1 Introduction

In recent years, the integration of AI into the industry has gained increasing significance (Zhang et al. 2022). The advantages of AI solutions, such as improved production control and optimization, better prediction of machinery failures, and enhanced quality assurance, have driven companies to incorporate AI into their operational processes. Therefore, according to several researchers, AI is considered to be one of the most important elements of Industry 4.0 (Barton et al. 2022; Bencsik 2020). They argue that AI is the only way to make efficient use of the large volumes of newly acquired data (Szedlak et al. 2020). This is especially true in manufacturing (Bettoni et al. 2021).

However, the new potential of AI is mainly realized by multinational enterprises (MNEs) (Mittal et al. 2018; Teerasoponpong and Sopadang 2021). This is also reflected in a study by Szedlak et al. (2020) according to which around 90% of SMEs did not have any applications of AI. But in the past few years, an increasing number of SMEs are recognizing the potential of AI solutions. In survey of 2021, 77% of respondents said that they see great possibilities for using AI in SMEs as well. The same study showed that this potential is not limited to individual business activities but can be applied to almost all areas (Ulrich and Frank 2021).

With the increasing recognition and pursuit of AI implementation by SMEs, the issue associated with the impact of varying company sizes becomes more apparent. The existing insights and experiences predominantly focus on large enterprises (Welte et al. 2020). However, their applicability to SMEs is largely limited. The varying starting conditions of the different company sizes have a particularly substantial impact, especially when it comes to the implementation of AI. These differences include, for example, the financial capabilities of a company, the availability of data, or the desired complexity of an AI application (Agerri et al. 2014; Jain et al. 2021). Hence considering AI publications based on enterprise size provides a distinct advantage and therefore it is crucial to illustrate AI as closely aligned with a company's specific circumstances (Bauer et al. 2020; Hansen and Bøgh 2021). Besides the importance of the distinction according to enterprise size has already been highlighted in reviews on other new technologies (Cotrino et al. 2020; Mittal et al. 2018).

Nevertheless, there is a lack of research that explicitly addresses the specific characteristics and needs of SMEs when it comes to implementing AI applications (Bhuvaneshwari Alias Sunita Kulkarni and Mishrikoti 2019). This is also underlined by two identified open research gaps from previous publications: (i) “This study recommends more systematic reviews […] across business functions that could benefit more from the AI techniques such as accounting, quality management and human resources management “ (Cubric 2020, p. 11) and (ii) “literature research shows that no relevant studies exist that systematically address the issues and requirements of SME in the application of ML technologies “ (Bauer et al. 2020, p. 8).

Therefore, our goal is to use a systematic literature review (SLR) to analyze the current state of research on AI with an explicit focus on SMEs and to identify the perceived barriers of SMEs to AI implementation. We refer to AI holistically, without restricting our focus to specific AI methods. The overview of the state of the art aims to highlight the current deficit in this research domain. Simultaneously, it concretizes certain areas, for example regarding to business activities, that have been overlooked so far. Identifying the challenges faced by SMEs in AI implementation enables exact and targeted support from both academics and government in the future. This, in turn, facilitates a faster and more widespread adoption of this technology in smaller enterprises. At the same time, this raises awareness of the disparity in perceived obstacles among different company sizes. In our SLR, we focus on manufacturing companies in the industry. This focus is important due to the distinctiveness from other economic sectors such as consulting or services. These companies have entirely different starting conditions, for instance, in terms of data capture and potentials regarding AI.

From this, our two subsequent research questions arise:

  • RQ1. What is the current state of the art of research on AI in the industrial sector regarding SMEs?

  • RQ2. What are the challenges for SMEs in the implementation of AI in the industrial sector?

The remainder of this article is structured as follows: The next section deals with the theoretical background of the terms AI and SME. Afterwards, we describe the used methodology in detail before we present our findings to the research questions. Next, we discuss our outcomes including an interpretation and critical reflection. In the end, we summarize the results in a conclusion and highlight the limitations.

2 Background to the study

For a uniform and deeper understanding of our study, definitions and their historical development over the time are shown initially for the most important terms. This is, on the one hand, the concept of AI and, on the other hand, the categorization of company sizes. In the latter case, the focus is on the group of SMEs.

In 1956 the term “artificial intelligence” was proposed for the first time by McCarthy at the Dartmouth Conference (McCarthy et al. 2006), which today is often regarded as the birth of AI (Copeland 1993). However, it was not until the 1970s that the development and research of AI solutions began on a meaningful level (Marr 1977; Waterman and Newell 1971). But even then, the possibilities were limited due to the available data and the low level of technical development at that time.

But it is only with industry 4.0 that AI has gained momentum, as it led to a massive increase of technical devices generating, collecting, and making data available. This was the basis for complex and extensive AI applications. In addition, networking has made it possible to exchange large amounts of data with each other, and technical progress has enabled computers to use these large volumes of data effectively (Sousa et al. 2019). Since that time, more precisely from around 2010 onwards, the publications on AI have increased considerably. This applies to many different AI methods such as machine learning (ML), natural language processing (NLP) or computer version (Abioye et al. 2021). Nevertheless, in our study we focus on AI as a whole and refrain from emphasizing a selected AI method.

Nowadays, the utilization of AI has become widespread, whether it is interacting with a chatbot, doing a web search or using a voice assistant such as Alexa (Brill et al. 2019). AI can be found in almost all areas of life, although the development of AI is still at the beginning (Nikitas et al. 2020). However, the scope of applications is not only expanding in everyday life, but also in industry, in the field of education, in medicine and many more. This is due, among other aspects, to the wide multiplicity of research fields around AI. It is remarkable that almost 60% of the publications are assigned to the sector education and only about 5% focus on companies (Zhang et al. 2022).

Despite the intensive research in the field of AI, there is still no consistent definition. Rather, it is a collective term for applications that can solve tasks which previously required human intervention (Zhang and Lu 2021). A common definition by Hashimoto et al. (2018) states: “Artificial intelligence (AI) can be loosely defined as the study of algorithms that give machines the ability to reason and perform cognitive functions such as problem solving, object and word recognition, and decision-making.”

