1 Introduction

The analysis of the learning outcomes of failure experiences has increasingly drawn academic attention (see, e.g., the survey by Dahlin et al., 2018). Building on organizational learning as the dominant theoretical narrative, prior work underlines that failure is a source of knowledge for learning as important as success because failure triggers change by stimulating the search for new strategies rather than reinforcing existing routines (Desai, 2015). The academic debate on learning from failure has primarily focused on how firms and individuals learn from failures and on the mechanisms supporting this learning process (Baum & Dahlin, 2007; Madsen, 2009; Madsen & Desai, 2010).

These studies have significantly expanded the learning literature; however, they mostly analyze how large businesses learn either from failure experiences—regardless of the type of failure—or from failures vis-à-vis successes. In the case of small and medium-sized enterprises (SMEs), prior research has analyzed how they learn from successful experiences. The work by Voss and Voss (2013)—which analyzes the role of learning on product and market domains—and Runst and Thomä (2022)—focused on the innovation outcomes of learning-by-doing processes—are notable cases of this literature. Nevertheless, the specific analysis of how SMEs learn from different types of failure remains, to the best of my knowledge, unexplored in the literature.

To address this gap in the literature, this study investigates the learning patterns of SMEs from different types of operational failure, i.e., work accidents: “non-severe” and “severe and fatal” (hereafter, NS and SF, respectively). This study’s focus on SMEs’ own past accident records and the adoption of tools—i.e., a safety management system—contributes to developing a structured framework for knowledge management and sharing to improve work safety practices (Argote et al., 2021, p. 5405). This study also explores the ambidextrous effect of simultaneously exploiting cumulative experience from NS and SF accidents and the acquisition of tools that modify internal safety protocols.

Organizations do not realize the generally positive learning effects of failure at the same intensity, and failure heterogeneity plays a decisive role in explaining the variations in SMEs’ abilities to learn from failure (Desai, 2016; Madsen, 2009). The distinctive characteristics of failure motivate managers and employees to question existing routines and adopt corrective actions (Haunschild & Rhee, 2004). This study proposes that the lessons drawn from failure will be more beneficial at higher levels of experience and that the diverse consequences of failures influence SMEs’ learning capacity. Also, work accidents are operational failures, and SMEs can learn how to reduce similar events by adopting safety management systems or modifying their memory system (Baum & Dahlin, 2007; Madsen & Desai, 2010; Desai, 2016).

Nevertheless, not all organizations are properly equipped to learn from failure. Successful learning entails important challenges for SMEs because both exploitation and exploration result from different processes and compete for resources which might create tensions between managers or business units (Argote et al., 2021; Dahlin et al., 2018; Raisch & Birkinshaw, 2008). Furthermore, the complexity of ambidextrous learning might amplify in the case of SMEs because they are often subject to liabilities of smallness (e.g., Runst & Thomä, 2022; Teruel-Carrizosa, 2010), which not only hinder knowledge-acquisition processes but also put to the test SMEs’ capacity to match new knowledge with existing tasks (Kim et al., 2016; Voss & Voss, 2013).

To test the study’s hypotheses, the empirical exercise uses a unique dataset of 108 Spanish SMEs during 2006–2009. The multi-industry sample is attractive for a number of reasons. First, the sample includes SMEs that have adopted the analyzed tool in different years (OHSAS 18001 standard: to be defined below in Section 3.2) and showed marked differences in work accident figures in terms of both frequencies (count) and magnitude (NS and SF). Second, in line with the call made by Dahlin et al. (2018, p. 271), the study’s analysis provides the opportunity to assess how SMEs learn from failure in contexts characterized by different routines and where work accidents are heterogeneous. Related, the sample includes SMEs operating in more accident-prone industries (i.e., manufacturing and construction) as well as sectors that traditionally report few work accidents (i.e., professional services) (EU-OSHA, 2017; Lafuente & Abad, 2018). The study of SMEs’ learning patterns in contexts of high and low risks at work enriches the proposed analysis.

Third, the focus of this research on SMEs contributes to unveiling how properties inherent in SMEs’ functioning—in this case, liabilities of smallness and organizational structure and communication processes—might condition their learning potential from different types of work accidents.

Building on the organizational learning frame, learning is defined as SMEs’ temporal performance improvements resulting from organizational processes. In the specific context of this study, the outcome of learning from failure is measured in terms of accident cost reductions, whereas the cumulative experience from NS and SF accidents and the adoption of a safety management system are considered proxies for learning processes. This approach to learning from failure is in line with the process perspective on learning that emphasizes the role of both cumulative experience and different tools for improving internal routines and, subsequently, learning outcomes (see, e.g., Baum & Dahlin, 2007; Desai, 2015, 2016; Haunschild & Rhree, 2004).

Within the learning literature, the increased weight of work safety in the agenda of scholars and policy makers corroborates the relevance of the proposed analysis. Specifically, from an academic perspective, safety management has evolved in the last two decades from a narrow view mostly connected to a costly administrative burden to an operational priority with large economic and social effects (Lafuente & Abad, 2021; Murrell-Blanco, 2023). Also, by analyzing how the accidents’ magnitude—i.e., NS and SF—influences learning, this study helps uncover boundary conditions in the relationship between learning and performance among SMEs.

From a policy viewpoint, the mitigation of work accidents is gaining space among policy makers. Estimations by the European Agency for Safety and Health at Work (EU-OSHA 2017) reveal that the costs of work accidents to workers, businesses, and public administrations roughly represent 3.9% of the global GDP (3.3% of the EU’s GDP). In light of the growing awareness of work safety, European authorities have endorsed different policies within the EU-2020 strategic plan aimed at supporting work safety plans (European Commission, 2007; Lafuente & Daza, 2020).

This is the first study that specifically evaluates SMEs’ learning patterns from diverse operational failures (i.e., work accidents), and the proposed analysis generates valuable insights with relevant implications. First, this study moves away from generic experiential learning and instead analyzes the learning implications for SMEs of different types of work accidents. Thus, this research increases the understanding of learning beyond the focus on cumulative experience and contributes to the further development of the learning-from-failure literature (Baum & Dahlin, 2007; Dahlin et al., 2018; Desai, 2015, 2016; Eggers & Suh, 2019; Kc et al., 2013; Madsen & Desai, 2010).

Second, by examining how a tool designed to improve safety practices supports exploratory learning in the context of SMEs, the study extends previous work on the fit between the acquisition of codified knowledge that alters firms’ organizational memory systems and tasks. Furthermore, by showing that SMEs learn from frequent and rare work accidents through different mechanisms, this research contributes to better comprehending the conditions under which SMEs capitalize on the knowledge generated by different learning processes (Baum & Dahlin, 2007; Kim et al., 2016; Voss & Voss, 2013). Third, the proposed analysis of SMEs’ learning from accidents with different magnitudes (i.e., NS and SF) complements the stream of qualitative research on high-reliability organizations (e.g., space shuttles, nuclear power plants, and air traffic control systems) focused on how large firms learn from mostly catastrophic failures (e.g., Madsen, 2009; Weick et al., 1999).

The structure of the paper is the following. Section 2 presents the theory and hypotheses. Section 3 deals with the methodological aspects of the study, while Section 4 offers the empirical results. Finally, the concluding section (Section 5) presents the discussion, implications, and future research avenues.

2 Background theory and hypotheses

2.1 Learning from failure

Organizational learning represents a critical resource that contributes to explaining long-run organizational success (see, e.g., the recent comprehensive survey by Argote et al. (2021)). Organizational learning theorists propose that businesses create, accumulate, and transfer knowledge that they learn from experience, and that increased knowledge stocks and experience, whether generated within the organization or brought from outside, enhance future performance by improving current practices and developing new routines (Levitt & March, 1988; March, 1991).

A large body of literature rooted in the theory of organizational learning has documented how learning occurs at different levels, including individuals (e.g., Narayanan et al., 2009; Staats & Gino, 2012; Fang He et al., 2018), teams (e.g., Reagans et al., 2005; Wilhelm et al., 2019), and organizations (e.g., Desai, 2015; Myers, 2018).

