The impact of multiagent systems on autonomous production and supply chain networks: use cases, barriers and contributions to logistics network resilience

Automation and autonomy in supply chains

Increasing automation and autonomy shape the present development of LSCM and will continue to do so in the foreseeable future (Nitsche et al., 2021). Automation has already been a key focus of the third industrial revolution, and autonomy is playing a leading role in the ongoing fourth industrial revolution, triggered in the wake of the IoT (Schwab, 2017). Both concepts, in essence, refer to processes happening without direct human intervention and tend to be used interchangeably. However, there is an important difference: automation refers to systems that are programmed by humans to act in an exactly specified way and under certain circumstances. In contrast, autonomous systems can act on their own and are able to change initially designed paths of action within boundaries defined a priori by designers (Wooldridge and Jennings, 1995; Wooldridge, 2002). Consequently, autonomous systems have greater flexibility and the ability to self-govern.

Although automation has been achieved in many processes in recent decades, autonomy in production or logistics networks has not yet lived up to its initial promise. When it comes to the differentiation between automation and autonomy, various authors have developed their own taxonomies over the years (Sheridan and Verplank, 1978; Miller and Parasuraman, 2007; Johnson et al., 2009). Some suggest classifying a system according to its level of autonomy, which typically ranges from full human control to a fully functional autonomous system that works without any human interference. Dumitrescu et al. (2018) propose five evolutionary stages in the autonomy of technical systems that range from remotely controlled systems to fully autonomous systems. At each stage, the amount of human user control is reduced, whereas the level of control by the system, its decision-making and ML gradually increase, leading to autonomous systems that are capable of solving complex problems and interacting with other systems without the need for direct human interaction. The advanced stages of this evolutionary process are closely linked to the deployment of AI, which leads to systems that have agency of their own (Baryannis et al., 2019). Recent advancements in AI have made the vision of fully autonomous systems increasingly realistic (Baryannis et al., 2019). As AI can still face computational boundaries due to the complex nature of many problems, DAI has been suggested as a solution that would divide the computational load across a network of smaller agents. MAS therefore offer a goal-oriented solution for agent collaboration within the field of DAI (Balaji and Srinivasan, 2010).

Multiagent systems

Based on Wooldridge and Jennings (1995), Wooldridge (2002, p. 5) defines an agent as “a computer system that is situated in some environment, and that is capable of autonomous action in this environment in order to meet its design objectives.” Any control system can therefore be seen as an agent, and the more complex the environmental control system becomes, the more complex and comprehensive the decision structures become. Taken together, individual intelligent agents can become part of a larger system, labeled as MAS. Such a decentralized system consists of multiple autonomous agents that each pursue their individual objectives and execute activities in parallel (Wooldridge, 2009). They can solve complex problems by dividing them into simpler subproblems and offering companies features such as decentralization, modularity, flexibility and robustness (Skobelev and Trentesaux, 2017), which help to increase the resilience of the overall system (Vistbakka and Troubitsyna, 2021). Different levels of autonomy within MAS exist, ranging from machines that are able to solve simple mathematical functions to fully autonomous individual agents that evolve over time through ML (Dorri et al., 2018; Moyaux et al., 2006).

MAS have existed for several decades, and previous research has already confirmed their importance in areas such as real-time manufacturing through the modeling of competitive manufacturing systems, the capturing of manufacturing systems’ evolutionary development and the designing of rational agents, which together allow for improved qualitative analysis and the development of simulation frameworks as well as real-life applications (Dominguez and Cannella, 2020; Lee and Kim, 2008).

MAS have gained renewed attention following the rise of data-driven technologies at the beginning of the current decade. The overall number of MAS implementations is increasing, and numerous application areas have been identified, including scheduling, coordination between enterprises, information sharing, order fulfillment processing, collaborative production planning, provider selection, remanufacturing and resilience (Dominguez and Cannella, 2020). However, the technology has not yet reached its predicted levels of use (Karnouskos and Leitão, 2017; Xu et al., 2021). Existing studies on specific industries, such as the oil and gas industry, have found that several companies already apply MAS systems, but existing implementations strive to solve rather well-defined supply chain problems, such as logistics planning or process optimization. In contrast, implementations for more complex problems covering the whole supply chain are rare (Hanga and Kovalchuk, 2019).