For a clear understanding of our study, it is crucial to recognize not only the term AI, but also the theoretical background of SMEs. As with AI, there is no standard definition for the grouping of company sizes. This led to more than 200 meanings for SMEs being already found in 1981 by Nieschlag (1981). These could be standardized in the following years, at least to a certain extent. For example, the European commission agreed on a common definition in 1996 (European Commission 1996). After its revision in 2005, all companies with fewer than 250 employees and either less than 50 million in annual revenue or an annual balance sheet total of less than 43 million are covered by the term SME (European Commission 2003). In this context, the European Union (EU) categorizes the term SME as follows: (i) micro enterprises (< 10 employees, < 2 million turnover), (ii) small enterprises (< 50 employees, < 10 million turnover), (iii) medium-sized enterprises (< 250 employees, < 50 million turnover). Although the definition distinguishes between micro, small and medium enterprises, in practice and in research the combined term SME is used almost exclusively (Gibson and van der Vaart 2008). However, many countries and organizations outside the EU define this term very differently. Berisha and Shiroka-Pula (2005) showed an overview to illustrate the difference in the definitions of SMEs based on the maximum number of employees. But a more detailed classification of large companies is requested also quite often. The European Commission (2022), for example, proposed a further category, the so-called middle-sized enterprises (or mid-caps), for companies with between 250 and 1.499 employees. In comparison to alternative definitions, the EU classification has become more established in the respective countries, leading to its utilization in diverse contexts such as the definition of limits for funding as well (European Commission et al. 2022). For this reason, coupled with the high prevalence of SMEs that possess the potential for AI applications in European (Riillo and Jakobs 2022), the definition provided by the EU serves as the foundation for the conducted SLR.

The importance of SMEs for a national economy cannot be underestimated. Although individual companies do not have a major impact, the large number of enterprises that fall into this group means that they are often referred to as the backbone of the economy (Kaymakci et al. 2022). The World Bank assumes that, depending on the definition of SME, at least 40% of national income (GPD) is generated by SMEs in emerging countries (World Bank 2015). At the same time, 50–60% of the world’s population work for SMEs (Ben Abdelaziz et al. 2020).

In relation to AI, the progress of development is not only important for large enterprises anymore, but also for SMEs. A substantial contribution to this has been made by the substantial reduction in the costs of developing or introducing an AI solution. For example, according to Zhang et al. (2022) the costs of an image classification system fell by almost 64% from 2018 to 2022. As a result, the applications are also becoming increasingly interesting for companies with lower financial resources. Furthermore, Barton et al. (2022) presented that, in addition to big data, AI is also a key attribute for SMEs in the context of Industry 4.0. The potential is not limited to individual areas. Instead, AI can support applications in almost every department. Nevertheless, various studies concluded that AI is being used so far in less than 10% of SMEs (Bauer et al. 2020; Iftikhar and Nordbjerg 2022; Szedlak et al. 2021).

This background is the starting point of our SLR, which shows that both AI and SMEs have an important role to play in successful economic development. The following selection describes in detail the used research method, the process, and the scope of the study.

3 Research methodology

To answer the research questions, we conducted a systematic literature review based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) protocol (Moher et al. 2009). A systematic review highlights the current state of the start of a specific field, shows agreements and disagreements of different studies, and identifies research gaps (Kitchenham 2004). Based on its transparency and replicability, the systematic approach was selected (Leidner and Kayworth 2006), which enables to verify the results in an easy way and reduce the risk of selection bias (Knobloch et al. 2011).

The PRISMA protocol divides the process in four steps: (i) identification, (ii) screening, (iii) eligibility, and (iv) inclusion (Moher et al. 2009). The identification phase is needed to define the scope of the research. To get an overall overview, it is necessary to include all appropriate studies that can help to answer the research questions (Keele 2007).

For this reason, in a first step we analyzed the used search strings of 77 systematic reviews about AI (a lineup with the reviews can be accessed at: http://bit.ly/41ac4lX) and the most frequent terms of the Artificial Intelligence Index Report 2022 of Zhang et al. (2022). We adopted all keywords that we found in at least five percent of the reviews (see Table 1) and the two most common terms of the index report in our own search strategy. Based on our research questions, we defined two further keywords groups, one by enterprise size and one by sector. Due to the already shown limited prevalence of AI publications concerning SMEs and the resulting lack of research (Žigiene et al. 2019), we intentionally adopted a broad approach in selecting keywords. Furthermore, in our search, we intentionally refrained from adding a further category of search strings related to challenges. This decision was made to include case studies as well that may not explicitly address challenges in their title or abstract but nonetheless highlight barriers during the description of their case study. Therefore the search was conducted with following terms: (“Artificial intelligence” OR “machine learning” OR “deep learning” OR “neural network” OR “AI” OR “natural language processing” OR “data mining” OR “big data” OR “robot” OR “chatbot” OR “machine intelligence” OR “reinforcement learning” OR “deep mining” OR “data science” OR “support vector machine” OR “algorithm” OR “machine version” OR “training data” OR “dataset”) AND (“SME*” OR “small and medium enterpris*” OR “small and medium-sized enterprises “ OR “small and medium businesses “ OR “small and medium-sized businesses “ OR “small and medium companies “ OR “small and medium sized companies “ OR “micro small and medium enterprises “ OR “micro small and medium sized enterprises “ OR “MSME “ OR “micro enterpris* “ OR “micro-sized enterpris* “ OR “micro compan* “ OR “micro-sized compan* “ OR “micro business* “ OR “micro sized business* “ OR “micro firm* “ OR “micro-sized firm* “ OR “small enterpris* “ OR “small-sized enterpris* “ OR “small compan* “ OR “small-sized compan* “ OR “small business* “ OR “small sized business* “ OR “small firm* “ OR “small-sized firm* “ OR “medium enterpris* “ OR “medium-sized enterpris* “ OR “medium compan* “ OR “medium-sized compan* “ OR “medium business* “ OR “medium-sized business* “ OR “medium firm* “ OR “medium-sized firm* “) AND (“industrial” or “manufacturing” or “secondary sector”).

Table 1 Search strings AI
Full size table

The databases Scopus, Web of Science, EBSCO, and ProQuest were used. All of them comply to the quality requirements for a suitable academic search system of Gusenbauer and Haddaway (2020). To eliminate duplicates, the software Mendeley Reference Management was utilized after downloading the publication title, author, and abstract from the databases.

