Modeling Organizational Resilience in Human-Cyber-Physical Systems (Industry 5.0) Through Collective Dynamics, Decision Scenarios and Crisis-Aware AI: A Multi-Method Simulation Approach

Abstract

Supply chain disruptions during the COVID-19 pandemic exposed structural vulnerabilities of centrally controlled manufacturing systems, motivating renewed interest in organizational resilience within the context of Industry 5.0 human–cyber–physical systems. This study investigates how organizational decision-making paradigms and crisis-aware artificial intelligence (AI) jointly influence performance, crisis response, and recovery. An agent-based modeling (ABM) framework is developed to compare centralized, distributed, and self-organized organizational structures across 650 simulation runs under a controlled supply side disruption. A crisis-aware Q-learning architecture enables AI agents to shift from efficiency-oriented to stability-oriented strategies when resource scarcity is detected. To avoid baseline-dependent bias, resilience is evaluated using an absolute, capacity-normalized metric. Results indicate that self-organized systems consistently outperform centralized and distributed structures in baseline performance, crisis throughput, and recovery speed. The integration of crisis-aware AI further increases absolute resilience by approximately 10.7% and enables substantially higher throughput during disruption compared to hierarchical control. Enhanced performance is primarily driven by adaptive coalition formation, proactive resource conservation, and rapid post-crisis recovery supported by preserved coordination structures. These findings provide quantitative support for Industry 5.0’s human-centric principles and show that decentralized decision-making augmented by context-adaptive AI offers a robust organizational design strategy for volatile manufacturing environments.

1. Introduction

The manufacturing and organizational landscape is undergoing a profound transition from the automation-driven paradigm of Industry 4.0 to the human-centric vision of Industry 5.0 [1,2]. While Industry 4.0 focused on digitalization, cyber-physical systems, and large-scale automation, Industry 5.0 emphasizes human–AI collaboration, adaptive decision-making, and organizational resilience in environments characterized by turbulence, uncertainty, and frequent disruption [3]. Recent global crises, from the COVID-19 pandemic to semiconductor shortages and geopolitical supply chain shocks, have revealed the structural fragility of traditional, centrally controlled production systems [4,5]. The practical consequences of such structural fragility are well documented. During the 2021 global semiconductor shortage, automotive manufacturers relying on highly centralized procurement and production planning experienced prolonged plant shutdowns, with average production stoppages exceeding four weeks and estimated global losses surpassing USD 200 billion [6]. In contrast, firms operating with more distributed supplier networks and decentralized decision authority were able to sustain partial production and reallocate resources more rapidly, maintaining substantially higher operational capacity throughout the disruption [6,7]. Similarly, during the COVID-19 pandemic, centralized pharmaceutical manufacturing systems required several months to reconfigure production lines in response to demand shocks, whereas loosely coordinated networks of smaller manufacturers and contract producers achieved faster reconfiguration through flexible task allocation and emergent collaboration [7,8]. These cases illustrate that organizational structure is not a neutral backdrop but a decisive factor shaping crisis response and recovery. Organizations that endured such shocks most successfully exhibited distributed autonomy, emergent team formation, and rapid reconfiguration [9], yet management science still lacks a unified, quantitative framework explaining why certain organizational structures recover faster and maintain performance under identical environmental pressures.

In this context, human–cyber–physical systems (HCPS) have emerged as a foundational concept for Industry 5.0, integrating human expertise, artificial intelligence (AI), and physical assets into adaptive decision networks [10]. From a technical perspective, HCPS provides concrete enablers for adaptive organizational designs. Digital twin infrastructures offer real-time visibility into production states, task queues, and resource availability, reducing information asymmetries that traditionally justified hierarchical oversight. Low-latency communication architectures enable distributed actors to access synchronized operational information, supporting local decision-making without coordination delays. These systems promise substantial gains in resilience and flexibility, but they also expose a fundamental tension: centralized AI-driven optimization can maximize efficiency in stable environments, while decentralized human autonomy preserves adaptability during crises [11]. Existing research does not resolve how these forces should be balanced, nor how different organizational structures interact with AI in dynamic conditions.

Three specific gaps limit current understanding. First, the literature typically evaluates organizational decision paradigms in isolation [12]. Second, existing AI applications in manufacturing, focused on predictive maintenance, scheduling, or quality control [13], generally assume stationarity and lack crisis-aware architectures capable of shifting strategies when disruptions invalidate historical patterns. Third, widely used resilience metrics calculate proportional recovery relative to each system’s baseline performance [14]. This approach creates a systematic bias against high-performing organizations, which experience larger proportional drops even though they maintain higher absolute throughput during crisis events.

Recent advances partially address these gaps. Capacity-anchored resilience indices have been proposed to normalize crisis performance relative to operational capacity rather than individual baselines [15], while empirical studies demonstrate that absolute metrics better predict supply chain failure and bankruptcy outcomes than proportional measures [16]. In AI domains, multi-objective reinforcement learning agents capable of shifting between efficiency and robustness objectives have been shown to improve crisis retention [17], and adaptive confidence calibration has been demonstrated to enhance AI recommendation acceptance rates under uncertainty [18]. However, these advances remain fragmented across disciplinary silos, and no unified framework integrates organizational structure, crisis-aware AI, and absolute resilience measurement.

Research on collective dynamics identifies three canonical organizational decision structures. Centralized systems rely on hierarchical command and control, achieving efficiency under stability but suffering information bottlenecks and decision overload during disruptions [19]. Distributed systems deliver authority across semi-autonomous departments, offering greater flexibility but risking coordination failures when cross-unit resource reallocation becomes critical [20]. Self-organized systems, in contrast, rely on emergent coordination, dynamic communication networks, and autonomous coalition formation [21], properties associated with high adaptability but insufficiently validated through controlled, comparative experimentation.

Parallel advances in AI for organizational decision support have largely focused on efficiency maximization. Reinforcement learning (RL) offers a promising alternative for dynamic environments [22], yet organizational adoption faces challenges, including partially observable states, delayed rewards, and shifting trust dynamics, particularly during crises when workers may be skeptical of algorithmic recommendations [23]. The need for crisis-aware AI agents capable of adjusting strategies, emphasizing stability over short-term efficiency, and integrating with human decision processes remains largely unexplored.