Although there are considerable benefits from MAS, they also have weaknesses that inhibit their practical deployment (Dorri et al., 2018). A major issue is that decentralized systems only provide local optimization and, when combined, may underperform compared with a centralized solution (Treiblmaier, 2018a). The high interdependency of the different agents makes finding an optimal solution difficult, and a great number of interactions between agents can also lead to long computation times. Problems that cannot be divided into subproblems are also unsuitable for MAS. Implementation and monitoring are costly, and the output of the system can be incomprehensible and unpredictable, as MAS can lead to emergent and unpredictable behavior (Jennings, 2000). These downsides result in six barriers, as summarized by Karnouskos and Leitão (2017), which hinder the widespread implementation of MAS in industry. These are change management – describing the complex process of MAS adaptation, lack of technological maturity, costs for development, maintenance and implementation, lack of standardization, problems finding suitable and feasible practical applications and the ability to measure the total benefits of the system.

As an evolving technology, the widespread industrial adoption of MAS depends on how well these barriers can be dealt with. Although several benefits can be relatively easily shown at the level of proof-of-concept, showing true business value in the long run is much harder (Bergenti et al., 2004). Furthermore, the business benefits of MAS are notoriously difficult to quantify in advance. The technology requires acceptance by all relevant stakeholders, which is particularly difficult to achieve in tightly controlled, hierarchical industrial settings. To ensure this acceptance, it is therefore crucial to identify and evaluate specific MAS use cases and to understand the barriers hindering their adoption.

Resilience and the role of multiagent systems

In general, resilience can be defined as the capability of a supply chain to return to normal operations or even to an improved state after being disrupted (Christopher and Peck, 2004; Pettit et al., 2010). Building on this core objective, a broad field of research has emerged, and several definitions have been developed that highlight various characteristics of resilience (Shishodia et al., 2021). An overview of resilience definitions in logistics and supply chain research is given in Table 1. To achieve resilience, proactive strategies can be established to ensure that the probability of disruptions is lowered, but reactive strategies also have to be in place to increase an organization’s capability to react to any disruptions (Tukamuhabwa et al., 2015). Although the proactive part is of importance, a major portion of resilience research focuses on the reaction to disturbances. What most definitions have in common is the ability to recover from disruptions and return to normal operations. In this regard, the speed of this process is crucial to becoming resilient (Brandon-Jones et al., 2014; Sheffi, 2005; Tukamuhabwa et al., 2015).

Pettit et al. (2010) list several vulnerabilities in the supply chain (i.e. turbulence, deliberate threats, external pressures, resource limits, sensitivity, connectivity, supplier/customer disruptions) and derive important capability factors (e.g. flexibility, capacity, efficiency, adaptability, collaboration) that can be leveraged to address these vulnerabilities. MAS in LSCM can help alleviate several of these issues by allowing lengthy and complex processes, which often include negotiations with other entities, to be run autonomously in a faster and more efficient way. Besides gains in the productivity of processes, MAS also offer the potential to react to unforeseen changes in a network in a timely manner and leverage the knowledge stored in the network as opposed to tacit knowledge in individual units.

As a contribution to logistics network resilience, MAS can increase organizational flexibility through modularity and are able, through decentralization, to offer new solutions that cannot be found in purely centralized systems (Huhns and Stephens, 1998). Moreover, MAS reduce the need for computational power, as only subsystems need to be optimized. In the wake of Big Data, highly complex centralized systems are consuming ever more resources (Schwartz et al., 2020). In this regard, MAS offer an effective way to design efficient distributed computer systems (Huhns and Stephens, 1998). Additionally, the division into subsystems makes the programming of MAS relatively easy for developers owing to the reduced complexity of the system. One resulting benefit is the ease of making modifications, as a modular system makes it easier to change a single agent, thereby reducing configuration costs and increasing system reusability. Finally, the multidimensionality of MAS simplifies problem-solving in complex environments, and the combination of decentralization and modularity lends itself to situations that are likely to change frequently (van Parunak, 1998; Treiblmaier, 2018a).