In line with PRISMA, the following inclusion criteria were applied: (i) focus on AI in industrial sector by small and medium companies, (ii) publications as of 2010, as suggested by Hansen and Bøgh (2021), and (iii) written in English. On the other hand, the exclusion criteria were also considered, which are the following: (i) papers not specifically related to AI but to technology in general, (ii) “false positive” papers, for example papers that use the acronym SME for smart manufacturing environments (Yan et al. 2017) and (iii) full article not available. For a sufficient quality of the included articles, the following conditions were taken into account: (i) only journal articles or conference papers and (ii) only peer-reviewed publications.

These criteria were used on the title and abstract in the screening phase and on the full text in the eligibility phase. In total, 71 papers met the criteria and were therefore included in the systematic literature review. As shown in Fig. 1, initially the result of the keywords search was overall 1395 records. After removing 335 duplications, 1060 titles and abstracts were read. 911 of them did not comply with the required criteria, so that a full text evaluation took place for 149 papers. There 78 research works were excluded which finally results in 71 selected papers. The predominant reason (947 articles) for excluding the publications was the lack of focus on AI. The high number can be attributed to the broad search strings related to AI, as some terms only have an indirect connection to the AI theme. To gain a comprehensive overview of AI publications in SMEs, these publications were initially selected during the search but were excluded during the screening or eligibility phase if no connection to AI was identified, or if AI only served as an example. An additional 23 publications were eliminated due to 'false positive' abbreviations that had no relevance to AI or SMEs. The reason for the exclusion of the remaining 19 articles from our SLR was the unavailability of the full article.

Fig. 1
figure 1

PRISMA flow chart

Full size image

In consideration of the research questions, the data extraction was carried out to highlight the key findings of the separate articles, as recommended by Kitchenham (2004). For this reason, for each paper the following characteristics were recorded into Excel: (i) title, (ii) topic, (iii) source of publishing, (iv) year, (v) study design, (vi) method of AI, (vii) function based on Porter’s value chains (Porter 1985), (viii) definition of SME, and (ix) challenges for AI adoption regarding to SMEs. With the help of charts and spreadsheets commonalities and contrasts, research gaps and new ways of looking were elaborated (Torraco 2005). Wohlin et al. (2000) emphasized the risk of biases of (i) construct validity, (ii) external validity, (iii) internal validity, and (iv) conclusion validity in a SLR. To reduce these as far as possible, a particular attention was paid to the biases in the whole processing phase. Furthermore, the structure was chosen very precisely, for example the choice of databases according to Gusenbauer and Haddaway (2020). Nevertheless, a complete avoidance of the biases is not possible. For example, the focus of most publications is on case studies and on enterprises that already dealt with AI. Only few articles included data of businesses and people without any points of contacts to AI. Moreover, the evaluations of the results and patterns of the conducted studies are often a limitation. Due to the RQ2, special attention was paid to the challenges. Therefore, we counted in how many and in which publications the barriers are mentioned. The classification of the obstacles was orientated on the PESTEL model, which divides the categories into (i) political, (ii) economic, (iii) social, (iv) technological, (v) environment, and (vi) legal (Yüksel 2012). The PESTEL model furnishes a comprehensive and systematically organized overview of the challenges identified by companies. This categorization has demonstrated its efficacy on multiple occasions, even when applied to the internal barriers of an enterprise, as evidenced in different publications addressing other new technologies (Alaloul et al. 2020; Oesterreich and Teuteberg 2016).

4 Results

This section presents the main findings, which can be divided into two groups: (i) characteristics of the studies, and (ii) challenges of SMEs in the implementation of AI.

4.1 Characteristics of the selected studies

First, the most important characteristics of the selected publications were analyzed to present the current state of the art of research on AI in the industrial sector regarding SMEs and to answer RQ1. The assignment of the papers to the distinctive features for each category can be accessed at: http://bit.ly/41ac4lX.

4.1.1 Chronology of the research

Initially, the selected studies were examined according to the year of publication. Thereby three phases can be recognized. In the initial phase spanning from 2010 to 2017, just one article per year was published about AI and SMEs in the industrial sector, indicating that this topic was confined to a niche domain during this period. A reason is the customary initial examination of emerging technologies in a broad, general context. Only over time, a more nuanced evaluation will develop, with a focus on specific attributes such as company size (Welte et al. 2020). With the increase in publications on AI in manufacturing companies in the following years (Rathore et al. 2021), publications on AI in SMEs also increased in the second phase (2018–2020). Smaller companies are now also beginning to address this topic more in individual cases (Kim et al. 2019b). However, initial experience has shown the need for a separate research focus on the differences between enterprise sizes due to distinctive initial situations (Barton et al. 2022; Mittal et al. 2018), which is reflected in the increase in corresponding publications in the second phase. In the third phase of 2021, there has been a sharp increase, almost tripling compared to the previous year. This surge is attributable to the expanding media prominence of AI and the advancement of AI applications, progressively tailored to the requirements and starting conditions of smaller enterprises (Teerasoponpong and Sopadang 2021). Consequently, these companies are increasingly recognizing and using the potential of AI (Zhang et al. 2022). This has led to a substantial rise, notably observed in the heightened number of published case studies. Ongoing technological advancements coupled with the expanding visibility of AI in the public imply the likelihood of a sustained increase in this trend in the forthcoming years (Fig. 2).

Fig. 2
figure 2

Number of publications over the years

Full size image

4.1.2 Categorization according to study design

When considering the methodological design, first a distinction is made between qualitative, quantitative, and mixed methods. Thereby, the strong dominance of qualitative studies is remarkable. Over three-quarters (55) of the articles used this approach. One reason for the overweight of the qualitative studies is the heaping use of case studies. 33 of the selected articles have chosen this approach among others. This can be ascribed, in part, to the limited prevalence of AI adoption within SMEs (Willenbacher et al. 2021). While case studies can be conducted and analyzed in technologically pioneering companies, the implementation of quantitative methods, such as surveys, is more difficult to implement due to the high demand for participating SMEs. This arises from the observation that numerous potential companies have not yet sufficiently engaged with the AI domain, making it difficult for them to participate in surveys (Seseni and Mbohwa 2018). This underscores as well why many of the case studies described the AI solution from a view of information technology (IT), for example, to explain the algorithm in detail. In the nascent stages of novel technology, the engagement primarily emanates from IT-proficient enterprises and researchers, directing their attention toward the technical factors (Putnik et al. 2021; Singh and Desai 2022). At a later juncture, broader perspectives as business aspects, subsequently are added.