In parallel to the development of this literature, a stream of research on organizational learning emphasizes that failure experiences are also conducive to knowledge creation and learning (see, e.g., the survey by Dahlin et al. (2018)). Studies in this tradition propose that, compared to other types of experience, failures have distinctive characteristics which motivate organization members to question existing routines and adopt corrective actions (e.g., Baum & Dahlin, 2007; Desai, 2015, 2016; Eggers & Suh, 2019; Haunschild & Rhee, 2004; Kc et al., 2013; Madsen, 2009; Madsen & Desai 2010). Thus, failure not only points to the presence of a knowledge gap in the organization but also points in many cases which provides an indication of where that gap may be. This literature supports the view that learning from failure creates the conditions to alter current knowledge through self-assessment mechanisms or knowledge-acquisition processes.

Prior work has also compared the learning effects of success versus failure and of different types of failures. In the case of the former, the results are inconclusive. While some research reports that organizations learn more effectively from failures than successes (Madsen & Desai, 2010), other studies report that the learning effect of past success is superior to that of past failure (Desai, 2015; Kc et al., 2013). In the latter case, Dahlin et al. (2018) document that failure heterogeneity conditions direct and vicarious experiential learning as well as the learning rates from different types of failures. It is therefore plausible to argue that not all failure experiences are of equal value in supporting learning and that failure heterogeneity produces asymmetries in organizations’ learning rates.

The core of this analysis is to match problems related to work safety with learning-related solutions available to SMEs while acknowledging that differences in work accidents and relevant SME properties condition learning outcomes. The rest of the theory section develops hypotheses relating learning processes to a learning outcome, in this case, the unit cost of work accidents.

2.2 Asymmetric learning from failure: the role of SME properties and accidents’ heterogeneity

Following the organizational learning tradition, learning processes are supported by three dominant approaches (Argote et al., 2021; Levitt & March, 1988; March, 1991). Inspired by the work of March (1991, p. 85), exploitation is conceived as the refinement of existing knowledge via systematic task repetition. Learning begins with experience, and ongoing exploitation of accumulated knowledge and experience features incremental learning, continuous improvement, and, subsequently, superior performance. Exploration focuses on the search for new knowledge that contributes to discovering new alternatives to improve a firm’s operations and routines. Finally, organization learning literature describes ambidextrous learning as the simultaneous exploitation of current knowledge and routines and the exploration of new knowledge (Levinthal & March, 1993).

Building on organizational learning work (Dahlin et al., 2018; Desai, 2015; Haunschild & Rhee, 2004), this study considers cumulative experience from different types of work accidents (NS and SF) and the adoption of a safety management system proxies for learning processes while SMEs’ learning outcome is measured via a widely accepted safety performance measure, namely, the economic cost of work accidents (e.g., Baum & Dahlin, 2007; Lafuente & Abad, 2018). Work accidents are a form of operational failure (e.g., Desai, 2016; Lo et al., 2014), and all organizations are exposed to these events and their consequences. Accident heterogeneity plays a pivotal role in this study. The magnitude of failure matters for learning (Dahlin et al., 2018), and compared to work accident counts, the proposed analysis focused on accident costs allows to capture how SMEs learn from different types of accidents, in terms of their magnitude (minor accidents viz-à-viz severe injuries and fatalities).

For SMEs, the choice between focused learning—i.e., exploitation or exploration—or ambidextrous learning to reduce the cost of work accidents brings about important considerations. There is little empirical work that specifically addresses SMEs’ learning (e.g., Runst & Thomä, 2022; Voss & Voss, 2013). Instead of presenting a line of reasoning that explains why learning is different in SMEs viz-à-viz large firms, this section forwards arguments linking learning processes (i.e., exploitation, exploration, and ambidexterity) to characteristics of SMEs and failure heterogeneity.

2.2.1 Focused learning and SME properties

The execution of any learning process implies coping with various organizational challenges linked to the development of specific structures and processes, as well as the friction between managers for internal resources (Argote et al., 2021). In the specific case of SMEs, focused learning may prove itself a valid process once two inter-connected characteristics of SMEs that may condition their learning rates are accounted for, namely, liability of smallness, organizational structure, and communication processes.

First, in consonance with the concept of liability of smallness (Alrich & Auster, 1986; Fackler et al., 2013), most SMEs face important constraints in accessing financial and human resources, which might compromise their operational and economic potential and increase their vulnerability to internal and external shocks (Runst & Thomä, 2022). Conceptually, this implies that SMEs are limited in their ability to create structures that facilitate the adoption of new tasks or buffers that conceal the tensions that may arise between managers or employees for scarce resources (Karlsson, 2021).

Second, the organizational structure and internal communication processes of SMEs might also affect their ability to learn. SMEs are generally characterized by flat hierarchies, strong interconnections between departments, short communication paths, high alignment between management and employees, and less bureaucratic decision-making (Grözinger et al., 2022). The low degree of structural differentiation and more rapid communication and decision-making provide SMEs with a high degree of organizational flexibility, which facilitates a faster adaptation to changing conditions than their large counterparts and, consequently, increases their performance potential at all levels (Karlsson, 2021). It can therefore be said that the internal structure of SMEs offsets their limited resource basis (Grözinger et al., 2022).

From a learning perspective, the primary advantage of focused learning—i.e., exploration or exploration—for SMEs relies on managerial coherence and operational effectiveness (Desai, 2015). Although both exploitation and exploration entail organizational costs (e.g., obsolescence of knowledge and technology and the cost of excessive experimentation, respectively) (Argote et al., 2021), prior work suggests that focused learning is especially suitable for SMEs with low structural differentiation because this approach is likely accompanied by consensus between management and employees about business’ priorities. Also, focused learning does not necessarily imply the deployment of resources to create or organize units and routines to support learning (Desai, 2015; Voss & Voss, 2013). Therefore, SMEs might have strong incentives for engaging in focused learning as this process does not add extra pressure on the business’ (scarce) resource stock (Desai, 2015; Runst & Thomä, 2022; Voss & Voss, 2013). This argument is indirectly supported by learning studies in large firms. For example, Raisch and Birkinshaw (2008) and Desai (2016) found that structural differentiation enables the pursuit of ambidextrous learning by creating subunits (e.g., departments and regional divisions) that are not necessarily tightly coupled via internal processes or shared resources.

For exploitative learning, SMEs that have accumulated greater experience with past operational failures (i.e., work accidents) have had the opportunity to create and store knowledge that can be used to learn how to mitigate and avoid similar failures in the future by altering internal practices. Effective learning, measured via reductions in accident costs, is therefore attributed to self-assessment processes that might result in an increased knowledge stock that can contribute to reducing inefficiencies through the refinement and adaptation of internal practices (Argote et al., 2021; Dahlin et al., 2018).

Despite experience is an imperfect teacher, it can be argued that SMEs’ accident costs will follow a learning-curve trajectory based on prior information on businesses’ accident-reducing experiential learning (Baum & Dahlin, 2007; Haunschild & Rhee, 2004; Madsen & Desai, 2010). This experiential learning process will likely reduce SMEs’ current accident costs.

In the context of this study, employees’ commitment is critical to disseminating the knowledge and lessons they draw from experience (Wilhelm et al., 2019). SMEs strongly rely on their entrepreneurs’ or managers’ knowledge because they have a direct influence on employees’ involvement compared to their counterparts in large firms (Andries & Czarnitzki, 2014; Runst & Thomä, 2022). Thus, SME managers are agents of change and their drive is critical for transferring the knowledge generated from past accidents to employees and implementing a learning culture within the business.

Besides employees, routines, and the internal memory system, businesses’ knowledge can also be embedded in tools (Kim et al., 2016; Narayanan et al., 2009). Viewed through the exploratory lens, the adoption of tools is often seen as a response to weak internal practices, and incremental learning is expected once the new knowledge is fully integrated into the business’ tasks (Madsen & Desai, 2010).