Recent Large Language Models (LLMs) offer alternative paradigms for AI-assisted coordination through natural language interfaces and semantic reasoning. However, this study employs reinforcement learning for three reasons: (1) explicit reward optimization aligns with quantifiable manufacturing objectives; (2) numerical state representations enable real-time adaptation without LLM inference overhead; and (3) interpretable action spaces facilitate trust calibration in high-stakes settings. Future research should explore hybrid architectures combining LLM-based semantic reasoning with RL-driven tactical coordination.

The measurement of organizational resilience further complicates evaluation. Classical metrics compute proportional recovery (retention × recovery speed), unintentionally penalizing high-performing systems that maintain higher absolute throughput despite larger proportional drops [17]. This study addresses this bias through capacity-normalized resilience metrics.

This research addresses these gaps by developing an integrated multi-method simulation framework that unifies organizational structure, individual and collective decision-making, and crisis-aware AI assistance. Specifically, three organizational paradigms (centralized, distributed, and self-organized) are examined across multiple AI conditions (no AI, generic AI support, and crisis-aware AI), under a controlled supply chain disruption. In this study, multi-method simulation denotes the integration of agent-based modeling for organizational structure, reinforcement learning for adaptive AI behavior, and dynamic network analysis for coalition formation and rewiring. Together, these methods enable a multi-level analysis of organizational resilience that spans micro-level decision rules, meso-level coordination patterns, and macro-level performance outcomes.

• This study pursues three specific research objectives aligned with the challenges of Industry 5.0 organizational resilience: To quantitatively compare organizational resilience across decision-making paradigms by evaluating baseline performance, crisis throughput, and recovery dynamics in centralized, distributed, and self-organized structures under identical disruption conditions;
• To identify the mechanisms through which crisis-aware AI enhances organizational resilience, distinguishing between direct performance effects and indirect coordination effects such as coalition formation and workload redistribution;
• To assess the validity of absolute, capacity-normalized resilience metrics in comparison with traditional proportional measures, examining their ability to produce rankings consistent with operational viability during disruptions.

These objectives guide the design of the simulation framework, the experimental protocol, and the interpretation of results presented in the following sections.

The remainder of the article is structured as follows. Section 2 reviews the literature on Industry 5.0, organizational resilience, collective dynamics, decision paradigms, and AI in socio-technical systems. Section 3 details the simulation methodology, agent models, decision structures, disruption protocol, and resilience metric. Section 4 presents results on performance, crisis response, recovery, network dynamics, and AI impact. Section 5 discusses theoretical and practical implications, limitations, and opportunities for future research. Section 6 concludes with insights for organizational design in human–AI collaborative environments.

2. Literature Review

2.1. Organizational Resilience: Concepts and Frameworks

Organizational resilience is increasingly recognized as a defining capability in environments characterized by uncertainty and disruption. The concept has evolved beyond the early engineering view of rapid recovery [24,25] toward ecological and socio-technical perspectives that emphasize the capacity not only to withstand disturbances, but also to adapt and reorganize while maintaining essential functions [26,27]. Contemporary frameworks describe resilience as a multistage process involving preparation, impact absorption, recovery, and long-term adaptation [28]. Despite extensive theorization, empirical validation remains limited because many studies rely on retrospective observations of organizations that survived crises [9], making it difficult to isolate causal mechanisms.

A persistent methodological challenge concerns how resilience is measured. The dominant approach uses proportional metrics that compare crisis performance to each system’s pre-disruption baseline [14]. Although widely adopted, this method introduces a systematic distortion: high-performing systems appear less resilient solely because their proportional decline is larger, even when they maintain substantially higher absolute output during disruptions. Recent critiques argue that such metrics conflate resilience with baseline performance and obscure meaningful comparisons across organizational structures [29]. Methodological advances propose capacity-anchored resilience indices that normalize crisis performance to operational capacity rather than individual baselines [15]. Empirical validation demonstrates that such absolute metrics outperform proportional measures in predicting supply chain failure and bankruptcy outcomes [16]. However, these alternative frameworks remain underutilized in comparative organizational studies.

2.2. Decision-Making Paradigms in Organizations

Decision-making structures strongly influence how organizations respond to disruption. Classical centralized systems concentrate authority at the top of the hierarchy and rely on a small number of decision-makers with global visibility [30]. Under stable conditions, centralized coordination can be efficient and ensure strategic alignment. However, extensive literature on bounded rationality and complex systems highlights the fragility of such structures during crises [19,31]. Decision bottlenecks, cognitive overload, and delayed sense-making limit the ability of centralized organizations to react quickly to unanticipated events.

Distributed organizational structures attempt to mitigate these limitations by delegating decision rights to semi-autonomous units [32]. This design promotes faster local responses, specialization, and adaptability. Modularity theory suggests that such architecture can contain disruptions within subunits, enhancing robustness [33]. Nevertheless, distributed systems face significant coordination challenges. Divergent local objectives, information asymmetries, and negotiation delays can hinder cross-unit collaboration [20], particularly when crises require rapid resource reallocation.

Self-organized systems represent an alternative paradigm inspired by complex adaptive systems theory [34]. Instead of relying on hierarchical control or formal coordination mechanisms, these systems enable autonomous agents to adjust their behavior based on local information, forming spontaneous coalitions and dynamically updating communication structures [21]. This emergent coordination can produce remarkable adaptability in volatile environments. Empirical examples from decentralized innovation teams and agile software development illustrate these benefits [35]. Yet unmanaged decentralization can also generate fragmented strategies, duplicated efforts, and capacity planning inconsistencies [12].

The question of whether crises favor centralized or decentralized decision-making remains contested. Some traditions emphasize the value of decisive hierarchical leadership under high stress [36], while others argue that complex crises exceed the cognitive capacity of any single decision-maker, making decentralized sense-making essential [37]. A growing consensus suggests that the effectiveness of each paradigm depends on crisis complexity rather than a simple preference for hierarchy [38].

2.3. Artificial Intelligence in Manufacturing Systems

AI has become integral to modern manufacturing systems, particularly through applications in predictive maintenance, quality inspection, and production scheduling [13]. These applications generally rely on historical data and assume stable operating conditions. Disruptions, such as supply chain shocks or abrupt demand shifts, violate this assumption and lead to degraded performance due to concept drift [39]. The instability of static machine-learning (ML) models during the COVID-19 pandemic [4] highlighted the need for AI techniques capable of adapting to non-stationary environments.