4.1.3 Classification of the publications regarding to business activities

The thematic classification of the papers into business activities was undertaken according to Porter’s value chain. It divides a company into primary activities, which includes (i) inbound logistic, (ii) operations, (iii) outbound logistics, (iv) marketing and sales, and (v) service as well as secondary activities with (vi) procurement, (vii) human resource management, (viii) technological development and (ix) infrastructure (Porter 1985). This facilitates an analysis across the entire process chain of a company, incorporating all crucial activities (Stabell and Fjeldstad 1998). As a result, activities that have been overlooked or underrepresented so far, become conspicuously visible.

Several of the selected papers could not be assigned specifically to one of the activities, because these addressed the topic of AI for the entire enterprise. The pronounced emphasis observed in the remaining articles (44) on operational aspects is noteworthy, with other activities, aside from technological development (7), either mentioned only occasionally or not mentioned at all. This shows that, to date, only a fraction of the diverse activities within a business concerning AI in SMEs have been scrutinized through publications, revealing substantial research gaps. For the predominant focus on operations, various reasons are given. On the one hand, this domain in manufacturing enterprises typically constitutes the core area with the highest number of employees. Consequently, the optimization potential afforded by AI is frequently the most substantial there (Barton et al. 2022). On the other hand, this area often includes the same or very similar activities, making it easier to change something and enable stronger control through key figures. This enhances transparent to the weak points and the success of AI adjustments easier to measure (Karl and Reinhart 2015) (Fig. 3).

Fig. 3
figure 3

Distribution of publications across Porter´s Value Chain

Full size image

Due to the large number for operations, Fig. 4 refines the subdivisions for this activity. Almost half of these papers (12) dealt with the optimization of the quality of the product. Regardless of the size of the company and the specific initial situation, his concern is pertinent to every manufacturing enterprise (Durana et al. 2019). This is where the quality differs from machine optimization or predictive maintenance, for instance, which is irrelevant for small companies with no or only old machines (Baars et al. 2021). A similar consideration extends to production planning when conducted within the company without the use of technical data. Consequently, quality appeals to a considerably broader target group within the context of SMEs.

Fig. 4
figure 4

Distribution of publications with the focus on operations

Full size image

4.1.4 Analysis based on the utilized SME definitions

About AI, the different company sizes play a key role in the point of view of this topic and the perceived challenges for implementation. SMEs often have completely different starting conditions, such as a lack of expertise or data availability, than large companies (Agerri et al. 2014). As a result, various scholars see the need or an added value to differentiate by company size regarding AI (Bauer et al. 2020; Hansen and Bøgh 2021). Nevertheless, the customary categorization of SMEs and large companies between distinctive countries substantially diverges owing to distinct definitions. In the EU, for instance, SMEs encompass companies with up to 250 employees, while the United States of America definition extends to companies with twice as many employees (Anastasia 2015).

For this reason, the selected papers were analyzed by the given meanings of the term SME (see Fig. 5). Despite the varying understanding, only a few articles outlined the criteria for their applied definition. Most of them (61) simply used the term without any explanation. When the criteria were explained, the majority (7) of the articles were based on the regulations issued by the EU. Two more publications defined the limit at 500 employees, and one was orientated at the regulations issued by the Chinese government. The rationale for the absence of the used definition of SME in many of the selected articles is generally not provided.

Fig. 5
figure 5

Number of publications by means of SME definition

Full size image

4.2 Challenges regarding AI implementation in SMEs

This section explains the obstacles for SMEs in the implementation of AI in the industrial sector, see RQ2. In the process, we have identified 27 different challenges mentioned in the papers. This shows the diversity of the challenges and the associated complexity of removing these obstacles. It is not merely one challenge that requires addressing but, rather, an interplay of many different factors. For a better overview, the barriers are summarized and structured in line with the PESTEL model. This involves presenting both the number of distinct articles addressing a specific challenge and the number of different articles of each PESTEL category. For the evaluation in categories, each paper is only counted once for a group. The results are presented in Table 2.

Table 2 Challenges for SME (S)
Full size table

The predominant challenge identified across the selected papers is knowledge. In 35 articles, this obstacle is mentioned and thus much more often than all other. Nevertheless, the authors of the selected studies define different challenges as the most important obstacle to the introduction of AI in SMEs. While many researchers go hand in hand with the absolute number and see knowledge as the most decisive hurdle (Kumar and Kalse 2022; Ulrich and Frank 2021), there are also other publications of the review that consider the costs (Barton et al. 2022), the acceptance (Joerg and Carlos 2022), or the data (Bauer et al. 2020) to be the most crucial factor. Concerning knowledge, two principal domains can be delineated. Firstly, the discourse addresses the scarcity of skilled employees (Basri 2020; Brezani et al. 2022), and secondly, it encompasses the general understanding of AI, such as knowledge about the procedure for AI implementation and knowledge about existing AI application possibilities (Brandalero et al. 2020; Welte et al. 2020).

The second most common reason for the hesitation to implement AI in SMEs are the costs (number of articles that addressed this challenge: 24). This is unanimously attributed to the limited financial resources, which are generally much more limited in SMEs than in MNEs (Chen et al. 2019; Willenbacher et al. 2021). Another complicating factor for SMEs lies in their often-limited awareness of their financial capacities or an inability to accurately assess them (Žigiene et al. 2019). This also has consequences for the knowledge challenge, as SMEs are often unable to afford external AI consultants (Szedlak et al. 2020).

The maturity level within their own organization stands out as the third most frequently referenced barrier (17) for SMEs when embarking on AI adoption. Thereby researchers concurred that most SMEs have a substantially lower maturity level of digitalization than MNEs (Bettoni et al. 2021; Teerasoponpong and Sopadang 2021). The great majority of SMEs are only in level one. Mittal et al. (2018) even introduced another maturity level (level zero) to emphasize that some SMEs are at the very beginning of digitalization. As an example, the authors mentioned the lack of wireless fidelity (WiFi). The low maturity level respectively the missing IT infrastructure are therefore an additional challenge for SMEs. Szedlak et al. (2020) pointed out that around 70% of the SMEs without AI ambitions cite the lack of digitalization as a reason.