Given the core characteristics of autonomous MAS, it can be concluded that they can support companies in increasing their organizational productivity and making their networks more resilient. Furthermore, gaining independence from individual human knowledge has become an important driver for LSCM, especially during the COVID-19 pandemic (Modgil et al., 2022a; Modgil et al., 2022b; Nitsche and Straube, 2021). During this pandemic, the need for more resilience in future supply chains became evident, and process automation is now seen as one important cornerstone in achieving it (Kiers et al., 2022). Process automation can be seen as another evolutionary step in the industrial evolution that substitutes humans with machine labor and, in doing so, reduces the number of jobs needed in the industry (Schmidpeter and Winter-Ebmer, 2021; Petropoulos, 2021).

Expert selection

Selecting the right experts is crucial for the success of the Delphi method, as participants who are highly qualified in a subject contribute the best possible insights and predictions for a research topic (Alarabiat and Ramos, 2019). To assemble a suitable group of industry experts, potential candidates from research and industry with known backgrounds in LSCM as well as autonomous technologies were contacted and asked if they considered themselves knowledgeable regarding MAS or, if this was not the case, to forward the request for participation in the Delphi panel to more suitable colleagues. Although there were no geographical restrictions, the participants were required to be able to complete the questionnaires in English or German. A group of 18 participants took part in all three rounds. Our number of participants slightly exceeds the recommended number of 10–15 experts for Delphi studies to gain meaningful results (Delbecq et al., 1975; Adler and Ziglio, 1996; Alarabiat and Ramos, 2019). To spark creativity and allow for well-rounded opinions, the Delphi panel also included two consultants and two academic researchers.

Table 2 shows the characteristics of the Delphi panel participants, including their country of origin, industry type, number of employees, annual turnover, management level and years of experience in LSCM. On average, the study participants had 13 years of professional experience, which considerably exceeds the recommended domain experience of at least five years (Durach et al., 2017).

Delphi study design

After confirming their qualifications of relevance to autonomous systems, the experts were briefed about the goals and the planned rounds of the Delphi study, which are detailed in the sections below (see Figure 1).

First round: The aim of the first round of the Delphi study was threefold. First, the experts were tasked with identifying and describing specific use cases for MAS in LSCM. These use cases should be complex and ideally included interactions with other supply chain participants, as well as exhibiting a high potential for autonomization. Second, the experts needed to name potentials that they would expect from an MAS implementation and that they would use to evaluate a specific application. Third, they identified barriers that hinder the widespread implementation of such MAS. The experts were asked to answer each question openly and to name as many use cases, potentials and barriers as they could think of. To synthesize the results, Q-methodology was applied to each question individually (Ellingsen et al., 2010). This method enables a structured synthesis of a variety of answers and has previously been applied in LSCM research to synthesize large numbers of qualitative answers (Durach et al., 2015; Nitsche and Durach, 2018). This was done by writing down each use case mentioned by a study participant on an individual card. Subsequently, two researchers individually performed the Q-methodology process, reading one card after another and either opening up a new use case group in case no thematically similar use group already existed or placing it into an existing and similar group of use cases. By executing this procedure, each researcher individually came to a synthesis of use cases that were presented to the other researcher. Afterwards, the similarities in assignments were identified and differences discussed to find a common understanding. As a result, a set of 11 use cases was synthesized that provided the basis for the further assessment of the potential of the use cases in the subsequent rounds.

