCA, challenging the notion of full automation in quality management (Muaz et al., 2025). This reflects ongoing debate (e Oliveira et al., 2023). Reluctance to use AI stems from concerns about validation and job loss (Li et al., 2025). Yet, e Oliveira et al. (2023) found that most authors support automating root-cause analysis for efficiency. Resistance mirrors trust and interpretability issues noted by Hickman (2025). Findings also support Schwaeke et al. (2025), who claimed SMEs adopt AI more readily in non-production areas. The reluctance in core operations underscores a cultural and operational conservatism.

Environmental factors, market volatility, customer and regulatory pressure, vendor reliability, and resource management were also influential. Strong interest in resource and energy optimisation reflects external shocks (Amarasinghe et al., 2023; Chen et al., 2023). As noted in prior work (Bevilacqua et al., 2017; Andrei et al., 2022; Shuford, 2024), AI drives energy efficiency. SMEs are progressing from basic systems to platforms using machine learning for energy optimisation. Implementing AI to schedule operations, detect anomalies, and select energy providers remains challenging.

Efforts to reduce environmental impacts through AI remain limited, with most firms relying on IoT and real-time monitoring for compliance. Yet studies (Beier et al., 2022; Adewuyi et al., 2024; Hasan et al., 2024) show AI enables proactive environmental assessment and predictive incident management. Respondents prioritised compliance and responsiveness over prediction, favouring traditional control systems. These mixed views echo Dubey et al. (2019), suggesting AI's sustainability impact depends on context.

Finally, the difficulty of accessing reliable AI providers and innovation hubs supports Del Giudice et al. (2021), who emphasised the value of external support networks. The lack of tailored SME solutions remains a significant adoption barrier.³

The findings suggest that AI adoption in manufacturing SMEs is most effective when focused on applications with immediate operational benefits, especially resource optimisation, energy efficiency, predictive maintenance, and production routing simulation. Managers should prioritise AI investments that align with strategic goals and measurable returns, while recognising that applications such as scheduling and root-cause analysis remain constrained by instability, limited data readiness, and the continuing importance of human expertise. Consultants and system integrators should therefore adapt solutions to SME realities, particularly legacy systems, skill shortages, and the need for phased implementation. Moreover, the study emphasises the importance of dynamic, generative AI applications that can simulate and adjust production processes in real-time. This points to a future where SMEs can leverage AI not only for efficiency but also for resilience and responsiveness. However, without addressing foundational issues, such as skill gaps, financial constraints, and cultural gaps, AI integration risks remain superficial. Thus, a phased, context-aware approach is essential, supported by an accessible innovation ecosystem and targeted capacity-building initiatives.

This study addresses two central research questions: (RQ1) How do manufacturing SMEs evaluate and prioritise AI applications in production management, and what factors influence their interest in specific technologies? and (RQ2) How do technological, organisational, and environmental challenges hinder the adoption of AI in SME production processes?

We found that SMEs exhibit a selective and pragmatic interest in AI applications. In response to RQ1, the findings reveal that SMEs prioritise AI applications that offer immediate and tangible benefits, particularly in areas such as resource consumption optimisation, energy efficiency, predictive and preventive maintenance, and production routing simulation. These applications are perceived as directly addressing operational efficiency and cost reduction, especially in the context of recent economic and geopolitical disruptions. Conversely, applications such as production scheduling optimisation, root-cause analysis, and predictive quality control are met with scepticism, often due to perceived complexity, lack of data readiness, or the belief that human expertise remains superior in these domains.

Addressing RQ2, the study identifies a range of technological, organisational, and environmental barriers that hinder AI adoption. Technologically, data quality, legacy systems, and integration challenges are persistent issues. Organisationally, the absence of strategic planning, limited internal expertise, and financial constraints significantly impede progress. Cultural resistance, fear of job displacement, and a lack of trust in AI systems further exacerbate these challenges. Environmentally, while external pressures such as regulatory compliance and market volatility drive interest in certain AI applications, they also contribute to a cautious approach to adoption.

This research introduces several novel contributions. First, it provides a granular, application-specific understanding of AI adoption in SME production management, moving beyond generic discussions to assess nine distinct AI use cases. Second, it highlights the divergence between expert expectations and SME perceptions, particularly in areas like scheduling and root-cause analysis. Third, it reveals the nuanced role of generative AI in facilitating dynamic simulations and adaptive maintenance strategies, indicating a shift from static to more autonomous and context-aware systems. Finally, the study proposes an actionable conceptual framework that integrates technological, organisational, and environmental dimensions, offering a structured pathway for SMEs to assess and enhance their AI readiness.

This study is primarily limited by its focus on Italian manufacturing SMEs, which may constrain the generalisability of the findings. National differences in industrial structure, digital maturity, and policy environments could lead to divergent AI adoption patterns in other countries. Additionally, the initial identification of AI applications was based on expert input, which may reflect specific professional biases. Future research should explore cross-national comparisons to assess how contextual factors influence AI integration in SMEs. Longitudinal studies could also track the evolution of adoption over time, while further investigation into the synergy between Lean practices and AI, as well as the role of innovation ecosystems, would provide valuable insights for both scholars and practitioners.