The particular AI-specific challenge pertaining to data availability (16) is also discussed intensively. Its importance attributed to SMEs is thereby viewed very differently in the selected publications. Some scholars see a clear lack of data in SMEs that needs to be rectified before implementation (Agerri et al. 2014; Putnik et al. 2021) meanwhile other researchers believe that the current level of data in many SMEs is already sufficient (Chen et al. 2019). While the former emphasizes that AI cannot be viewed separately from big data as they only make sense together (Kim et al. 2019a), the latter highlights case studies of successful AI implementations in SMEs that are based on little data at or at least only a small amount of their own data (Kulkarni et al. 2021; Thiagarajan et al. 2018). Thereby they underline the possibility of collecting data together with other companies or introducing AI applications that have been trained on external data.

14 different papers considered management to be a decisive factor for the implementation of AI. In contrast to the predominantly cited barrier of knowledge, fewer than half as many publications identify management as an impediment. Nevertheless, this alignment Barton et al. (2022) survey, underscored management as a significant hindrance to AI adoption in SMEs. Only with their support can the new technology be successfully introduced (Iftikhar and Nordbjerg 2021). The management's hesitation is based on various reasons. According to Bunte et al. (2021), many leaders think that their business is too small for AI. Others cannot imagine any areas of applications (Husson et al. 2021) and others again do not see the need for a change, especially when the enterprise is doing well (Joerg and Carlos 2022).

Besides productivity evaluation is a further often mentioned challenge for SMEs (13). Many enterprises are often unsure about the return on investment (ROI), because they cannot assess the consequences of an implementation (Prem 2019; Žigiene et al. 2019). The given reasons can be categorized into two groups, on the one hand into factors that apply regardless of the size of the company and on the other hand into factors that have a major influence on SMEs in particular. General difficulties in assessing the ROI are the uniqueness and the difficulty of calculating the customization costs of many AI solutions for your own company (Bunte et al. 2021). This complicates the derivation of insights from the experiences of other companies. Numerous AI solutions also necessitate multiple attempts and approaches before achieving functionality (Hansen and Bøgh 2021). This further complicates the anticipation of the implementation timeline and, consequently, the realization of cost-saving effects. Moreover, for smaller enterprises, there exists the additional challenge that the investment costs of an AI solution often need to be distributed across a lower quantity of sold products. Therefore, each must bear a higher share of it (Husson et al. 2021). Furthermore, the current absence of publications featuring case studies about the introduction of AI in SMEs that calculate the ROI also makes it difficult for these companies to estimate the ROI in advance (Iftikhar and Nordbjerg 2021).

Next to the amount of data, many researchers (12) emphasized the data quality as a demanding requirement as well. This holds particular more importance to SMEs compared to MNEs, given that SMEs often do not have any data collection standards or ways to clean up data trash (Lu et al. 2022). Hence the preparation of the data would be a time-consuming task for many SMEs. For this reason, it is difficult to use these for an AI application, even if enough data are collected (Jain et al. 2021). The importance of data quality should not be underestimated, as redundant and unclear data has already caused implementation projects to fail in selected case studies (Bender et al. 2022).

Compared to other technologies, AI has a much higher complexity. This is another important challenge that is often (11) seen in SMEs regarding an implementation. The perceived complexity results from three causes. First compared to other technologies, AI is considerably more IT-intensive, and there are more prerequisites that must be fulfilled, such as the existence of high-quality data (Bender et al. 2022). Secondly, AI applications are frequently challenging to transfer from other companies or necessitate extensive customization, thereby augmenting the perceived complexity (Kant and Johannsen 2022; Lu et al. 2022). The more the companies differ from each other in their structures or sizes, the greater the effort required to adapt (Seseni and Mbohwa 2018). Thirdly, this perceived intricacy is increased by highly technical case studies on AI in SMEs, for instance, with detailed descriptions of individual algorithms. Only rarely publications simplified to an understandable level for managers without IT expertise (Willenbacher et al. 2021).

As already described, AI implementation is often a trial-and-error process with an unclear outcome at the beginning (Prem 2019). In addition to the resulting difficulty in determining the ROI, many scholars (10) also see the risk of failure as a challenge, especially for SMEs. Since these enterprises are often unaware of their own risk profile and also have fewer financial resources, failed attempts are a major problem that strongly influence the company’s performance (Willenbacher et al. 2021). For this reason, those enterprises rely more frequently on mature technologies, which prevents or slows down the adoption of AI in SMEs at the moment (Bettoni et al. 2021).

In addition to the challenges in their own companies, various studies have also identified barriers on the market side (8). The available AI solutions that explicitly focus on the requirements and initial situations of SMEs are very limited (Kaiser et al. 2021; Velmurugan et al. 2021). Most applications are aimed at large companies and are tailored to their structures. Although there are now initial approaches for SMEs (Marco et al. 2021; Singh and Desai 2022), these are still in the clear minority. This is particularly contradictory to the fact that there are significantly more companies falling under the SME definition than there are larger enterprises (Kaymakci et al. 2022; Perera and Chand 2015). However, one example of an AI solution with an explicit focus on SMEs is developed by Brezani et al. (2022). In this case study, a program was designed to detect objects on camera images with the help of AI. During the development phase, special attention was paid to ensuring that this AI program can be adopted by other SMEs without any major adaptation effort.

Besides the frequently cited challenges, certain impediments to AI implementation in SMEs received limited attention in the selected publications. These include the obstacles of acceptance (6), trust (3), data security (7), and data privacy (4). First acceptance and trust, play a minor role in AI implementation in SMEs in the selected publications, although, Joerg and Carlos (2022) considered acceptance to be the most important challenge for SMEs and Kumar and Kalse (2022) emphasized that the Technology Acceptance Theory (TAM) must be taken into account. Another conspicuousness is the issue of data security and data privacy, but only seldom addressed in the selected articles about SMEs. Scholars attributed this to the generally less attention of SMEs to these topics, given the lower volumes of data and limited data utilization until the present (Empl and Pernul 2021). However, some authors assume that this will also change in SMEs with the introduction of AI and the associated collection and evaluation of large amounts of data. As a result, these challenges will become more important as smaller companies engage more intensively with the topic of AI (Joerg and Carlos 2022).