Second round: The goal of the second round was to assess the potential of the identified MAS use cases. Building on the results of the first round, the potential of each MAS use case was identified in the synthesis as either improving the resilience of the logistics and supply chain network or increasing organizational productivity. Additionally, we asked for a thorough assessment of the implementation complexity of each use case. In the second and third rounds, the 11 use cases identified were assessed using a five-point differential scale with endpoints 1 (low) and 5 (high) using the following three assessment criteria:

1. Potential for logistics network resilience (PLNR): The potential of a use case to increase the capability of a logistics/supply chain network to go back to regular operational functioning within a tolerable time interval after being disrupted.

2. Potential for productivity increase (PP): The potential of a use case to perform at least one organizational process more efficiently (e.g. faster or at lower cost).

3. Reduction of the complexity of implementation (CI): A decrease in the level of resources needed to deploy a project.

Third round: In the third round, the experts were given the aggregated results for each of the three metrics (i.e. PLNR, PP and CI) applied to assess each use case and were given the chance to adjust their initial answers based on this new information and to provide additional information to outline their decisions. Based on the results of this final round, we calculated important statistics for each use case to determine whether a group consensus had been reached. Such consensus is the case if the interquartile range (IQR) lies beneath a certain threshold. A consensus confirms the results and indicates high forecast accuracy (Hahn and Rayens, 1999; Raskin, 1994).

Identification of promising MAS use cases

Table 3 outlines the 11 use cases (UC) that the experts identified as having the most potential for MAS-supported automation. For all use cases, it was considered beneficial to include multiple entities with partially different target systems in decisions related to problem-solving. These use cases illustrate the broad applicability of MAS. In UC1, search agents check databases to discover suitable suppliers for required tasks and materials. Interested suppliers can register in the databases to be visible to search agents. Once the agent finds a suitable supplier, it notifies the responsible employee or another agent to initiate a contract process. In UC2, renegotiating expiring contracts or creating new ones can be automated by letting agents take the place of negotiators. These agents are configured to represent the goals of individual companies. For a more complex problem, agents from different companies can come together to achieve an optimal collaborative planning scenario. UC3 describes how agents can be used to monitor demand and coordinate purchasing. All involved agents are in constant communication with one another, in accordance with the metrics agreed. Demand and stock are constantly evaluated, and new orders are generated automatically. In UC4, agents check truck load sizes, optimal transport routes and available transport offers for specified routes. Should a match with a third-party logistics (3PL) or fourth-party logistics (4PL) provider occur, an agent can automatically book the optimal transportation solution. UC5 illustrates how two separate agents can optimize production and transport plans in accordance with one another, finding the optimal solution for the entire supply chain. Should one of the two subsystems be disrupted, the agents can work together to create an automated, viable solution in real time. In UC6, an agent constantly monitors whether the current production plan is still optimal, based on new information that can be added in real time to the system. Should such data lead to a new optimal plan, a production agent is informed, who checks the plan for feasibility before implementing it as an updated production plan. In UC7, different agents take into account staff availability, workload and flexibility to optimize personnel scheduling. The information is regularly passed on to another agent responsible for generating an optimal schedule. UC8 involves different agents that track the demand and availability of intralogistics components and match them accordingly, which includes the scheduling of forklifts and automated guided vehicles (AGVs). In UC9, an agent regularly updates the list of outgoing orders, sorting them by urgency. This information is given to another agent, which plans the sequence of commissioning and outbound goods. In UC10, an agent scans databases, such as news sites, for disruptive events. Relevant instances are reported to an event management agent, which then reconfigures the supply chain to limit the events’ impacts and implements the new optimal solution. Finally, in UC11, each agent becomes part of the supply chain and can be used to create a simulation to optimize and verify solutions for the whole value network.

As described in the introduction, MAS are used when separate entities need to be involved in complex decision-making processes that require automation. Participants were aware of the crucial role of ML in solving complex MAS problems; this role was also explicitly included in the definitions of the use cases. The extent to which the agents should be equipped with ML capabilities or whether predefined rules are sufficient was not explicitly prescribed for each use case. However, due to the complexity of the use cases, it can be assumed that their automation would frequently require intelligent agents that can learn and adapt over time.