Such formal knowledge tools—e.g., checklists, computerized systems, virtual work rooms, and repositories of encoded knowledge (e.g., management systems)—can be conducive to exploratory search and positively influence businesses’ learning by facilitating the acquisition, storage, and dissemination of new knowledge, regardless it was generated by their employees or brought from outside (Dahlin et al., 2018, p. 268). Studies on the effect of tools on learning have primarily focused on information technology or knowledge management systems. In their analytical model, Lee and Van den Steen (2010) found that knowledge management systems generate more learning benefits among large firms that deal with similar failures repeatedly; and that recording both successful and unsuccessful experiences supports exploration. Additionally, Kim et al. (2016) documented that the adoption of a knowledge management system in a retail grocery business helps managers with limited access to knowledge sources to search for novel solutions to improve operational performance.

Failure triggers different processes that support learning by unveiling deficiencies in current operations and the business’ current knowledge (Dahlin et al., 2018, p. 262). Failure unquestionably offers valuable knowledge; however, its learning potential will remain untapped without an extra effort to fully comprehend the causes of the event. When managers engage in reflection to improve learning, they might conclude that business practices are flawed and that, because of the characteristics of failures (e.g., complexity and magnitude), designs that produce flattering solutions are no longer valid. Thus, looking for knowledge in new places becomes a viable option to generate potentially novel solutions to improve performance (Kc et al., 2013; Kim et al., 2016).

Concretely, the adoption of knowledge management systems (in this study, OHSAS 18001) entails the introduction of a number of tasks as well as changes in business operations that constitute an information-based opportunity to learn. This process seeks to enhance decision-making by equipping managers with codified knowledge from the repository associated with the management system (Dahlin et al., 2018; Kim et al., 2016). For instance, in the context of this study, managers can use the management system to generate analyses to anticipate new risks at work resulting from technology renewal or operations’ updating and knowledge that alerts them of the need to alter or adjust the business’ safety practices. This way, the schematic configuration of management systems not only facilitates information processing but also helps to develop an analysis of business operations to implement action plans aimed at reducing accidents’ frequency and, subsequently, their personal and economic consequences (Baum & Dahlin, 2007; Lafuente & Abad, 2018; Lo et al., 2014). Also, the analysis of the causes of failure can assist managers in identifying best practices and increasing business’ know-how, and the procedural nature of management systems facilitates knowledge transfer to employees via specific training programs (e.g., drills and protocols) (Kim et al., 2016).

In sum, focused learning incorporates one single process and unified operational routines, thus contributing to mitigating managerial conflicts for limited resources as well as facilitating communication and coordination across business levels (Runst & Thomä, 2022; Voss & Voss, 2013). Taken together, this theory and evidence lead to propose that focused learning—i.e., exploitation or exploration—is positively associated with SMEs’ learning. More formally

  • Hypothesis 1a (H1a). For SMEs, cumulative experience with work accidents will reduce the unit cost of work accidents.

  • Hypothesis 1b (H1b). For SMEs, the adoption of knowledge tools designed to improve safety practices will reduce the unit cost of work accidents.

2.2.2 Failure heterogeneity

Failure challenges businesses’ status quo by pointing to the existence of inadequacies in their internal processes and routines. Although exploitation and exploration are a natural reaction to failure, failure heterogeneity—in this study, measured by the magnitude of work accidents—may have a differentiated impact on SMEs’ capacity to learn from work accidents.

Rooted in learning studies, the “small losses” hypothesis, which proposes that businesses learn more from small failures than from large failures, is the dominant narrative explaining learning as a function of failure heterogeneity (Hayward, 2002; Sitkin, 1992; Weick, 1984). In this tradition, two main objectives motivate the businesses’ response to any failure: to study the causes of the failure to reduce the probability of future failures and to unfold the causes and the responsibility of the failure so that those members responsible for it can be held accountable (Desai, 2016; Dahlin et al., 2018).

When a business experiences a small failure, learning is the primary objective for managers, and the low magnitude of the failure attenuates the value of determining accountability. On the contrary, following failures high in magnitude, both learning and accountability become relevant goals. In this scenario, managers and employees can be reluctant to share information about the failure in order to protect themselves from the consequences of internal scrutiny, and they can opt for a “threat-rigidity response,” that is, they might be unwilling to alter current practices (Desai, 2016).

If the focus is on SMEs, exploitative learning (i.e., learning-by-doing) is typically automatic and tacit (Dahlin et al., 2018), which increases the appealing of this type of learning among generally resource-constrained SMEs that have little incentives to invest in tools that alter their structure and communication system (Andries & Czarnitzki, 2014; Runst & Thomä, 2022). Besides, non-severe work accidents (i.e., small failures) likely share causes that are relatively easier to identify (Desai, 2016). In this context, managers and employees of SMEs may become more adept at extracting information from non-severe (more frequent) accidents as their businesses gain experience (Lafuente & Abad, 2018). With greater operating experience, SMEs’ organizational members become more familiar with their business’ environment and are better able to make sense of useful information from past minor accidents and implement preexisting solutions to deal with new accidents.

As a result of this exploitative process, the impact of non-severe work accidents, in this case, measured by the unit accident cost, is expected to decrease as SMEs accumulate experience with this type of accidents that allows them to increase their knowledge stock and adopt corrective actions, and this can be interpreted as a learning outcome (Baum & Dahlin, 2007; Desai, 2016). This learning pattern, which reinforces SMEs’ learning from their own experience with past non-severe work accidents, emphasizes that exploitation will result in successful learning if, after prior work accidents drive modifications in existing knowledge and routines, SMEs alter their practices to better evaluate and anticipate potential work risks, thus reducing the negative consequences of future work accidents proxied by the cost of these events (Baum & Dahlin, 2007; Desai, 2016; Lafuente & Abad, 2018).

  • Hypothesis 2 (H2). For SMEs, cumulative experience with non-severe work accidents produces greater learning outcomes, in terms of reductions in the unit cost of work accidents, than does cumulative experience with severe and fatal work accidents.

2.2.3 Ambidextrous learning

Work on organizational learning underlines that for the benefits of learning to materialize, firms should maintain a balance between exploration and exploitation: “The basic problem confronting an organization is to engage in sufficient exploitation to ensure its current viability and, at the same time, to devote enough energy to exploration to ensure its future viability,” but, “the precise mix of exploitation and exploration that is optimal is hard to specify” (Levinthal & March, 1993, p. 105). Decision on the balance between exploration and exploitation is especially aggravated in resource-constrained SMEs (Runst & Thomä, 2022). Arguments on how SMEs’ ability to learn through ambidexterity might be conditioned by failure heterogeneity are presented below.

Failure heterogeneity implies differences and complexity in the origin of failures. As with any failure, accidents are caused by a combination of human- and machine-led errors (e.g., incorrectly executed tasks, operational violations, and chance), and businesses’ ability to learn relies on their capacity to identify and understand the failure leads so that corrective actions to prevent future accidents and minimize their impact can be implemented (Desai, 2016). The business’ information processing is critical to generate the deliberate reflection that precedes a learning process (Dahlin et al., 2018). As mentioned above, exploitation of experience from non-severe accidents is especially attractive for SMEs because this approach does not entail restructuring investments that might deteriorate their position or add complexity to their operations (Andries & Czarnitzki, 2014; Runst & Thomä, 2022).

A different picture emerges when severe and fatal accidents are at the center of the analysis. These accidents are rare, so; therefore, businesses have to invest purposely in order to develop a comprehensive knowledge stock that can be used to learn how to prevent these failures (Desai, 2016; Madsen, 2009). Despite the intuitive appeal of exploitation processes, learning from past severe and fatal accidents can produce counterproductive outcomes. When faced with work accidents, managers tend to adopt preexisting solutions rather than develop new ones (Eggers & Suh, 2019). Thus, firms with experience in minor accidents are tempted to handle severe and fatal accidents by reusing the same actions learned from past experience in minor accidents, hoping to obtain similar outcomes (Dahlin et al., 2018; Desai, 2015). Yet, because severe and fatal accidents are rare and require unique solutions, extrapolating corrective actions from minor accidents to injuries and fatalities without a clear grasp of the problem at hand can be problematic. Blindly implementing solutions designed to tackle minor accidents might result in a mismatch between knowledge generated by experience with minor accidents and the underlying causes of the severe or fatal accident, thus leaving the genuine threat of severe accidents unresolved (Eggers & Suh, 2019; Madsen & Desai, 2010).