RL is a suitable alternative because it learns through interaction rather than static datasets [22]. However, organizational settings introduce complexities rarely present in robotic or control applications. Organizational environments are partially observable, rewards are often delayed, and human operators may accept or reject algorithmic recommendations based on trust rather than objective performance [23,40]. This dynamic creates a paradox in which AI systems are often least trusted precisely when they could provide the most value during disruptions.

Recent scholarship argues for the development of crisis-aware AI systems that can detect environmental shifts and adjust their strategies accordingly [41]. Sharma et al. (2024) demonstrate that multi-objective reinforcement learning agents capable of shifting between efficiency and robustness objectives achieve higher throughput retention during supply disruptions [17]. Similarly, Liu and Chen (2024) show that adaptive confidence calibration improves AI recommendation acceptance rates under uncertainty [18]. Such systems would prioritize stability and robustness during crises while pursuing efficiency under normal conditions. Although this idea aligns with the broader goals of Industry 5.0 and human–AI collaboration [1,2], few empirical implementations exist in the literature, leaving an important research gap.

The rapid emergence of LLMs and generative AI in manufacturing contexts introduces additional considerations. While LLMs offer natural language interfaces and semantic reasoning capabilities, their application to real-time operational coordination faces computational overhead and interpretability challenges. The present study focuses on RL-based approaches that provide explicit reward-driven optimization aligned with quantifiable manufacturing objectives, enabling faster adaptation cycles and more transparent action spaces for trust calibration in high-stakes settings.

2.4. Multi-Agent Systems and Emergent Coordination

Agent-based modeling (ABM) provides a powerful framework for studying how organizational resilience emerges from micro-level interactions among heterogeneous agents [42]. Unlike equation-based approaches, ABM allows researchers to simulate decision-making, communication, and collaboration processes that unfold over dynamic networks [43]. This makes it particularly suitable for examining coordination patterns, workload distribution, and adaptive behavior under disruption.

Coalition formation is a particularly relevant emergent mechanism [44]. Agents may join forces to pool skills or share workload, and the structure of such coalitions can have substantial effects on performance. The size, stability, and diversity of coalitions change dynamically during disruptions, influencing both crisis absorption and recovery. Network topology also plays a significant role. Small-world networks support efficient communication [45], modular networks promote local specialization, and scale-free networks can exhibit robustness to random failures [46]. Importantly, real organizational networks evolve in response to task demands and disruptions, reshaping communication flows and collaboration patterns. However, relatively few studies examine how these adaptive network dynamics interact with decision-making paradigms and AI assistance during crises.

Multi-agent reinforcement learning (MARL) extends RL to multi-actor environments [47] but remains rarely applied to organizational decision-making. Social dynamics such as authority gradients, trust, and negotiation norms complicate MARL adoption and require novel architectures that account for human acceptance of algorithmic recommendations.

2.5. Research Positioning and Gap Synthesis

Across the reviewed fields, several persistent gaps emerge. First, centralized, distributed, and self-organized paradigms are rarely compared within a unified modeling framework, preventing controlled evaluation of how structural differences influence resilience [12]. Second, existing AI systems in manufacturing are optimized for stable environments and lack crisis-aware mechanisms capable of adjusting strategies during disruptions [13,39]. Third, commonly used resilience metrics rely on proportional measures that systematically penalize high-performing systems and obscure meaningful comparisons [14,29], though recent proposals for absolute metrics [15,16] have not been systematically applied in comparative organizational studies. Fourth, the micro-level behavioral mechanisms underlying resilience, such as coalition formation, network rewiring, and trust-based adoption of AI recommendations, remain underexplored [44]. Finally, very few studies integrate ABM with RL [47], limiting understanding of how humans, AI agents, and organizational structures jointly influence resilience in dynamic environments.

These gaps collectively motivate the present study, which integrates structural, behavioral, and computational components into a unified framework to examine how decision-making paradigms and crisis-aware AI affect resilience in HCPS.

3. Materials and Methods

This study employs ABM as its primary methodology and enables examination of micro-level decision rules, network topologies, coalition formation, and emergent behaviors under controlled conditions [42]. Prior ABM work identifies the importance of network structure [43], but the interaction between communication topology, AI assistance, and crisis-triggered rewiring has not been systematically investigated. Furthermore, organizational settings introduce unique social dimensions—trust, cooperation, authority—requiring new architectural approaches [47].

The simulation framework integrates multiple organizational decision structures within a unified agent-based environment. A crisis-aware Q-learning AI agent dynamically shifts reward structures from efficiency optimization to stability preservation during disruptions. To enable fair comparison, an absolute resilience metric eliminates baseline-dependency bias by normalizing to system capacity and operational thresholds. The framework supports micro-level analysis of coalition formation, network evolution, and emergent coordination.

3.1. Simulation Framework

The organizational model was implemented in Python (v3.10) using the Mesa (v2.1.5) agent-based modeling framework [48], which supports discrete-time simulation of complex adaptive systems. The simulated organization consists of 100 agents divided into three functional classes: 70 worker agents executing production tasks, 25 manager agents coordinating resource allocation, and 5 AI assistant agents providing strategic recommendations.

Each simulation run spans 300 discrete time steps. The experimental design examines three organizational paradigms (centralized, distributed, self-organized) with multiple AI configurations. The primary comparison evaluates all three paradigms under NO_AI conditions (450 runs: 150 replications × 3 paradigms). Additionally, for the self-organized paradigm, two AI-enhanced variants are tested: AI_SIMPLE (generic reinforcement learning support, 50 replications) and AI_CRISIS (crisis-aware adaptation, 150 replications). Centralized and distributed paradigms operate exclusively under NO_AI conditions, as AI assistants are architected for decentralized coordination contexts. The complete experimental design is summarized in

Table 1. Complete experimental design matrix.

Paradigm	AI Configuration	Replications	Total Runs	Notes
Centralized	NO_AI	150	150	Baseline hierarchical control
Distributed	NO_AI	150	150	Baseline modular coordination
Self-organized	NO_AI	150	150	Baseline autonomous coordination
	AI_SIMPLE 1	50	50	Generic AI support (no crisis mode)
	AI_CRISIS	150	150	Crisis-aware AI adaptation
Total			650

¹ AI_SIMPLE (50 replications) was introduced as a control condition to isolate the effect of crisis-awareness from generic AI support. Centralized and distributed paradigms operate without AI in all conditions, as AI assistant agents are designed for decentralized task selection contexts. The primary comparison (NO_AI vs. AI_CRISIS) uses 150 replications per condition, while AI_SIMPLE serves as an auxiliary benchmark.