Assessment of MAS use cases

To answer RQ2a and RQ2b pertaining to the different potentials of the respective use cases, as well as the applicability of these use cases to the strengthening of supply chain resilience and productivity, the experts used the assessment criteria PLNR and PP, as presented above. Additionally, we asked about the CI. The results can be found in Table 4, which is ordered by the resilience potential (PLNR) of the respective use cases. For each case, the arithmetic mean of the assessment criteria is shown, as well as the IQR and the convergence rate (CV). The IQR is used to measure the consensus of the expert group and is calculated by subtracting the results of the first quartile (25%) from the third quartile (75%). The CV is the decrease in standard deviation between the rounds of the Delphi study, in this case, between rounds two and three. A negative CV indicates a convergence toward a group consensus. Furthermore, we calculated the difference between the scores of the PP and the PLNR and weighed these scores against the CI. These deltas can be used to assess and compare the relative benefits of the corresponding use cases. Those with positive scores can be seen as high-value use cases and should be prioritized when it comes to MAS implementation in practice, as their potential benefits (e.g. increases in either productivity or resilience) outweigh the CI. It should be noted that a use case does not necessarily have to address both benefits; rather, LSCM managers need to start thinking about and prioritizing applications in light of their respective demands to determine whether the goal of autonomizing the process concerns either increases logistics network resilience or productivity. Although Table 4 illustrates that all use cases have above-average assessment scores for each of the assessment criteria, the results still diverge, depending on the nature of the use case itself.

Concerning the potential of PLNR, automated risk and event management is the most promising use case. However, every use case that automates material flow achieves a high score, as MAS allows for the rescheduling of material flows in real time and swift responses to unforeseen events. Notably, automated supplier negotiations (the only MAS that received a score below 3 for PLNR) are not seen as very relevant to resilience, as they are regular, singular events with limited potential for automation.

When it comes to productivity increases, the use cases that exhibit the highest potential are those that relate to automated rescheduling and production planning, synchronized transport and production planning and inbound supply planning and purchasing. Since production and transport are at the core of any supply chain and require substantial coordination effort, they benefit the most from autonomous processes. In comparison, automated negotiations are also not highly ranked for PP, since negotiations have a distinct social component and can be hard to automate, which is especially true if two or more parties are pursuing conflicting goals.

The CI is highest for those processes that deal with a lot of data or involve multiple different parties. The automated rescheduling of production plans and their synchronization with transportation are examples of such complex problems. In comparison, in-house activities, such as automated order processing in warehouses and the planning of intralogistics, are comparatively low in complexity. These use cases can be used to test and gain experience for companies starting out with deploying MAS, as they offer sufficient potential while simultaneously having low implementation complexity.

The deltas show that swift implementation is not desirable for all use cases from a resilience or productivity perspective, as their benefits may be simply too low or the complexity too high. These include automated rescheduling and production plan automation (UC6), the simulation of supply chain networks for optimization purposes (UC11), automated personnel scheduling (UC7) and automated supplier negotiations (UC2). The optimization of transport and production plans (UC5) as well as inbound supply planning and purchasing (UC3) score positively when it comes to organizational productivity but negatively from a resilience perspective. Automated risk/event management (UC10) is the only use case that has a negative score from a productivity perspective, but has a positive resilience assessment. Two cases have two positive deltas, identifying them as applications that should be prioritized by companies wanting to implement MAS solutions. These include automated transport planning (UC4) and automated order processing/prioritization (UC9).

We defined a threshold value for the IQR of ≤25% for the three criteria (PP, PLNR and CI) to test consensus in group opinion, which is in line with the recommendations of previous Delphi studies (Hahn and Rayens, 1999). Given the use of a five-point scale, consensus is thus reached if IQR ≤ 1.0. In total, consensus was achieved for 29 out of 33 predictions (88%), which is an above-average group consensus for a Delphi study, as most results vary between 22% and 85% (Hahn and Rayens, 1999).