Accidents with more heterogeneous causes create more learning (Desai, 2015; Haunschild & Sullivan, 2002). In the presence of little prior experience with severe and fatal accidents, the use of acquired tools may constitute a valid option to better channel the knowledge generated by these rare and complex accidents and, consequently, minimize their occurrence in the future.

The specific properties of severe and fatal accidents—i.e., low frequency and complex causes—defy corrective actions based on preexisting solutions and force managers to creatively search for new solutions (Madsen, 2009). Yet, in their attempt to understand the large failure’s causes, managers do not always develop beliefs that correspond with the true causes of such events (Dahlin et al., 2018). The effective handling and analysis of the large volume of information generated by a large failure is an essential prerequisite to fully extracting the “lessons learned” from these events (Desai, 2015).

In this sense, knowledge management tools (including the OHSAS 18001) improve businesses’ internal information system, and their structured nature offers the opportunity to clearly (and objectively) identify all individuals, processes, and technologies involved in all work accidents. This is especially relevant in the case of severe or fatal accidents as such tools ease the access to more organized information to enhance managers’ decision-making which, in turn, supports a more meaningful learning that relies on richly and thoroughly investigating these failures and their root causes (Kim et al., 2016; Madsen, 2009). This way, management systems equip managers with valuable input on large failures’ causes and help them to adequately account for the complex interactions that take place among individuals, technologies, and processes (Dahlin et al., 2018). The primary objective of this strategy is to identify gaps in safety practices and create structured safety-enhancing roadmaps that lead to the development of new solutions to mitigate the consequences of severe and fatal accidents.

These arguments are reinforced by prior studies. Madsen (2009) found that coal miners learn from both minor accidents and disasters but through different mechanisms: frequently occurring minor accidents support learning by warning employees on the importance of adhering to the safety routines set by the business, whereas rare disasters activate a process of drastic change in the business’ safety practices. Similarly, Stan and Vermeulen (2013) show that fertility clinics that admit complex cases report more failures and learn faster than clinics that admit simple (less complex) cases. Failure complexity enables clinics to enhance their understanding of different knowledge areas, explore new solutions, and develop tools and routines to store and transfer knowledge. Related, Clay-Williams and Colligan (2015) found that the adoption of new tools, such as highly detailed checklists, has contributed to improving work safety conditions in the aviation and health care industries. Additionally, Lafuente and Abad (2018) report that, regardless of firm size, the properties of business operations (more or less systematic) condition the joint effect of work accidents’ recent record and the OHSAS 18001 on the rate of work accidents. In the specific context of SMEs, Voss and Voss (2013) underline that contingent upon business age, management systems, and processes can be instrumental in managing the complexities of ambidexterity and enhancing learning.

Following these arguments and evidence, for SMEs’ task performance, experience will produce learning if this process is based on past experience with non-severe work accidents, and the adoption of tools offers the opportunity to modify businesses’ organizational structure and information system to obtain richer information about large failures, which, in turn, leads to design more efficient remedies to complex problems, such as severe and fatal accidents (Desai, 2016; Kc et al., 2013). Therefore, the value of past experience with severe and fatal work accidents might be better capitalized through an ambidextrous approach that matches the knowledge generated from these rare accidents and the drastic organizational changes that follow the adoption of tools that modify internal routines and, therefore, help minimize the consequences of severe and fatal accidents. More formally,

  • Hypothesis 3 (H3). For SMEs adopting tools to improve work safety practices, cumulative experience with severe and fatal work accidents produces greater learning outcomes—in terms of reductions in the unit accident cost—than does cumulative experience with minor work accidents.

3 Research program

3.1 Sample

The information used in this research was collected specifically for the purpose of this study, and data came from three sources. First, it was possible to identify SMEs that obtained the OHSAS 18001 as well as the certification year through cooperation established with the Spanish Association for Standardization and Certification (AENOR). This agency is the leader in the certification of work safety management systems based on the OHSAS 18001 in Spain. Second, in collaboration with one of the largest Spanish occupational injuries insurance companies—FREMAP—it was possible to obtain information about variables related to occupational health and safety performance. In consonance with Spain’s Occupational Risks Prevention Act of 1995, FREMAP is one of the many occupational injuries insurance companies accredited by the Spanish government.

The two databases obtained from AENOR and FREMAP were merged on the basis of the business’ unique identification code. As a result of this matching procedure, detailed information about the OHSAS 18001 certification date as well as data about key work safety measures was obtained for 152 SMEs during 2006–2009. Concretely, the dataset includes specific information about the number and the economic cost of work accidents reported by the sampled firms. At this point, notice that 20 businesses were removed from the sample because data on total work accident costs was not available or incomplete. In addition, the data allows to differentiate work accidents by the degree of injury making a distinction between non-severe (minor) accidents and injuries and fatal accidents. Following the Spanish legislation and EU standards (Lafuente & Abad, 2021), the data on work accidents includes accidents at work as well as accidents on the way to and from work; however, the obtained information does not allow to distinguish between these two types of accidents.

Notice that reliable and complete data on the breakdown of work accidents by severity was not available for 24 firms; therefore, these businesses were excluded from the sample in the interest of following a rigorous methodology. As a result, data availability limits the final sample to 108 SMEs.

Third, based on the unique identification code, accounting and organizational data were obtained for the period 2005–2009 from the Spanish database SABI (Sistema de Análisis de Balances Ibéricos). This database is provided by the Bureau Van Dijk© and contains detailed balance and income statement information, as well as qualitative data for Spanish firms. An examination of the size distribution of the sample reveals that 31.54% of businesses are small (less than 50 employees), while 68.46% of businesses fall into the medium-sized firms category (between 51 and 250 employees). Also, construction organizations represent 46.59% of the sample, while the proportion of SMEs in the manufacturing and professional services sectors stands at 27.30% and 26.11%, respectively.

3.2 Variable definition

3.2.1 Dependent variable

Safety performance is measured through the unit accident cost which is calculated, for each firm (i), as the total cost of work accidents (indexed to thousands of 2009 euro) divided by the number of accidents. Safe working conditions are a priority for any firm, and this variable directly captures safety outcomes by measuring the consequences of poor safety at work (Baum & Dahlin, 2007; Cabrera-Flores, 2023; Delgado-Sánche, 2023; Lafuente & Abad, 2018). The proposed approach to unit accident cost enhances estimation accuracy by accounting for the possibility of multiple accidents by the same employee. This variable was logged to reduce skewness. Figure 1 illustrates the unit accident cost by the total number of work accidents. The functional form of the empirical data suggests that SMEs that have accumulated greater accident experience have learned how to minimize the consequences of similar work accidents, thus reducing their unit accident cost. The unit accident cost follows the expected learning-curve pattern reported by prior work on organizations’ learning from failure (Baum & Dahlin, 2007; Desai, 2016; Madsen, 2009; Madsen & Desai, 2010).

Fig. 1
figure 1

Source: authors’ elaboration based on the study data

The relationship between the unit cost of accidents and the number of work accidents.

Full size image

3.2.2 Accident experience

Following the learning-curve tradition, learning refers to organizational processes that enhance subsequent performance, and cumulative experience is considered a proxy for exploitative learning (Baum & Dahlin, 2007; Argote et al., 2021). Concretely, cumulative experience is examined via two variables. First, total accident experience \(({AE}_{it})\) is measured as the accumulated number of work accidents reported by the focal organization (i) in a given period (t).

The second variable of accident experience deals with the severity of work accidents. Accident severity—as a proxy measure of failure heterogeneity—generates valuable information which varies with the consequences of accidents, thus affecting learning. The visibility and severity of work accidents might motivate their reduction through the intensive refinement of existing routines and safety practices. To evaluate the learning effects of different types of accidents, the accumulated accident experience for each business and each period was split into two variables: accumulated experience with non-severe accidents \(({AE}_{it}^{NS})\) which is calculated as the cumulative number of minor accidents and the accumulated experience with severe and fatal accidents \(({AE}_{it}^{SF})\) which is computed as the cumulative number of injuries and fatal accidents.