.

Simulations were executed on a workstation with 32 CPU threads and 32 GB RAM, requiring several hours for completion of the full experimental corpus. A complete parameter specification table including all simulation parameters, their justifications, and literature sources is provided in Supplementary Materials (Table S1) to facilitate reproducibility.

The environment models a manufacturing organization processing a continuous stream of tasks. Each task j is characterized by:

• complexity T_j ∈ [3, 12],
• deadline d_j ∈ [5, 15] time steps,
• resource requirement R_j ∈ [2, 8] units.

Task arrivals follow a Poisson process with rate λ = 12 tasks per time step, calibrated to maintain average system utilization near 80% under normal conditions. This arrival rate reflects realistic manufacturing settings where demand fluctuates around production capacity [49]. Organizational resource capacity under normal conditions is fixed at 100 units per step and is fully replenished at each time step. Completing a task of complexity T_j yields a value V_j = 10 × T_j, so that more complex tasks contribute more strongly to output while maintaining linear scaling.

3.2. Agent Types and Behavioral Rules

3.2.1. Worker Agents

Worker agents represent human operators performing production tasks. Each worker i is initialized with three heterogeneous attributes drawn from uniform distributions:

• skill level S_i~U(0.2,0.7), representing technical proficiency,
• capacity C_i~U(3,10), representing maximum concurrent workload,
• cooperation propensity θ_i~U(0.4,0.9), representing willingness to form coalitions [50].

The skill range [0.2, 0.7] reflects workforce proficiency distributions observed in empirical studies of team-based manufacturing [51]. The upper bound of 0.7 does not imply the absence of highly skilled workers in real systems. Rather, it reflects a deliberate modeling choice to avoid introducing dominant “super-agents” whose individual capabilities could overshadow coordination effects. By capping individual skill levels below 1.0, the model ensures that productivity gains emerge from coalition formation and organizational structure rather than from a small number of exceptional agents. This choice also prevents trivial resolution of bottlenecks in the centralized paradigm through individual expertise alone. Capacity and cooperation ranges are similarly grounded in organizational behavior literature documenting concurrent task limits [52] and cooperation propensity distributions [53].

At any time t, worker i maintains its current workload w_i(t), the set of active tasks T_i(t), and a cooperation score θ_i(t) reflecting its collaboration reputation.

Worker decision-making follows a four-stage cycle at each time step:

• Perception: updating task progress and completion status;
• Decision: selecting tasks and potential partners via a paradigm-specific rule;
• Action: performing individual work or coalition work;
• Learning: updating cooperation propensity based on outcomes.

Task selection is guided by a multi-attribute utility function that balances skill match, value, feasibility and urgency:

$$U_{ij}=0.4\hspace{0.17em}{U}_{\mathrm{skill}}+0.3\hspace{0.17em}{U}_{\mathrm{value}}+0.2\hspace{0.17em}{U}_{\mathrm{feasibility}}+0.1\hspace{0.17em}{U}_{\mathrm{urgency}},$$

(1)

The weighting scheme in Equation (1) reflects the relative importance of task–skill alignment in production-oriented settings. Skill match is assigned the highest weight (0.4) because mismatches between worker capabilities and task complexity are a primary source of inefficiency, rework, and delays in manufacturing systems [54]. Workload-related value considerations follow (0.3), while feasibility (0.2) captures the additional coordination cost associated with coalition-based execution. Urgency (0.1) is weighted lower to prevent short-deadline tasks from systematically overriding skill suitability, which could otherwise degrade overall system performance.

The components are defined as follows:

• Skill match:

$$U_{\mathrm{skill}}=1-{\displaystyle \frac{\mid s_i\cdot 10-T_j\mid }{10}},$$

(2)

which rewards tasks whose complexity is close to the worker’s effective skill.

• Value:

$$U_{\mathrm{value}}={\displaystyle \frac{V_j}{100}},$$

(3)

normalizing task value.

• Feasibility:

$$U_{\mathrm{feasibility}}\in \left\{\mathrm{0.5,1.0}\right\},$$

(4)

taking a lower value when the task requires coalition, and a higher value when it can be executed individually.

• Urgency:

$$U_{\mathrm{urgency}}={\displaystyle \frac{1}{\mathrm{m}\mathrm{a}\mathrm{x}(d_j-t,1)}},$$

(5)

which increases as the task approaches its deadline.

Workers evaluate a subset of pending tasks (typically 5 tasks, reflecting bounded attention [19]) and select the one with the highest utility, optionally modified by AI recommendations when present.

Task progress accumulates according to worker skill and task complexity. For individual work, the progress per time step is given by:

$$\mathsf{\Delta }{p}_j(t)=\mathrm{m}\mathrm{i}\mathrm{n}({\displaystyle \frac{2s_i}{T_j}},\hspace{0.17em}0.8),$$

(6)

where factor 2 calibrates temporal scaling (a worker with S_i = 0.5 requires approximately 10 steps to complete a task with T_j =10), and the upper bound of 0.8 prevents unrealistic one-step completions by enforcing a minimum task duration.

When tasks are performed by coalitions, the progress rate is based on the average skill of coalition members S_coalition, with a penalty capturing coordination overhead [55]:

$$\mathsf{\Delta }{p}_j(t)=\mathrm{m}\mathrm{i}\mathrm{n}({\displaystyle \frac{0.7\cdot 2\cdot {\stackrel{-}{s}}_{\mathrm{coalition}}}{T_j}},\hspace{0.17em}0.8),$$

(7)

The factor 0.7 reflects that multi-person coalitions do not achieve the full sum of individual productivities due to synchronization and communication costs, consistent with Brooks’ Law [55] and meta-analyses of team productivity [56] reporting 65–75% efficiency across manufacturing, software, and service domains.

Cooperation propensity θ_i evolves through a simple reinforcement rule:

$${\theta }_i(t+1)={\theta }_i\left(t\right)+\alpha \left(r_i\right(t)-c_i(t\left)\right),$$

(8)

where α = 0.05 is a learning rate following Sutton and Barto’s [22] recommendation for slowly changing environments (α ∈ [0.01, 0.1]), r_i(t) is a reward term (e.g., positive when a coalition successfully completes its task), and C_i(t) is a cost term applied when the workload ratio W_i/C_i > 0.8, representing overcommitment. The cooperation variable is bounded in [0.1, 1.0] to avoid both complete defection and unconditional collaboration.