Barriers to MAS use cases

To answer RQ3, the participants were tasked with naming the relevant barriers to the implementation of MAS. Four main barriers emerged, which we summarize in a qualitative manner:

Standardization: Current MAS applications are highly customized solutions implemented by single companies. Consequently, standardization has not been achieved yet, and one-size-fits-all solutions are not readily available on the market. This makes the scalability of current MAS applications very difficult, and connection with third parties is also impeded by the lack of widespread standards. Even though there have been efforts to standardize MAS, these have mainly come from academia and do not specifically consider industrial requirements.

Technological maturity: This barrier addresses the technological complexity of solutions and the readiness of these technologies for wider use. The experts indicated that companies may doubt the technological readiness of MAS; thus, a MAS solution can always be questioned, as the problem-solving process is complex and may not always be transparent. As a result, potential use cases might be discarded or canceled due to their complexity. Most companies have no or only a few MAS experts who are qualified to develop these complex solutions and dispel doubts about the technology’s maturity. However, the growing number of academic MAS publications and the proliferation of industry projects indicate that MAS might reach a sufficient level of maturity in the foreseeable future that will then be acknowledged by industrial actors.

Costs: In addition to the high costs for software acquisition and development, deployment, management and maintenance, personnel training and hardware acquisition, the generation of high-quality data can be very expensive, a problem that is not exclusive to MAS. Setting up the master data or migrating it from legacy systems can consume a lot of effort and time. Automated data acquisition needs to be implemented in the entire process so that the whole MAS system can function fully autonomously. These requirements can easily lead to soaring costs.

Change management: Decision-makers and management are often unaware or skeptical of MAS applications, which can be partly attributed to the barriers discussed above. Furthermore, MAS might lead to a loss of control because of decentralization and autonomization. Apart from the previously mentioned need for successful MAS projects as positive role models for decision-makers, systems need to be designed in a way that allows for human interaction and intervention, at least in the first years of MAS application. This would prevent a feeling of fully handing over control to an automated system and could help to increase trust in the technology. Employees may also fear being replaced by an MAS solution. Good change management is therefore essential, with employees being involved in the development and implementation of MAS. They also need to be prepared to take on new tasks when their current work is replaced by automated processes. During change management, it is important to pay particular attention to the behavioral factors that prevent employees from supporting technological change. Personal attributes, values and fears must be addressed; otherwise, this barrier can lead to a situation in which a technologically mature solution has been developed but cannot be implemented. On a broader scale, this also raises the issue of finding employment opportunities for humans, who are increasingly being replaced by technology.

Theoretical implications

In this study, we investigate the current and future development of MAS in LSCM and identify and assess possible use cases. In a rigorous three-round Delphi study, a total of 18 experts compiled a comprehensive list of use cases, which presents a starting point for other researchers to focus on one or several of the identified cases or to add further applications to the list. More in-depth knowledge of the impacts and benefits of MAS applications in LSCM can reduce the CI and lower the threshold for further industry integration.

We approached the assessment of the cases from the viewpoint of resilience theory, which, in a network setting, refers to the capability of a supply chain to return to normal operations or even to an improved state after being disrupted (Christopher and Peck, 2004; Pettit et al., 2010). In this regard, our study makes an incremental contribution to previous research by identifying capability factors that can be leveraged to address supply chain vulnerabilities, with the goal of increasing resilience (Pettit et al., 2010). MAS might, by definition, be unable to identify black swan events in advance, but they can offer increased flexibility when such events occur and enable companies to react to unforeseen events more quickly, thus supporting faster recovery. Our findings identify those MAS use cases with the highest potential to increase logistics network resilience, namely, automated risk/event management, transport planning, rescheduling and production plan optimization and combination/optimization of transport and production plans. Furthermore, we also take into account the implementation complexity of all use cases and have been able to pinpoint applications that might be advantageous for companies to try out first. Future theory-based research can build on our findings by:

• taking a behavioral viewpoint and investigating adoption antecedents and consequences of a specific implementation; or

• taking a design-oriented perspective and helping to create solutions that help companies to make their own networks more efficient.