Among the sampled SMEs, the mean work accident rate (accidents per worker) is 7.37%, and accidents are mostly non-severe: SMEs reported 7.15 accidents per year (range 0–76), while severe injuries and fatalities represent 1.70% of the total number of work accidents. Further scrutiny of the data reveals that only eight SMEs did not report a work accident during the study period, and for these firms, \({AE}_{it}^{NS}=0\) and \({AE}_{it}^{SF}=0\). Also, the number of “zero-accident” cases is relatively low, ranging between 18 (16.67%) in 2007 and 21 (19.44%) in 2009. At the industry level, construction businesses show the highest rate of work accidents (11.03%), whereas the work accident rate in the manufacturing and professional services sectors was 8.50% and 5.50%, respectively. It should be noted that for the study period manufacturing and construction are more accident-prone industries than professional services activities: proportion of “zero-accident” manufacturers: 12.50%, proportion of “zero-accident” construction firms: 13.18%, and proportion of “zero-accident” cases in professional services: 25.80%.

Similar to prior work (Baum & Dahlin, 2007; Desai, 2016; Madsen & Desai, 2010), the variables related to accumulated accident experience (\({AE}_{it}\), \({AE}_{it}^{NS}\), and \({AE}_{it}^{SF}\)) are depreciated by a discount rate \((\lambda )\) which depreciates past failure experience as a function of business age. This way, the variables account for the possibility that the potential benefits of past accidents experience decay over time due to organizational forgetting or knowledge antiquation (Argote et al., 2021). The value of accident experience for these variables is the sum of prior accidents reported for the focal SME, each divided by the depreciation factor \((\lambda )\), that is, \(\sum_{t=1}^{t-1}\frac{{AE}_{it}}{\lambda }\) for total accumulated accident experience (\({AE}_{it}\)), while the experience variables take the form \(\sum_{t=1}^{t-1}\frac{{AE}_{it}^{NS}}{\lambda }\) and \(\sum_{t=1}^{t-1}\frac{{AE}_{it}^{SF}}{\lambda }\) in the case of the accumulated experience with non-severe (\({AE}_{it}^{NS}\)) and severe (\({AE}_{it}^{SF}\)) accidents, respectively.

At this point, two considerations are in order. On the one hand, due to data availability on work accident figures, the first year used to compute the variables \({AE}_{it}\), \({AE}_{it}^{NS}\), and \({AE}_{it}^{SF}\) is 2005. Some caution is advised due to the potential bias caused by the lack of data over a longer time period. It should also be noted that the three variables are computationally reliable, and tracking work accidents from 2005 only implies that this research evaluates how organizational memory accumulated during the study period influences SMEs’ learning. On the other hand, keep in mind that four values of \(\lambda\) were tested to determine the discount rate and, consequently, the learning variables related to accumulated experience with accidents: \(\lambda =1\), which indicates perfect organizational memory (knowledge does not depreciate); \(\lambda =\sqrt{{\text{experience}}}\), a less-than linear depreciation rate; \(\lambda ={\text{experience}}\), which assumes that knowledge depreciates linearly; and \(\lambda ={{\text{experience}}}^{2}\), a faster-than-linear depreciation rate (Baum & Dahlin, 2007; Desai, 2016). Parameter estimates for all accident experience variables using \(\lambda =\sqrt{{\text{experience}}}\) yield higher coefficient t-statistics and goodness of fit measures (chi-square statistic and R2 values). This suggests that, regardless of the severity of work accidents, the value of cumulative accident experience for reducing unit accident costs depreciates slowly. All accident experience variables were logged to reduce skewness, and the analysis presented in Section 4 uses estimates based on this specification.

3.2.3 Safety management system (codified safety knowledge)

Firms adopting safety systems introduce a variety of controls seeking to minimize work risks (Lafuente & Abad, 2018; Lo et al., 2014). The objective of the Occupational Health and Safety Assessment Series (OHSAS) 18001 is to support and promote good practices in the area of occupational health and safety via systematic and structured safety practices (BSI, 2007). The safety system is granted to firms that develop and maintain a safe workplace, minimizing risks to their employees. The OHSAS 18001 was the reference safety system, and the number of certified firms has significantly grown worldwide (Lafuente & Abad, 2018). Notice that the new ISO 45001 standard was launched in 2018 to formally replace the OHSAS 18001.Footnote 1 The OHSAS 18001 can be seen as a soft self-regulatory norm related to the introduction of new safety practices. The full implementation of the OHSAS 18001 entails the development of different tasks which act as knowledge repositories of the new safety knowledge and are executed by employees of all organizational levels. In conclusion, the OHSAS 18001 can be seen as a strategic tool that facilitates the creation of safety knowledge through safety controls that not only help to reduce work accidents but also ultimately enhance firm performance (Lafuente & Abad, 2018).

In the empirical analysis, the introduction of new safety knowledge through the OHSAS 18001 was identified with a dummy variable that takes the value of one for SMEs adopting the OHSAS 18001, and zero otherwise. To enhance estimation accuracy, the study considers that the adoption of the OHSAS standard corresponds to a given period (t) if certification took place between the second half of year t − 1 (from July to December) and the first semester of period t (from January to June). This empirical design allows to examine the extent to which the new safety knowledge impacts subsequent unit accident costs. The sampled SMEs adopted the OHSAS 18001 at a different pace. For certified SMEs, three firms adopted the OHSAS in 2004, four in 2005, 15 in 2006, 20 in 2007, and 34 in 2008. Also, six SMEs adopted the safety standard in the second semester of 2009 which implies that, based on the approach adopted in this study, safety investments in these SMEs will likely not impact future accident costs.Footnote 2 Thus, the final sample includes a control group of 32 non-certified SMEs.

3.2.4 Control variables

The model specifications control for size, age, performance, industry, and time. Business size is measured by the number of employees, while business age is expressed in years. Business performance might affect safety outcomes by shaping managers’ priorities and attention to work safety conditions (Desai, 2015). Therefore, all models control for labor productivity measured as the ratio of sales divided by employees (indexed to thousands of 2009 euro). In all specifications, the variables size, age, and performance were logged to reduce skewness and enhance estimation accuracy. Also, industry-specific features might promote safety learning at different intensities. Thus, a set of industry dummies is included in all models (i.e., manufacturing and professional business services, while construction is the omitted industry variable). Finally, a set of time dummy variables rule out the potential effect of time trends and other environmental changes on the cost of work accidents.

3.3 Method

Panel data techniques are used to estimate the proposed learning-curve model which emphasizes exploitative learning from own accident experience and exploratory learning from knowledge acquisition processes linked to the adoption of the OHSAS 18001. Pooling repeated observations on the same SMEs violates the assumption of independence of observations, resulting in autocorrelation in the residuals. First-order autocorrelation occurs when the disturbances in one time period are correlated with those in the previous time period, resulting in incorrect variance estimates, rendering ordinary least squares (OLS) estimates inefficient and biased (Wooldridge, 2002). Therefore, random-effects panel data generalized least squares (GLS) models with robust standard errors are used to correct for autocorrelation of disturbances due to constant firm-specific effects (Greene, 2003).

To evaluate the role of experiential and exploratory learning empirically, the following baseline learning curve models businesses’ unit accident cost as a function of both accumulated accident experience and the adoption of the OHSAS 18001. More formally,

$${\text{ln}}{\mathrm{unit accident cost}}_{it}={\beta }_{0}+{\beta }_{1}{\text{ln}}\sum_{t=1}^{t-1}\frac{{AE}_{it}}{\lambda }+{\beta }_{2}{\mathrm{safety system}}_{it}+{\beta }_{12}{\text{ln}}\sum_{t=1}^{t-1}\frac{{AE}_{it}}{\lambda }\times {\mathrm{safety system}}_{it}+{\beta }_{3}{{\text{controls}}}_{it}+{T}_{t}+{\eta }_{i}+{\varepsilon }_{it}$$
(1)

In Eq. (1), \({\beta }_{j}\) are parameter estimates for the jth independent variable, \({\eta }_{i}\) is the time-invariant business-specific effect that controls for unobserved heterogeneity across SMEs (i) and that is uncorrelated with parameter estimates, and \({\varepsilon }_{it}\) is the normally distributed error term that varies cross-observations and cross-time (t). Control variables include business size, business age, performance, and industry, while T refers to the set of (T − 1) time dummy variables. All time-varying variables are lagged one period to avoid potential endogeneity problems due to reverse causality.