If the challenges are viewed from a higher level, based on the categories of the PESTEL model, a bisection is remarkable. The categories economic, social, and technological are weighted much higher than the groups of political, environmental, and legal. The category with the most mentions of challenges is economic (68 challenges in 36 different publications). However, the social (58/36) and technological (66/33) barriers are almost identically high. On the contrary, the remaining three categories political (8/5), environmental (3/3), and legal (1/1) are also quite close to each other, but on a variant scale. The observation that the three predominant challenges—knowledge, costs, and IT infrastructure—originate from three separate categories (social, economic, technological) further exemplifies the diversity and complexity inherent in addressing barriers to AI implementation in SMEs.

The political challenges can be split into (i) lack of advice, (ii) lack of research, and (iii) lack of funding. Even if all of them are rather less discussed, Žigiene et al. (2019) highlighted the importance of solving these barriers. Furthermore, within the political context, the lack of advice opportunities (3) and academic research (3) are currently seen as more of a challenge than the lack of funding (1).

In the economic area, the barriers are (i) cost, (ii) productivity evaluation, (iii) risk to fail, (iv) solutions available on the market, (v) lack of resources and machines, (vi) data exchange, (vii) lack of standardization, (viii) process adjustment, (ix) duration of implementation, and (x) project planning. Costs emerge as the most frequently cited economic challenge, with noteworthy emphasis on the fact that the second (ROI) and third (risk of failure) most prevalent obstacles are also intricately connected to financial considerations. This underscores its meaning for an AI introduction (Bunte, 2021), simultaneously highlighting a disparity with the rare mentions of the necessity for action in political funding.

The two least frequently mentioned challenges in the economic category, project planning and duration of the project can be assigned to project realization. This shows that the concrete execution of an AI project in SMEs is viewed rather uncritically (Teerasoponpong and Sopadang 2021).

The social difficulties in AI adoption are (i) knowledge, (ii) management, (iii) acceptance, (iv) AI corporate strategy, and (v) trust/ethics. In this category, in particular, two challenges are emphasized: the knowledge required and convincing management. However, knowledge is by far the most common social challenge for SMEs (Bauer et al. 2020). In contrast, moral and ethical aspects are rarely seen as an obstacle (Lu et al. 2022).

Besides (i) IT infrastructure, (ii) data availability, (iii) data quality, (iv) complexity/individuality, (v) data security, (vi) data privacy, and (viii) model/method/application selection, there are the technological obstacles. This category is more balanced between the individual barriers than the others. All appeared in at least three different articles. The focus of challenges in the technological domain is more on general barriers that must be met irrespective of specific AI applications (data, IT infrastructure), rather than application-specific challenges such as complexity or the selection of individual AI solutions (Iftikhar and Nordbjerg 2021). Overall, this category is strongly characterized by challenges related to data. More than half of the listed challenges can be assigned to this topic. In this context, the focal point primary centers in data generation (Kim et al. 2019a).

In the environment category, only (i) environmental protection is depicted. Conspicuously, the only perceived hindrance concerning the environment is the high energy requirements of the IT hardware for the procedure of AI solutions (Kant and Johannsen 2022). Other potential environmental barriers remain unaddressed.

The last category is legal with (i) regulatory compliance. Due to the large amount of data collected and used, AI poses new legal challenges for the companies. For many, there is still no satisfactory solution how to deal with it (Lu et al. 2022). As already described, many SMEs have not looked at any or only at uncritical data until now. With the expansion of data collection and analyses that inevitably accompanies the implementation of AI, SMEs must now, in addition to a stronger focus on the acceptance of data use and data security, also increasingly deal with the regulatory requirements (Žigiene et al. 2019).

5 Discussion and future research opportunities

The literature review shows the most important characteristics of the selected studies as well as the challenges of AI implementation for SMEs. These barriers and additional research gaps slow or prevent widespread AI adoption. Therefore, this section identifies and discusses potential solutions, practical implications, and open research questions. A summary of the results is shown in Table 3.

Table 3 Implications and research gaps
Full size table

5.1 State of the art of research on AI in the industrial sector regarding SMEs

One of the main results of our analysis of the state of the art research is that the subject area of AI with an explicit focus on SMEs is still underexplored. This can be attributed to various factors. First, there are only a few studies on this topic to date. Despite extensive keyword searching, only 71 publications were found for our SLR. Compared with the number of publications on AI in MNUs (Rathore et al. 2021), it shows that this topic area is still backward, despite the identified increase since 2021. This supports the statement by Hansen and Bøgh (2021) and Brandalero et al. (2020) that the overwhelming focus on AI studies to date has been on larger companies. The lack of studies also has practical impacts on SMEs. Although the effect of AI on a company’s competitiveness may not be decisive in most cases today, it is anticipated that this will change in the future. As the implementation of AI is a lengthy process (Prem 2019), it is crucial to initiate action now. Thereby academic research is the first step. Hence, we urge scholars to focus more on the topic of AI in SMEs, taking into consideration its unique characteristics in comparison to MNEs.

Furthermore, the analysis according to the business activities by Porter (1985) underscores the deficiency of AI papers for SMEs. Notably, there is a pronounced emphasis on the operational domain, while there is an absence of publications in numerous other business areas (see Fig. 3). Against this background, it is essential to acknowledge that a comprehensive overview of possible AI applications from different business areas holds great importance for enterprises. The current state of research on AI and SMEs conveys the impression that AI is mainly applicable in the domain of operations, leading to a perception of limited potential for AI development in other areas. Hence, there is a need for additional research papers on AI and SMEs that specifically focus on the underrepresented areas (Dobler et al. 2020; Žigiene et al. 2019). Given the recognition of the prevailing concentration of case studies on the operational domain, practical implications for businesses can also be derived. This provides SMEs that are newly delving into the field of AI with an initial orientation on where to commence within their enterprise. Even though the potential of AI is very company specific, this offers entrepreneurs at least an initial starting point. Alongside the previously highlighted advantages of AI in the operational domain, such as the high number of employees and the frequently repeated activities (Bunte et al. 2021), we therefore recommend that manufacturing SMEs initially focus on operational activities, when starting with AI.