As indicated above, the implications of our findings may have both intended and unintended consequences for the organization itself and its workforce. The main focus of our study was the investigation of how MAS can help improve an individual organization’s performance or strengthen the resilience of the network it is operating in, but the change processes that accompany these digital transformation processes must not be neglected. In this regard, the theory of organizational learning (Crossan et al., 2011) and organizational behavior theories (Miner, 2003) can help to get a better understanding of how organizations and their workforce might respond to the changes that are induced by MAS. Moreover, our findings can help to identify technologies that might impact organizational structure and processes in the not-too-distant future.

Another theory-building outcome of this study is the confirmation and refinement of previous research on MAS implementation barriers (Karnouskos and Leitão, 2017). Future research needs to focus in more detail on ways to overcome these barriers and enable easier MAS implementations. Barriers always need to be considered for specific scenarios and can potentially be removed as the industry matures and knowledge of MAS becomes more widespread. Some of these barriers take the form of obstacles resulting from digitalization in LSCM in general (Agrawal et al., 2019).

Managerial and societal implications

This study has several important implications for practitioners. Management can use the suggested list of MAS application scenarios to identify and assess implementation opportunities within their own supply chains. The assessment of the respective use cases gives practitioners guidance on which MAS cases to start with. If a company has no prior experience with MAS, use cases with a low CI and reasonable profits might be recommended to first try out the technology. The assessment criterion (i.e. delta) that we used to measure the difference between the expected benefits and implementation complexity also provides interesting insights for practice, since managers with little experience in MAS can identify those use cases that promise high benefits in relation to their expected complexity. Examples include the automation of in-house transport planning and order processing.

It is to be expected that with growing experience in the field of MAS, companies will be able to maximize the benefits and lower the costs of implementation over time. In case organizational subsystems have already been implemented that use MAS, developers in the company will have earned relevant experience, and management and employees will have gained trust in the technology. When implementing MAS, it is also advisable for management to keep a close eye on the barriers presented in our study, especially the lack of standardization, missing industry experience and unpredictable costs. Furthermore, managers need to establish change management processes that not only take into account the interests of a company’s shareholders but also its employees. Our list of barriers can thus help to enable mindful implementation, show management which challenges are likely to arise and give them some ideas on how to adequately address these.

Even though our study focuses on the potential of MAS for increasing resilience in logistics and supply chain processes, the societal implications of these changes must not be neglected. In general, optimized supply chain processes contribute to societal welfare and value generation, supporting the more effective and efficient use of existing resources (MacCarthy and Ivanov, 2022). However, the autonomization of previously manual processes could have a detrimental impact on employment levels and lead to deskilling in supply chains. The use cases developed by the Delphi panel illustrate that many of these processes are currently very labor-intensive and are conducted manually, with little technological assistance. Autonomizing such processes will have direct implications for the workforce and raise various questions that have not yet been adequately investigated. Several authors already highlighted the significance of the human factor in the introduction of AI systems (Dora et al., 2022; Dwivedi et al., 2021; Hoberg et al., 2020; Klumpp and Ruiner, 2022). However, most current research focuses on the technical implementation of specific UCs, as demonstrated in the literature reviews by Pournader et al. (2021), Riahi et al. (2021) and Toorajipour et al. (2021). So far, few articles have investigated the employment implications of AI systems in general and MAS systems in particular.

In highly digitalized and partially autonomous supply chains, the competence profiles of logistics managers will need to change. It is currently unclear to what extent systems can run autonomously or when employees will need to intervene in systems whose decision-making processes can only be understood to a limited extent. Consequently, Hoberg et al. (2020) point out that, despite technological advancements, it is unlikely that human decision-making will ever be completely eliminated. For the supply chains of the future, Hoberg et al. (2020) favor a combination of the unique capabilities of human decision-makers with AI-based guidance to enable complex problem-solving.