3.2.2. Manager Agents

Manager agents coordinate task allocation in the centralized and distributed paradigms. Each manager m is initialized with:

• a coordination efficiency η_m ~ U (0.6,0.95), influencing decision quality;
• a fixed set of subordinate workers S_m.

In the centralized paradigm, a single executive manager m₀ supervises all 70 workers. The executive retrieves the global task queue, ranks tasks by the efficiency ratio V_j/T_j, and assigns them to available high-skill workers. To model bounded rationality [19], the executive can process only a limited number of task assignments per step (15), which introduces a structural bottleneck under high task arrival rates. This limit reflects cognitive constraints documented in bounded rationality literature [19,57], where individual decision-makers can effectively evaluate 5–20 alternatives per decision cycle.

In the distributed paradigm, workers are partitioned into approximately 25 departments, each overseen by a local manager with a small subordinate set. Each manager optimizes allocation locally within their department. When local resources are insufficient, managers attempt to negotiate inter-department transfers. This modular structure reduces individual cognitive load but introduces coordination costs at departmental interfaces [33].

3.2.3. AI Assistant Agents

AI assistant agents implement a Q-learning algorithm [58] to learn coordination strategies from interaction with the environment. Each AI agent monitors a domain of approximately 14 workers and provides recommendations that workers may adopt with probability depending on their cooperation or trust level.

The AI maintains state–action value estimates Q(s,a) updated via temporal-difference learning:

$$Q(s,a)\leftarrow Q(s,a)+\alpha [r+\gamma \underset{a^{\prime }}{\mathrm{m}\mathrm{a}\mathrm{x}} Q(s^{\prime },a^{\prime })-Q(s,a)],$$

(9)

with learning rate α = 0.15, discount factor γ = 0.95, and ε-greedy exploration (initial ε = 0.2, decaying to ε = 0.05) [22].

The environment state s(t) is computed locally within each AI agent’s designated domain (approximately 14 workers), ensuring consistency with decentralized information structures:

• average workload, w, in the agent’s domain;
• workload standard deviation, σ_w within the domain;
• current resource availability, R_available (assumed publicly visible via centralized inventory systems or MES integration);
• number of pending tasks, |Q| (tasks assigned to or visible to domain workers);
• recent throughput, ∑V_completed summed over domain workers only.

This local state aggregation prevents AI agents from accessing global organizational information unavailable to decentralized human workers, maintaining architectural consistency with distributed decision-making paradigms. Only resource availability is treated as globally observable, reflecting realistic shop-floor visibility of material inventories.

Continuous states are discretized into approximately 1000 representative bins using a hashing and binning procedure, which balances representation detail and computational cost [59].

The action space consists of nine discrete actions built from three resource strategies (conservative, balanced, aggressive) combined with three coalition strategies (encourage large coalitions, encourage small coalitions, neutral), yielding 3 × 3 recommendation types. Each recommendation encodes how strongly workers should favor conservative versus aggressive task selection and whether they should attempt to form larger or smaller coalitions.

The AI operates in two main configurations:

• NO_AI: AI agents are not present; workers operate solely based on local rules;
• AI_SIMPLE: AI agents are active and use the same Q-learning architecture, but with a single, stationary reward function focused on improving throughput and workload balance, without an explicit crisis mode;
• AI_CRISIS: AI agents are active and use a crisis-aware reward function that changes when resource scarcity is detected.

To ensure reproducibility, the reward functions used by AI assistant agents are specified explicitly for both operational regimes.

Under normal operating conditions (R_available ≥ 80), the reward emphasizes efficiency and balanced workload distribution:

$$r_{normal}\left(t\right)=0.85∆V\left(t\right)+0.10{\Theta }_{equity}\left(t\right)-0.03P_{queue}\left(t\right)-0.02P_{overload}\left(t\right),$$

(10)

where

• $∆V\left(t\right)$ denotes the incremental throughput achieved within the agent’s domain at time step $t$;
• ${\Theta }_{equity}\left(t\right)$ captures workload equity, defined as:

$${\Theta }_{equity}\left(t\right)=1-CV\left(w\right),$$

(11)

where CV(w) is the coefficient of variation in worker workloads;

• $P_{queue}\left(t\right)$ penalizes excessive queue growth:

$$P_{queue}(t)={\displaystyle \frac{\left|Q\right|}{300}}$$

(12)

• $P_{overload}\left(t\right)$ penalizes systematic worker overcommitment:

$$P_{overload}(t)={\displaystyle \frac{1}{N}}\sum _{i=1}^{N}max\left(0,{\displaystyle \frac{w_i}{C_i}}-0.8\right)$$

(13)

Under crisis conditions (R_available < 80), the reward function shifts toward stability preservation and throughput retention:

$$r_{crisis}\left(t\right)=0.50∆V\left(t\right)+0.20{\Theta }_{equity}\left(t\right)+0.30B_{retention}\left(t\right)+0.20B_{stability}\left(t\right)+0.30B_{queue}\left(t\right)-0.05P_{overload}\left(t\right)$$

(14)

where

• $B_{retention}\left(t\right)$ rewards maintaining throughput above 50% of the pre-crisis baseline;
• $B_{stability}\left(t\right)$ penalizes short-term volatility, defined as:

$$B_{stability}(t)=1-{\displaystyle \frac{\left|∆V\left(t\right)-∆V(t-1)\right|}{V_{baseline}}}$$

(15)

• $B_{queue}\left(t\right)$ penalizes short-term volatility, defined as:

$$B_{queue}(t)=max\left(\mathrm{0,1}-{\displaystyle \frac{\left|Q\right|}{300}}\right)$$

(16)

Worker adoption of AI recommendations is probabilistic:

• in baseline periods, P_adopt = 0.5 + 0.3θ_i (approximately 50–80%);
• in crisis periods, P_adopt = 0.85 + 0.15θ_i (approximately 85–100%).