Another reason for the need for additional publications on AI in SMEs are the very technical case studies to date. Most of them described their approaches and results from an IT expert point of view (Kiangala et al. 2022; Lin et al. 2021; Manmohan and Shalij 2022). However, since the focus is on SMEs and they usually have less specialized knowledge in their company, especially in the IT area, it is debatable whether the case studies are too complex for many users. This applies particularly to publications that explain the program code or compare different algorithms (Brillinger et al. 2021; Murphy et al. 2019). We urge future researchers to evaluate the purpose and necessity of including detailed technical descriptions in publications about SMEs, and, whenever possible, to make case studies more comprehensible for a broader target audience or to explain the technical specifications in a separate section. In this way, more people can be reached, as their openness to AI is not negatively affected by the complexity. In line with Szedlak et al. (2021), we also perceived the risk that it could give otherwise the false impression that AI is only manageable for software engineers or at least IT experts. This would inhibit the acceptance and thus also the spread of AI in SMEs. Nonetheless, our SLR also highlighted examples where the technical aspects are not the primary focus (Seseni and Mbohwa 2018). These can serve as an entry point into the AI topic for entrepreneurs with limited IT expertise.

In the case studies it is important to critically question the extent to which these solutions and their benefits can be transferred to other use cases. In the field of AI in particular, researchers widely agree that it is often difficult to transfer or copy solutions that have been developed (Ersöz et al. 2022; Kant and Johannsen 2022). Rather, the very different initial situations, such as the amount of available data, cause the need for individual requirements and conditions. Consequently, in upcoming research papers utilizing a case study, it is vital to establish a clear and concise depiction of the initial situation. Moreover, researchers should highlight which aspects of their case studies are transferable to other companies and should specify the necessary conditions for such applicability. For businesses, this implies that many AI applications and case studies can be adopted only with often substantial customization efforts. Therefore, it is crucial to influence decision-makers to understand that the presented AI solutions typically cannot be directly adopted in a 1:1 manner. Otherwise, companies may recognize these divergent starting points too late or misjudge their impacts, resulting in a flawed business case for the AI applications.

It has already been stated that the selected papers generally only distinguish between SMEs and other company sizes, whereby the exact term of SME often remains undefined. In other areas of Industry 4.0, this might not be critical, but AI is substantially different compared to other technologies (Jöhnk et al. 2021). A major influence on an AI implementation arises from the initial conditions of the company, including factors such as available data and maturity level. For this reason, companies should always consider these factors in case studies whenever possible, especially when drawing conclusions from the publications for their own business. To enable companies to do this we believe that it is needed to clearly define the company sizes considered and show the starting conditions of the enterprises in the context of AI publications. Only in this way is it possible to differentiate the results. But the right level of detail and the right key figures are questionable, as some researchers have already shown (Loecher 2000). In addition to the widespread option of revenue and/or number of employees, scholars dealing with AI often distinguish according to the maturity level (Bauer et al. 2020; Mittal et al. 2018). Defining the most appropriate key figure for a subdivision still represents a research gap that would have to be investigated in more detail in the future.

5.2 Challenges for SMEs in the implementation of AI in the industrial sector

A main finding of our SLR is the great number of different challenges for an AI implementation by SMEs. The result of 27 identified challenges is surprising, as many case studies only focus on a few most important barriers (Kim et al. 2019b), and this diversity clearly passes the number of identified obstacles in reviews of AI in MNEs (Arinez et al. 2020; Regona et al. 2022). This indicates, on the one hand, that SMEs perceive substantially more challenges in AI implementations than larger enterprises, and, on the other hand, underscores the complexity and individuality inherent in AI applications. This is further supported by the researcher’s disagreement on the definition of the biggest challenge for an AI implementation by SMEs. Depending on the specific use case, different obstacles can be perceived as the greatest barriers to implementation. If, for example, the required data could be the biggest barrier for a complex ML solution (Bauer et al. 2020), acceptance may be the largest obstacle for a simple application that processes personal data. This renders standardization of the challenges and their prospective solutions arduous. Thus, companies should adopt an approach that considers the concrete AI application, evaluating the challenges and prerequisites associated with the specific use cases to be implemented. In this context, we recommend considering our identified challenges as guiding principles to avoid overlooking or underestimating potential barriers. Simultaneously, our SLR shows that due to the individuality of AI applications, there is no necessity for companies to satisfy all 27 identified challenges before an implementation. For companies with low data availability, for example, AI applications that precisely take these initial conditions into account and are designed accordingly can be interesting. Case studies from our SLR present this possibility (Kim et al. 2019b). For SMEs, this implies that even if their own company does not meet all usual prerequisites for an AI implementation, it does not necessarily mean a general unsuitability for AI. Rather, these companies should not be deterred by this and explore what is already achievable given their initial starting conditions.

When considering the challenges, the most common implementation barriers for SMEs are knowledge, costs, and IT infrastructure/digital maturity. It is noteworthy that these three most frequently cited obstacles can be assigned to three different PESTEL categories: (i) knowledge to social, (ii) costs to economic, and (iii) IT-infrastructure/digital maturity to technological. In our opinion, this shows that from an economic and political point of view, the dismantling of challenges for SMEs must take place in a triad. It is not enough to focus on just one category or even on one single obstacle. Instead, the barriers, especially economic, social, and technological, must be viewed and solved as a whole. Future political decisions regarding the support options for SMEs in AI adoption should be made based on these findings. A parallel improvement in the conditions across all these diverse categories is expected to have the most positive impact on the increase in the spread of AI in SMEs.

The three major economic challenges, cost, productivity evaluation, and risk of failure, can all be attributed to the topic of finance. This underscores the importance of financial aspects in the implementation of AI in SMEs. However, there is a disparity to the barely noticed political barriers. The authors of the selected studies see only an occasional need for action about the political obstacles to AI (Basri 2020), even though political support, especially through funding, has proven effective in overcoming barriers and expediting the dissemination of other new technologies (Yang et al. 2022).We assume that many companies perceive the financial barriers of adoption but may not be aware of the available support options. Therefore, the government should enhance the visibility of existing funding opportunities for SMEs and, if necessary, expand them accordingly. Nevertheless, companies should initially not view the financial barrier as a reason to forego AI solutions. Firstly, they should consider potential support offerings in their deliberations. Furthermore, case studies have already demonstrated that individual and efficient AI applications can be introduced with a budget of less than $500 (Kim et al. 2019b).