In terms of the study hypotheses, it is expected that \({\beta }_{1}<0\) (exploitation: H1a) and \({\beta }_{2}<0\) (exploration: H1b) to corroborate the benefits of focused learning for SMEs.

Equation (1) is extended to include the interactive effects of exploitation and exploration on the unit accident cost:

$${\text{ln}}{\mathrm{unit accident cost}}_{it}={\beta }_{0}+{\beta }_{1}{\text{ln}}\sum_{t=1}^{t-1}\frac{{AE}_{it}^{NS}}{\lambda }+{\beta }_{2}{\text{ln}}\sum_{t=1}^{t-1}\frac{{AE}_{it}^{SF}}{\lambda }+{\beta }_{3}{\mathrm{safety system}}_{it}+{\beta }_{13}{\text{ln}}\sum_{t=1}^{t-1}\frac{{AE}_{it}^{NS}}{\lambda }\times {\mathrm{safety system}}_{it}+{\beta }_{23}{\text{ln}}\sum_{t=1}^{t-1}\frac{{AE}_{it}^{SF}}{\lambda }\times {\mathrm{safety system}}_{it}+{\beta }_{4}{{\text{controls}}}_{it}+{T}_{t}+{\eta }_{i}+{\varepsilon }_{it}$$
(2)

Using this notation (Eq. (2)), it is expected that \({\beta }_{1}<0\) and \({\beta }_{1}<{\beta }_{2}\) to confirm that cumulative experience with NS accidents produces greater learning outcomes than does cumulative experience with SF accidents (H2), whereas a result of \({\beta }_{23}<0\) and \({\beta }_{23}<{\beta }_{13}\) will confirm the ambidextrous learning hypothesis (H3), that is, for SMEs adopting tools to improve work safety practices, cumulative experience with severe and fatal work accidents produces greater learning outcomes—in terms of reductions in the unit accident cost—than does cumulative experience with minor work accidents. The Hausman (1978) specification test was estimated to corroborate the appropriateness of the proposed random-effects models. Results, presented in Table 2 (Section 4), indicate for all models that the business-specific effects are uncorrelated with the regressors, thus confirming that the business-specific heterogeneity can be modeled as randomly distributed across cross-sectional observations (Greene, 2003, p. 285).

Table 1 presents summary statistics and correlations for the study variables. Each of the accumulated accident experience variables reported in the table was depreciated using the depreciation parameter \((\lambda )\), as described in Section 3.2. The correlations among the analyzed variables are generally in the low to moderate range, although the correlations among the accident experience variables are high. One reason for these fairly high correlations could be that, regardless of the severity of the accidents, accident experience increases with the business’ overall (market) experience.

Table 1 Descriptive statistics and bivariate correlations
Full size table

4 Results

This section presents the findings for the effect of exploitative learning from own accident experience and exploratory learning resulting from the adoption of the OHSAS 18001 on unit accident cost. Following Eq. (1), model 1 is the baseline specification which includes the control variables, the cumulative accident experience, and exploratory learning, whereas model 2 introduces the interaction term between cumulative accident experience and the exploratory learning variable. Based on Eq. (2), model 3 splits cumulative experience by accidents’ severity (“minor” (NS) and “severe and fatal” (SF)), while model 4 presents the full model with the interactions between the two types of cumulative experience and exploratory learning.

Prior to reporting the findings, notice that the potential presence of collinearity was tested in all model estimations. A low to moderate level of multicollinearity among explanatory variables can generate less-efficient parameter estimates (i.e., larger standard errors) for the correlated variables, but will not likely bias coefficients (Greene, 2003). To evaluate the potential presence of collinearity, the average variance inflation factor (VIF) was computed for all variables, and summary results are presented in Table 2. For the baseline models drawn from Eq. (1), the average VIF values are 1.74 (model 1 in Table 2) and 2.39 (model 2 in Table 2), and individual VIFs in these models ranged between 1.29 and 4.97.

Table 2 Random effects GLS models of the unit cost per work accident
Full size table

The average VIFs for models 3 and 4 in Table 2 (Eq. (2)) are 1.68 (ranging between 1.14 and 2.51) and 3.08 (ranging between 1.30 and 6.76), respectively. In model 4 the highest VIF values—which are below the generally recommended ceiling value of 10 (Cohen et al., 2003)—were observed for the variable accumulated experience in NS accidents and the interaction term between exploratory learning (OHSAS 18001) and the accumulated experience in NS accidents. The results of this diagnostic test do not raise collinearity concerns. Keep in mind that, even if theoretically correct, models including interaction terms typically show moderate to high collinearity levels (Greene, 2003).

Because the interaction effects between exploitation and exploration represent key theoretical predictions in this study, both the individual and the multiplicative terms are used in the regression models. Furthermore, instability in regression coefficients was not detected for the baseline model and control variables when the key learning variables and their interactions were individually added.

4.1 Baseline results: exploitation and exploration

Concerning the baseline specification (model 1in Table 2) analyzing the learning effects of cumulative experience with work accidents and the acquisition of safety knowledge, the coefficient for own work accident experience is negative and statistically significant (\({\beta }_{1}=-0.4174\) and p-value < 5%). This finding means that unit accident cost significantly declines with increases in own cumulative experience with work accidents. This result supports hypothesis 1a (H1a) that states that, for SMEs, cumulative experience with work accidents reduces the unit cost of work accidents.

On the contrary, the coefficient for the variable linked to the acquisition of safety knowledge (OHSAS 18001) is not significant in the models (\({\beta }_{2}=0.4133\) and not significant). Contrary to the study’s theoretical development in Section 2, hypothesis 1b (H1b)—which proposes that, for SMEs, the acquisition of knowledge tools designed to improve safety practices reduces the unit cost of work accidents—is not supported.

To complement the analysis, Fig. 2 plots the slope coefficients reported in model 2 (Table 2) for the cumulative experience with work accidents in SMEs that adopted the OHSAS 18001 viz-à-viz SMEs that did not introduce the safety standard. Control variables are set to their sample means.

Fig. 2
figure 2

Cumulative experience with work accidents (total) and exploration interaction

Full size image

These results corroborate that the sampled SMEs learn how to reduce accident costs from their cumulative experience; however, they do not do so from an exploratory approach. In other words, the strategic emphasis focused on the exploitation of cumulative experience with work accidents produces greater learning benefits, in terms of reductions in the unit cost of work accidents, than an exploratory strategy based on the acquisition of safety knowledge (i.e., the OHSAS 18001 standard).

4.2 Learning patterns from different types of work accidents

This section evaluates how accident heterogeneity—in terms of severity—conditions the relationship between learning models and the unit cost of work accidents. Results are presented in models 3 and 4 of Table 2. To aid in the interpretation of the estimations, Figs. 3 and 4 plot the effects of ambidextrous learning based on estimates from model 4. Control variables are set to their mean values. The figures illustrate how the adoption of the OHSAS 18001 impacts the relationship between cumulative experience with heterogeneous work accidents (NS and SF) and the unit accident cost. Notice that model 4 of Table 2 is used to interpret the coefficients and test hypotheses 2 and 3.

Fig. 3
figure 3

Cumulative experience with non-severe (NS) work accidents and exploration interaction

Full size image
Fig. 4
figure 4

Cumulative experience with severe and fatal (SF) work accidents and exploration interaction

Full size image

Overall, the findings underline the relevance of distinguishing between minor (NS) and severe and fatal (SF) work accidents when analyzing the role of ambidextrous learning on SMEs’ outcomes.