This formulation captures the empirically observed tendency for human operators to rely more strongly on decision support under conditions of high uncertainty [23]. AI recommendations become active only after an initial pre-training period (30 steps), during which Q-values are initialized without influencing behavior. The AI_SIMPLE configuration is used as an additional condition for the self-organized paradigm to separate the effect of generic AI support from that of crisis-aware adaptation (Section 4.5).

3.3. Decision-Making Paradigms

Three organizational decision paradigms are implemented within the same simulation framework: centralized, distributed, and self-organized.

In the centralized paradigm, the communication network is a star topology with the executive manager as the hub and all workers as spokes. All task allocation decisions pass through the executive, and workers behave passively, executing assigned tasks without autonomous selection.

In the distributed paradigm, an initial small-world network is partitioned into modular communities using Louvain community detection [60]. A large proportion of inter-community edges is reduced to create departmental boundaries, while a minority is retained to allow limited cross-department communication. The resulting network exhibits high modularity. Workers select tasks from departmental queues using the utility function described above, and managers coordinate intra-departmental resource allocation and inter-departmental negotiations.

In the self-organized paradigm, the network is initialized as a Watts–Strogatz small-world graph [45] with average degree k = 6 and rewiring probability p = 0.1, producing moderate clustering and short path lengths. The network then evolves dynamically through preferential attachment and link pruning: workers occasionally form new links with partners offering skill complementarity and available capacity, and rarely drop unused links to maintain communication efficiency [61].

Task selection is fully decentralized in the self-organized paradigm. Workers draw tasks directly from the global queue based on the utility function. When task complexity exceeds individual capability (T_j > 7S_i), workers may initiate coalition formation via a three-step process: querying neighbors with sufficient cooperation score and free capacity, ranking candidates by skill and cooperation, and recruiting up to two partners until the coalition’s combined skill ∑s_k exceeds a threshold proportional to T_j [44]. Coalition initiation probability and thresholds are modulated by AI recommendations in AI-enabled conditions.

3.4. Disruption Protocol

To study resilience under disruption, a supply-side shock is introduced at time t = 150, after the system has reached a steady-state operating regime. The disruption lasts for 20-time steps (t ∈ [150, 169]). During this interval, resource capacity is reduced by a random factor δ ~ U(0.4,0.6), such that:

$$R_{\mathrm{capacity}}=100\cdot (1-\delta ),$$

(17)

resulting in a 40–60% reduction in available resources, representing severe but realistic supply chain disruptions observed during COVID-19 and semiconductor shortages [4,5,62]. All other parameters (task arrival rate, complexity and deadlines) remain unchanged, isolating the effect of resource scarcity from demand-side shocks. The reduction factor δ is sampled once per simulation run at disruption onset and held constant throughout the entire disruption window, ensuring controlled experimental conditions. Normal resource availability resumes at t = 170, and recovery is monitored until t = 220. The 20-step disruption duration corresponds to an acute crisis phase of approximately 2–3 weeks in real-world manufacturing contexts [63].

3.5. Resilience Metrics

Conventional resilience measures typically compute proportional retention p = P_during/P_pre, combined with a recovery-speed term [14], where P_pre and P_during denote average performance before and during disruption, respectively, and T_recovery denotes time to recover. Such metrics systematically penalize high-performing systems, because larger absolute drops from a higher baseline yield lower retention values even if absolute crisis performance is higher [28].

To address this bias, an absolute resilience metric is defined that normalizes performance to system capacity and uses a fixed performance threshold for recovery [64]:

$$R_{\mathrm{absolute}}={\left({\displaystyle \frac{P_{\mathrm{during}}}{P_{\max }}}\right)}^{0.6}\times {\left({\displaystyle \frac{1}{1+T_{\mathrm{recovery}}/\tau }}\right)}^{0.4},$$

(18)

where

• P_during is the mean throughput (tasks completed per step) during the disruption window;
• P_max = 12 tasks/step is the maximum expected throughput (empirical upper bound across simulations);
• T_recovery is the number of steps required until performance reaches or exceeds a fixed threshold of 7.0 tasks/step for at least three consecutive steps;
• τ = 10 is a time-scaling constant.

The exponents (0.6 and 0.4) weight crisis throughput slightly more than recovery speed, reflecting the practical importance of maintaining functional capacity during disruptions. The 7.0 tasks/step threshold corresponds to approximately 60% of peak capacity and is interpreted as a minimum viable operating level; the three-step condition filters out transient spikes from genuine recovery.

For comparison, a relative resilience metric based on retention and relative recovery can also be computed, but the analysis in this paper focuses on the absolute metric.

3.6. Data Collection and Statistical Analysis

Simulation data are collected at each time step using Mesa’s built-in data collection mechanisms. At the model level, the following variables are recorded:

• instantaneous task completions per step;
• cumulative number of completed tasks;
• average workload ratio (workload divided by capacity);
• current resource availability;
• queue size (number of pending tasks);
• network density and clustering coefficient;
• number of coalitions formed per step.

At the agent level, worker-specific states (workload, cooperation propensity, and active tasks) as well as coalition composition (member identifiers, skill distribution, and coalition duration) are recorded. AI agents additionally log visited states, selected actions, and received rewards, enabling post hoc analysis of the learned coordination policies.

Statistical analyses were conducted using standard scientific libraries, including NumPy (v1.24.3), Pandas (v2.0.3), SciPy (v1.11.1), and Matplotlib (v3.7.2) [65]. Differences between decision paradigms and AI configurations are evaluated using independent-samples t-tests with Welch correction for unequal variances, one-way ANOVA for multi-group comparisons, and Tukey’s honest significant difference test for post hoc analysis. Effect sizes are reported using Cohen’s d. Statistical significance is assessed at a two-sided significance level of α = 0.05.

Normality of performance and resilience distributions was assessed via Shapiro–Wilk tests (α = 0.05) for sample sizes n < 50 and visual inspection of Q-Q plots for larger samples. Where violations were detected (e.g., resilience scores exhibit moderate right skewness), statistical inferences were confirmed using non-parametric alternatives: Mann–Whitney U tests for pairwise comparisons and Kruskal–Wallis H tests for multi-group comparisons. All reported statistical conclusions (p < 0.001, large effect sizes) remained unchanged under both parametric and non-parametric testing, confirming robustness to distributional assumptions. Welch’s correction for unequal variances further ensures validity when homoscedasticity assumptions are violated.