Another noticeable aspect, regarding the financial obstacles, is the frequent mention of productivity evaluation (quantity: 13). This suggests that it is not only the absolute costs of implementation, but rather the difficulty of assessing whether an investment in AI may have a good ROI. In all selected case studies, the calculation of the productivity evaluation is completely out of scope without further justification. Only Kim et al. (2019b) listed the exact costs for their case study. Given that the ROI is a crucial factor for SMEs when contemplating AI adoption (Iftikhar and Nordbjerg 2022), future case studies should address this aspect as well. However, since case studies are always very specific, there is also a necessity for additional frameworks customized for SMEs to determine the ROI of an AI solution in a way that is as simple and universal as possible. In the absence of a comprehensive framework for calculating and assessing the potential benefits, numerous companies tend to abstain from initiating AI implementation, or alternatively postpone the decision-making process, pending further experience-based insights. This reluctance is often due to the fact that SME managers without suitable aids lack the capacity to adequately estimate the potential benefits of AI solutions (Bunte et al. 2021). The development of a corresponding framework would enable the calculation of the benefits of AI applications, ultimately fostering greater acceptance and diffusion of the technology within companies.

The challenge of management also holds an important weight in the context of an AI introduction. This is particularly pronounced in SMEs, firstly, because decision-making authority rests especially with management, and secondly, due to the larger impacts of an AI implementation on the financial aspects of a smaller company. Therefore, we consider the conviction of management in AI technology as a prerequisite for the widespread adoption of AI in SMEs. The persuasion of this target group should therefore take priority. Addressing the concerns of management identified in our SLR, namely (i) the own company is too small, (ii) there are no potential use cases for the business, and (iii) there is no need to introduce AI as the company can function without it, is essential. To achieve this, we see an opportunity in enhanced communication and collaboration between businesses. Especially, the fact that many managers think that their companies are too small and that they cannot imagine concrete use cases (Husson et al. 2021), can be eliminated through successful experience reports. Direct exchange, for example with the help of networks, would also be useful here. These could be organized by academic institutions, among others. To help SMEs to understand the potential of AI, scholars should provide in addition a practical overview of possible applications and existing case studies, in an easily understandable manner. Furthermore, to address the challenges arising from the complexity and diverse starting conditions of AI implementation in SMEs, many managers from SMEs already have issues defining the first necessary steps for an AI introduction in their company (Lu et al. 2022). Thus, very small-scale procedures are needed that explain an AI introduction step by step and address the different starting conditions.

Both at publications on AI in MNUs and in the media, the topics of trust, ethics, and acceptance are frequently discussed in relation to AI applications. But in the articles on AI in SMEs, only limited reference is made to these points (Joerg and Carlos 2022). There is a need to clarify whether trust, ethics, and acceptance are generally less important in the context of AI solutions in SMEs or whether this is mainly because many SMEs are still in the early stages of introducing AI and therefore do not yet perceive these topics as a challenge. If the subject of ethical and acceptance considerations is not sufficiently pronounced in SMEs, it is important to explore the underlying reasons. This includes investigating whether SMEs have not yet fully recognized the consequences of these aspects in the context of AI and if measures need to be taken. One potential solution is to increase the focus on education and awareness-raising initiatives in this area. Regardless, SMEs should critically scrutinize these aspects for their specific use case before implementation to avoid being surprised by their potential implications later on. This is particularly crucial for AI applications.

6 Conclusion and limitation

The aim of our study was to answer our two research questions and to highlight open research areas, considering the limitations. To obtain a comprehensive overview of important search strings on the topic AI for our SLR, we chose a novel approach. Therefore, we determined the most frequent key words from 77 systematic reviews related to AI (see Table 1) and added the most frequent AI terms from the Artificial Intelligence Index Report 2022 of Zhang et al. (2022). With this, we were able to identify 71 relevant publications for our subsequent SLR. Out of these we figured several findings related to our research questions.

6.1 RQ1. What is the current state of the art of research on AI in the industrial sector regarding SMEs?

  • In order to address the current imbalance of AI research that predominantly focuses on MNEs, it is necessary to increase the number of studies concerning AI regarding SMEs. Furthermore, the scope of research should encompass the full range of business activities, since current studies often consider only a limited number or none at all in certain areas (see Fig. 3)

  • For manufacturing SMEs newly entering the AI domain, we recommend placing initial emphasis on operational activities. This domain exhibits the most advanced research on AI in SMEs, with successful case studies. Moreover, it commonly offers the greatest potential for optimization.

  • In the selected case studies the initial situation of an enterprise, such as the data collection process, is frequently unclear, rendering the transferability of the case study to other companies challenging. SMEs need to be aware of this issue. This can be further complicated by elaborate exposition of technical details, as often the target group of SMEs lacks technical expertise. Therefore, future publications should provide a comprehensive description of the starting situation and carefully consider which technical aspects are necessary and relevant to the topic.

6.2 RQ2. What are the challenges for SMEs in the implementation of AI in the industrial sector?

  • We could identify 27 different challenges for SMEs in total (see Table 2). This shows the great complexity and individuality of AI solutions. Henceforth, it is advisable for enterprises to consider embracing an application-oriented approach while integrating AI. Besides, SMEs can utilize the identified challenges to investigate these for their own AI projects, mitigating the risk of overlooking potential obstacles or risks.

  • The three most frequent challenges in our SLR were knowledge, costs, and the low maturity level in digitalization. Additionally, a pronounced emphasis on financial risks is evident in the economic domain. Many SMEs consider the lack of ROI assessment as an obstacle to deciding on AI implementation, but none of the selected case studies addressed this aspect. This represents a research gap for future studies.

  • For the management, we identified three primary reasons for hesitation: (i) the own business is too small, (ii) lack of AI use cases and (iii) there is no necessity to introduce AI. In response to these concerns, we recommend the conducting of knowledge-sharing rounds with other companies to address these concerns.

However, some limitations must be taken into account when considering the results. First of all, the utilized research method for this article results in certain limitations, as mentioned in other SLRs (Reis et al. 2020). The choice of search strings and exclusion criteria significantly influenced the result of the selected publications. Moreover, many primary studies often considered only a single country, industrial sector, or company (Ha and Jeong 2021; Ushada et al. 2017). This creates a risk of bias in the interpretation of the outcomes. To answer the second research question, the challenges addressed in the individual publications were listed and the number of mentions was counted. For this approach, there is a limit to the extent to which the quantity can be causally linked to the importance of the obstacles for SMEs. Furthermore, there was no different weighting of the challenges based on qualitative criteria.

7 Appendix 1

See Table 4.

Table 4 Summary of reviewed publications—SMEs
Full size table