Hypothesis 2 (H2) proposes that, among SMEs, cumulative experience with non-severe work accidents produces greater learning outcomes, in terms of reductions in the unit cost of work accidents, than does cumulative experience with severe and fatal work accidents. The findings in model 4 support this hypothesis. Specifically, by unfolding work accidents based on their severity, the results indicate that exploitation is conducive to greater learning outcomes—i.e., lower unit accident cost—only when such cumulative experience is related to minor accidents (\({\beta }_{1}=-0.4790\) and p-value < 5%), while this pattern does not occur for cumulative experience with sever and fatal accidents (\({\beta }_{2}=1.1802\) and not significant).

Following the theoretical arguments presented in Section 2, these results confirm the “small losses” hypothesis. Minor work accidents are a valuable source of information and knowledge because their comparatively low impact on worker’s health, relative to severe and fatal accidents, motivates managers to focus on the cause of the accident; this process facilitates information gathering and, consequently, enhances learning (Desai, 2016; Hayward, 2002). Additionally, this result is in line with existing studies on learning to show that the relatively high incidence rate of non-severe accidents, compared to that of severe and fatal accidents, allows one to gain more experience that increases the business’ knowledge stock and supports exploitative learning (Desai, 2016; Madsen, 2009).

Turning to the exploitation-exploration interaction estimates for model 4 points to a contrasting effect of adopting the OHSAS 18001 on the relationship between cumulative experience with work accidents and unit accident cost. Concretely, the coefficient for the joint effect of cumulative experience with minor accidents and the OHSAS 18001 dummy is not significant (\({\beta }_{13}=0.1087\) and not significant), whereas the estimate for the interaction between cumulative experience with severe and fatal accidents and the exploration variable linked to the OHSAS 18001 is negative and significant (\({\beta }_{23}=-2.0118\) and p-value < 5%).

To help interpret these results, Figs. 3 and 4 graphically represent the contrasting effect of cumulative experience with different types of work accidents in the relationship between the acquisition of safety knowledge (i.e., OHSAS 18001) and the unit cost of work accidents. The figures illustrate that, among the sampled SMEs, effective learning occurs from the exploitation of the experience with past minor accidents, regardless of whether the business searched for new safety knowledge (Fig. 3). On the contrary, Fig. 4 clearly shows that successful exploitation of the knowledge generated by past severe and fatal accidents requires complementary exploratory efforts that, in the case of this study, focus on the adoption of the OHSAS 18001.

Collectively, this pattern of interaction effects is consistent with hypothesis 3 (H3) that predicts that, among SMEs adopting tools to improve work safety practices, cumulative experience with severe and fatal work accidents produces greater learning outcomes—in terms of reductions in the unit accident cost—than does cumulative experience with minor work accidents.

Severe and fatal accidents are rare events that require a unique approach likely unconnected to preexisting recipes (Argote et al., 2021). These results are thus in line with theoretical predictions that exploration might represent an effective process to learn from severe and fatal accidents whose resolution relies more on the generation of knowledge that equips businesses with new solutions to these accidents (Desai, 2016; Haunschild & Sullivan, 2002; Madsen, 2009) and relies less on the mere replication of practices resulting from past experience with minor accidents (Eggers & Suh, 2019).

4.3 Robustness check: the role of learning on the dispersion of the unit cost of work accidents

So far, the empirical exercise has focused on how accumulated knowledge and safety knowledge impact the mean unit cost of work accidents. This section extends the proposed model by evaluating the effects of learning on the dispersion of the unit cost of work accidents among the sampled SMEs. Specifically, Eqs. (1) and (2) are re-estimated by performing a series of cross-sectional regressions—estimated via the OLS method—to verify if the accumulated experience with heterogeneous work accidents (NS and SF) and the acquisition of safety knowledge (i.e., OHSAS 18001) reduce the volatility of the unit cost of work accidents.

Data for the year 2009 was used to run the proposed OLS regressions. The dependent variable used in this section is the volatility of the unit accident cost measured by the standard deviation of the unit cost of work accidents reported by each SME over the study period (mean value: 3293.10 euro; range: 1001–105640 euro). This variable was logged to reduce skewness. Concerning the learning variables, the total values reported for the three metrics of accumulated experience with work accidents (total, NS, and SF) are used in the analysis. In line with the empirical design of this study, these variables lagged one period (i.e., accumulated data up to 2008) to avoid potential endogeneity problems. The OLS regressions also include the (lagged) control variables: business size, business age, labor productivity, and industry.

The cross-sectional regression results reported in Table 3 are in favor of the baseline model (Eq. (1)). The negative coefficients in models 1 and 2 confirm that cumulative experience with work accidents not only reduces unit accident costs but also contributes to narrow down the dispersion of accident costs among the sampled SMEs (model 2: \({\beta }_{1}=-0.3212\) and p-value < 5%).

Table 3 OLS results: dispersion in the unit cost per work accident (2006–2009)
Full size table

Once accident heterogeneity is accounted for (Eq. (2)), the results in models 3 and 4 show that the positive learning impact of accumulated experience on the dispersion of SMEs’ unit accident cost primarily comes from the effect of non-severe (NS) work accidents. This finding is consistent with the main results presented in models 3 and 4 of Table 2.

A different pattern is observed in the case of severe and fatal accidents (SF). From model 4, it can be seen that cumulative experience with SF accidents increases the dispersion of the unit accident cost (\({\beta }_{2}=1.3027\) and p-value < 5%). Because SF accidents are rare events with major personal and economic repercussions, it is plausible to argue that this result only reflects the economic gap that exists in SMEs reporting NS and SF accidents; the dispersion (i.e., standard deviation) of SMEs’ unit accident cost increases as businesses deal with a larger number of SF work accidents. A simple Pearson correlation test between the dispersion of the unit accident cost and the cumulative experience variables corroborates this intuition for both NS accidents (Pearson’s correlation: −0.4773, p-value < 0.000) and SF accidents (Pearson’s correlation: 0.1829, p-value = 0.042).

Finally, the consistently non-significant coefficient for the variable linked to the acquisition of safety knowledge suggests that the OHSAS 18001 is not a “gearing tool” between the dispersion of unit accident costs and exploitative learning, regardless of whether such experiential learning is connected to NS or SF work accidents.

5 Discussion, implications, and lines of future research

5.1 Discussion of results

Work safety cannot be overlooked. This study investigated the learning patterns of SMEs from their own experience with different types of operational failures, namely, minor (NS) and severe and fatal (SF) work accidents. This article also attempted to offer a fine-grained assessment of how failure heterogeneity, in terms of the magnitude of work accidents, conditions the effect of exploitation, exploration, and ambidexterity on learning outcomes, measured as the unit cost of work accidents.

Taken together, the findings reveal that the sampled SMEs learn from work accidents through different mechanisms, and the reported asymmetric learning process is driven by accident heterogeneity, in terms of frequency and severity: the positive effect of exploitation on the level and volatility of the unit cost of work accidents is significant only for direct experience with non-severe accidents (NS), whereas SMEs’ accident costs decline with the adoption of an ambidextrous strategy that involves the exploitation of past experience with severe and fatal accidents and the acquisition of knowledge management systems (i.e., OHSAS 18001) that facilitate the storage, analysis, and transfer of safety knowledge generated by these type of accidents (SF).

This research is relevant in various dimensions. First, the results of this study are generally consistent with the postulates of the organizational learning theory; however, they also pose some challenges to them. Specifically, the identification of an asymmetric learning pattern among SMEs constitutes a contribution to the literature on learning from failure (e.g., Dahlin et al., 2018; Eggers & Suh, 2019; Madsen, 2009). While the estimates for minor work accidents are consistent with prior work dealing with the “small losses” hypothesis, which confirms that the concentration of failures influences SMEs’ ability to learn from these events (e.g., Baum & Dahlin, 2007; Desai, 2015; Hayward, 2002; Kc et al., 2013), the findings for the coefficients linked to severe and fatal accidents opened the door for a more precise interpretation of how SMEs learn from severe and fatal accidents.