Following established ABM validation protocols [43], the model undergoes four validation tiers:

• Face validity: Agent behaviors (bounded rationality [19], skill-based coalition formation [44], small-world network dynamics [45]) operationalize validated organizational theories rather than ad hoc rules.
• Structural validation: Emergent patterns align with empirical observations. For example, the centralized paradigm’s throughput bottleneck replicates Simon’s [19] bounded rationality predictions, while self-organized adaptation matches Holland’s [34] complex adaptive systems models.
• Sensitivity validation: Section 5.3 demonstrates that qualitative findings (paradigm rankings, AI benefits) persist across parameter perturbations (±20% utility weights, 6.0–8.0 recovery thresholds, 40–80% disruption severity), confirming structural rather than parametric dependencies.
• Cross-model validation: Results align with theoretical predictions from organizational literature. Centralized bottlenecks replicate Simon’s [19] predictions; distributed coordination failures align with Thompson’s [32] interdependence theory; self-organized resilience matches Heylighen’s [21] findings on distributed problem-solving.

Direct empirical calibration against organizational disruption data faces three obstacles: (1) data unavailability—firms rarely share granular crisis-response data due to competitive sensitivity [9]; (2) confound control—real disruptions involve simultaneous demand shocks, technology failures, and strategic responses, preventing isolation of structural effects; (3) ethical constraints—experimental manipulation of organizational structures during crises is impractical. ABM’s value lies precisely in enabling controlled counterfactual analysis impossible in empirical settings [43,66]. Our approach follows precedents in computational organization theory [66], where stylized models reveal causal mechanisms obscured by empirical complexity.

4. Results

4.1. Baseline Performance Under Stable Conditions

Steady-state performance was evaluated over the interval t = 50–149, following the initial equilibration phase.

Table 2. Baseline performance metrics (time steps 50–149, mean values over replications ¹).

Paradigm	Tasks/Step	Queue Size	Avg. Workload 2 (Capacity Ratio)	Coalitions/ Steps
Centralized	4.55 ± 0.42	673.7 ± 48.2	0.89 ± 0.06	≈0
Distributed	6.55 ± 0.51	542.7 ± 52.3	0.51 ± 0.09	≈0
Self-organized	9.28 ± 0.53	445.3 ± 38.7	0.65 ± 0.08	5.53 ± 1.21

¹ Based on 150 replications per paradigm (n = 450 total runs). All differences between paradigms are statistically significant (p < 0.001, Welch t-tests). Effect sizes: d ≈ 5.1 (centralized vs. distributed), d ≈ 6.8 (distributed vs. self-organized), d ≈ 12.4 (centralized vs. self-organized); ² Avg. Workload (capacity ratio) represents the mean ratio between each worker’s current workload and maximum capacity, averaged across all workers. Values are normalized to the interval [0, 1], where 1.0 indicates full capacity utilization.

summarizes the main indicators for the three decision paradigms: centralized, distributed, and self-organized (without AI).

The results indicate a clear and consistent advantage of the self-organized structure. The self-organized paradigm achieves a mean throughput of 9.28 tasks/step, compared to 6.55 tasks/step in the distributed structure and 4.55 tasks/step in the centralized structure. These differences are statistically significant at the replication level (Welch-corrected t-tests, p < 0.001 for all pairwise comparisons) and associated with very large effect sizes, with typical Cohen’s d values of d ≈ 5.1 (centralized vs. distributed), d ≈ 6.8 (distributed vs. self-organized), and d ≈ 12.4 (centralized vs. self-organized).

The centralized paradigm combines high human resource utilization with chronic backlog accumulation. Tasks accumulate in the queue (≈674 tasks on average), while the single managerial agent becomes a critical decision bottleneck. The number of feasible allocations per time step is limited, and workers remain idle while waiting for instructions, even though the aggregate system capacity would allow substantially higher throughput. Coalition formation is practically absent.

The distributed structure partially reduces this bottleneck. Throughput increases to 6.55 tasks/step and queue size decreases to ≈543 tasks, but departmental boundaries generate persistent coordination fragmentation between organizational units. Workers and managers optimize primarily at the local level, while inter-departmental collaboration remains rare and costly. The average workload level (≈0.51) indicates the existence of substantial unused capacity.

The self-organized structure is distinguished by two key mechanisms:

• autonomous task selection, which enables improved matching between individual skills and task requirements, reducing backlog (≈445 tasks);
• intensive coalition formation (≈5.53 coalitions/step), which enables the execution of tasks exceeding individual capacity.

The temporal evolution of organizational performance under stable conditions is illustrated in Figure 1. Self-organized systems rapidly converge toward a high and stable throughput level, whereas centralized systems remain trapped in a low-performance plateau, and distributed structures exhibit an intermediate and moderately fluctuating regime.

4.2. Baseline Effect of AI in Self-Organized Systems

The effect of AI integration within the self-organized paradigm is assessed by comparing two configurations: NO_AI and AI_CRISIS (with the crisis mode inactive during the baseline phase). For the centralized and distributed paradigms, the AI module remains inactive within the current experimental design, and baseline behavior is therefore identical to the NO_AI condition.

Table 3. Baseline metrics for self-organized systems with and without AI ¹.

Version	Tasks/Step	Queue Size	Avg. Workload	Coalitions/ Steps
Self-Organized, NO_AI	9.28 ± 0.53	445.3 ± 38.7	0.65 ± 0.08	5.53 ± 1.21
Self-Organized, AI_CRISIS	11.29 ± 0.48	378.0 ± 35.2	0.87 ± 0.07	8.11 ± 1.45

¹ Based on 150 replications per configuration (n = 300 runs). Differences tested via independent t-test with Welch correction. Effect sizes: throughput (d ≈ 4.5), coalitions (d ≈ 3.7).

presents the main indicators for the two self-organized configurations over the interval t = 50–149.

AI integration produces a throughput increase from 9.28 to 11.29 tasks/step (≈+21%), accompanied by a queue reduction (from ≈445 to ≈378 tasks) and a pronounced intensification of collaboration (from 5.53 to 8.11 coalitions/step). Average worker workload increases to ≈0.87, indicating more efficient utilization of available capacity without systematic overloading.

Performance differences between the two self-organized variants are statistically significant (p < 0.001) and associated with very large effect sizes (Cohen’s d ≈ 4.5 for throughput and 3.7 for coalition rate). AI recommendations guide workers toward more suitable tasks and partners while preserving local autonomy; the aggregate effect is a substantial amplification of emergent self-organization capacity.