Although the data does not allow to explore the mechanism underlying SMEs’ learning process, the results are in line with the notion that learning from severe and fatal accidents is a complex process that entails the gathering, encoding, and analysis of new information generated by these rare events that will likely induce managers to modify or update internal routines and safety practices (Madsen, 2009). The null result for the coefficient linked to severe and fatal accidents should not be interpreted as evidence that SMEs do not learn from these events. Instead, by putting to the test the ambidexterity assumption, it was found that SMEs do not realize the generally positive effects of learning at the same intensity, and that under certain conditions, they can simultaneously benefit from experiential learning and knowledge tools that support exploration by modifying SMEs’ memory system.

Achieving ambidextrous learning is far from being a trivial managerial exercise. The findings of this study are in line with the debate fueled by Dahlin et al. (2018) on the need to further analyze the value of schemas and tools used to respond to failures. Codified safety management systems—in this case, the OHSAS 18001—may prove themselves economically fruitful for SMEs interested in capitalizing on their own experience with severe and fatal accidents, whose low frequency and complexity require more sophisticated information systems to fully understand their causes and, subsequently, adopt efficient corrective measures that reduce the incidence and consequences of such tragic events in the future.

Second, in a related manner, the proposed analysis of SME’s learning patterns also contributes to the small business literature by providing a context in which the ambidextrous learning hypothesis can only be partially supported because of the structural differences that characterize SMEs. Learning from heterogeneous failures is a challenging task for all businesses. From a small business viewpoint, the study’s results suggest that two factors—linked to structural and technical obstacles—might stand in the way of SMEs to learn from different types of work accidents. On the one hand, all estimations indicate that the sampled SMEs are better off at exploiting their own experience with minor accidents, which is in line with the argument that lessons learned from minor accidents can be incorporated into business routines at a relatively low economic and organizational cost (Dahlin et al., 2018). Also, these results support the notion that the archetypal structure of SMEs—which is characterized by low structural differentiation and more fluid communication—helps to neutralize the resource deficiencies faced by most SMEs (Andries & Czarnitzki, 2014; Karlsson, 2021; Runst & Thomä, 2022).

On the other hand, rather than support that ambidextrous learning is always desirable, the results stimulate the ongoing debate on the conditions under which ambidexterity adds value to SMEs (Grözinger et al., 2022; Voss & Voss, 2013). Exploration might entail technical obstacles related to the economic and organizational costs of investing in learning tools, such as the analyzed management system. In the context of this study, drastic organizational changes—in terms of evaluation processes, controls, and monitoring tasks—follow the adoption of the OHSAS 18001 (Lafuente & Abad, 2018), which might obscure the inferences drawn from work accidents. But, it was found that failure heterogeneity helps explain SMEs’ asymmetric learning pattern. This way, in settings where failures are complex and information is critical to put into action corrective measures, ambidexterity might represent a valuable approach for SMEs to offer new, unique solutions to severe and fatal accidents which are rare and often have multiple and complex causes (Desai, 2016; Dahlin et al., 2018).

In sum, the results of the study can be interpreted in terms of an asymmetric learning process in which, once accident heterogeneity is taken into account; knowledge exploitation is conducive to experiential learning when the costs of modifying safety practices are affordable and SMEs align employees’ interests toward a common goal, that is, learning from past work accidents. By contrast, for SMEs requiring specific solutions to severe and fatal accidents, ambidexterity might turn into successful learning if the acquisition of learning tools, in this case, the OHSAS 18001, facilitates the recognition of the value of the experience with this type of accident. This latter point is especially challenging for SMEs whose ability to learn relies on their resource availability and their capacity to manage the tensions that escalate between managers and employees when the business simultaneously pursues exploitation and exploration (Eggers & Suh, 2019; Voss & Voss, 2013).

5.2 Implications

This paper has implications for how SMEs can efficiently match learning solutions generated by exploiting experience with past work accidents with the organizational changes that follow the acquisition of safety knowledge (i.e., the OHSAS 18001).

5.2.1 Academic implications

This study contributes to the literature on organizational learning in, at least, two ways. First, this is the first study directly analyzing the learning pattern of SMEs from different types of failure, thus complementing existing research on learning from failure experiences (e.g., Baum & Dahlin, 2007; Desai, 2016; Kc et al., 2013; Madsen, 2009; Madsen & Desai, 2010). In addition, the direct comparison of the outcomes of learning from different types of operational failures revealed relevant insights into the hypothesized relationships between different learning processes and the learning patterns of SMEs. This analysis therefore contributes to increasing the knowledge stock dealing with learning in SMEs (e.g., Grözinger et al., 2022; Voss & Voss, 2013).

Second, in close connection to the latter point, studies on learning have examined how various forms of failure may affect learning in different ways (Dahlin et al., 2018). Not all SMEs are properly equipped to learn from different failures, and the findings of this study, which support the “small losses” hypothesis and partially support the ambidextrous learning hypothesis, contribute to advancing the strategic management literature. By showing and arguing why some forms of failure experiences produce greater learning outcomes than others in SMEs that engage in different learning processes, the framework presented in this study offers nuances to the literature dealing with the analysis of the managerial actions and operational processes that explain SMEs’ performance.

5.2.2 Implications for practice

The results suggest various implications for SME managers. First, we suggest that managers need to turn their attention to the characteristics of the business’ operational processes—which likely explain the incidence and consequences of work accidents—when considering the implementation of safety management systems. The prioritization of safety practices creates a safer working environment which fulfils the workers’ safety needs and allows them to pursue operational goals (Nahrgang et al., 2011). Any attempt to support learning by adopting a safety system—such as the OHSAS 18001—that helps mitigate work accidents should be coupled with internal information mechanisms that allow to sharing of information about work accidents while acknowledging the specific value of different types of work accidents for reducing operational losses linked to poor safety conditions. In a related manner, managerial short-termism can be detrimental to work safety investments (Lafuente & Abad, 2018), while the stigmatization of failure might obscure the inferences that can be drawn from work accidents (Madsen & Desai, 2010). Under these conditions, managers will be well advised to not ignore or stigmatize failure, and treat work accidents as opportunities to learn instead. Alternatively, SMEs should couple information-gathering processes with specific training that improves the business’ safety culture and generates greater collaboration and commitment among employees, as these elements allow learning processes to work for a long time.

Second, this research offers insights into how to cope with the potential incompatibilities between safety management systems and the properties of work accidents. Not all SMEs have the resources necessary to create the structures that facilitate ambidextrous learning. Based on results suggesting that SMEs learn from their own experience with minor accidents and that ambidextrous learning is a desirable approach to learning from severe and fatal accidents, the prescription is to link the adoption of safety systems to specific mechanisms that help to overcome potential barriers to knowledge creation and retention that properties of severe and fatal accidents (i.e., uniqueness and rarity) may create.

5.3 Lines of future research

As with any study exposing a theoretical framework to empirical scrutiny, this research presents various limitations that, in turn, represent avenues for future research. First, like other studies on learning, the data do not permit the direct analysis of the underlying learning process followed by SMEs. Various interpretations of how SMEs interlock exploitation and exploration are presented; however, this study does not evaluate how SMEs collect and store new knowledge from work accidents nor assess the processes through which cumulative knowledge is transferred to employees. Further research on these learning processes would be valuable. For example, future studies should evaluate the managers’ response to work accidents and determine if SMEs with more accurate documentation processes adopt a greater number of corrective actions and if these processes enhance learning and minimize the threat of superstitious learning (Argote et al., 2021). In a related manner, future work should analyze the conditions under which different ambidexterity models (e.g., sequential, simultaneous, or contextual) are more conducive to learning in SMEs (Raisch & Birkinshaw, 2008).

Second, although the study’s analysis distinguishes minor from severe and fatal work accidents, underlying my modeling strategy is the assumption that SMEs’ learning ability is conditioned by the consequences of work accidents, irrespective of their causes. Haunschild and Sullivan (2002) show that accidents with more heterogeneous causes create more learning. A promising future research area involves studying the learning patterns of SMEs according to the causes of work accidents.

Finally, businesses are increasingly integrating different management systems in order to reduce the duplication of managerial tasks as well as the economic losses that result from operating with multiple management systems in parallel (Abad et al., 2014). From a strategic perspective, specifically designed future research can address this point by evaluating whether exploratory safety learning is compatible with other management systems adopted by SMEs.