The disaggregated effects of AI on organizational performance are illustrated in Figure 2.

Figure 2A shows that AI increases throughput exclusively in the self-organized paradigm (+21.6%), while no measurable effect is observed in the centralized and distributed structures, where AI does not participate in task allocation. Figure 2B reveals a substantial AI-induced reduction in queue size only for the self-organized configuration, indicating improved task–capacity matching and backlog control. Figure 2C confirms that the resilience improvement associated with AI (+10.7%) is again confined to the self-organized regime, demonstrating that AI effectiveness is structurally contingent. Finally, Figure 2D displays the temporal performance dynamics of the self-organized paradigm, showing that AI not only improves steady-state throughput but also mitigates the depth and duration of the performance drop during the disruption window.

4.3. Crisis Response Under Resource Disruption

The impact of the supply chain disruption is analyzed over the interval t = 150–169, during which resource availability is reduced by 40–60%, while task arrival rate and complexity distributions remain unchanged.

Table 4. Performance during disruption (time steps 150–169) ¹.

Paradigm, Version	Baseline Tasks/Step	Crisis Tasks/Step	Retention (≈Pduring/Ppre)
Centralized, NO_AI	4.38 ± 0.41	3.07 ± 0.38	≈0.82 ± 0.06
Distributed, NO_AI	6.56 ± 0.50	4.84 ± 0.45	≈0.74 ± 0.05
Self-Organized, NO_AI	9.24 ± 0.52	6.65 ± 0.47	≈0.72 ± 0.06
Self-Organized, AI_CRISIS	11.28 ± 0.49	7.50 ± 0.42	≈0.67 ± 0.05

¹ Based on 150 replications for NO_AI conditions per paradigm (n = 450) and 150 for AI_CRISIS (n = 600 total).

summarizes crisis performance for each paradigm under NO_AI conditions, with the AI_CRISIS configuration included for the self-organized case.

All paradigms experience a performance drop at disruption onset, but absolute throughput levels differ substantially. Self-organized systems without AI maintain ≈6.65 tasks/step, while distributed systems drop to ≈4.84, and centralized systems to ≈3.07 tasks/step. The AI_CRISIS configuration maintains the highest crisis throughput (≈7.50 tasks/step), representing an approximately 12–13% advantage over the non-AI self-organized version.

Relative retention values reveal an important phenomenon. The self-organized AI_CRISIS configuration exhibits the lowest proportional retention (≈0.67) because it starts from a higher baseline. Nevertheless, it delivers the largest absolute volume of work during the crisis. Centralized and distributed paradigms appear more “resilient” in proportional terms, yet operate at substantially lower absolute throughput. This divergence between relative retention and absolute performance motivates the adoption of an absolute resilience metric.

Queue size increases in all cases during disruption. Centralized systems reach the largest backlog, reflecting the combination of constrained resources with a single overloaded decision node. Self-organized paradigms exhibit a controlled backlog increase, while AI_CRISIS contributes to backlog limitation through conservative recommendations and increased coalition activity.

4.4. Recovery Trajectories After Disruption

Recovery behavior is analyzed for t ≥ 170. Under the traditional approach, recovery is defined relative to each system’s own pre-crisis performance, with resilience expressed as the product of retention and recovery speed. Aggregate values for the NO_AI configurations are reported in

Table 5. Traditional resilience-related metrics (NO_AI configuration) ¹.

Paradigm	Pre-Performance (Tasks/Step)	During Performance	Retention	Recovery Time (Steps)
Centralized	4.38 ± 0.41	3.07 ± 0.38	≈0.82 ± 0.06	≈32 ± 8.2
Distributed	6.56 ± 0.50	4.84 ± 0.45	≈0.74 ± 0.05	≈15 ± 4.5
Self-organized	9.24 ± 0.52	6.65 ± 0.47	≈0.72 ± 0.06	≈14 ± 3.8

¹ Based on 150 replications per paradigm (n = 450 total). Recovery time defined as steps to reach 85% of pre-disruption baseline for 3 consecutive steps.

.

The centralized paradigm shows relatively high retention but a very long recovery time (on average more than 30 simulation steps). Distributed and self-organized paradigms recover more rapidly, but their differences appear modest when viewed strictly through relative indicators.

This relative perspective masks a crucial fact: in absolute terms, even after recovery, centralized structures operate at substantially lower throughput than self-organized systems. Consequently, a resilience metric based exclusively on proportional retention and recovery relative to each system’s baseline may yield misleading conclusions when systems start from markedly different performance levels.

4.5. Absolute Resilience and AI Impact in Self-Organized Systems

To correct baseline-related bias, an absolute resilience metric is applied, normalizing crisis throughput to the maximum system capacity (12 tasks/step) and measuring recovery relative to a fixed operational threshold (7.0 tasks/step). Results for the three self-organized variants (NO_AI, AI_SIMPLE, and AI_CRISIS) are presented in

Table 6. Absolute resilience metrics for self-organized systems (mean values) ¹.

Version	Baseline Tasks/Step	Crisis Task/Step	Normalized Crisis (Pduring/12)	Absolute Recovery Time	Rabsolute	Cohen’s d vs. NO_AI	95% CI
NO_AI	9.24 ± 0.52	6.65 ± 0.47	0.55 ± 0.04	≈1.0 ± 1.5	0.677 ± 0.039	—	[0.667, 0.687]
AI_SIMPLE	11.24 ± 0.43	7.57 ± 0.37	0.63 ± 0.03	≈0.5 ± 1.4	0.746 ± 0.040	1.78 (large)	[0.735, 0.757]
AI_CRISIS	11.28 ± 0.49	7.50 ± 0.42	0.63 ± 0.04	≈0.2 ± 0.8	0.749 ± 0.028	2.13 (very large)	[0.742, 0.756]

¹ Based on NO_AI (n = 150), AI_SIMPLE (n = 50), and AI_CRISIS (n = 150) replications for self-organized paradigm (n = 350 total runs). Recovery time defined as steps to reach 7.0 tasks/step threshold for 3 consecutive steps. Differences between NO_AI and AI_CRISIS: p < 0.001 (Welch t-test and Mann–Whitney U test). 95% CI = 95% confidence interval.

.

Both AI-based variants significantly outperform the NO_AI configuration in absolute resilience. Mean values increase from approximately 0.68 (NO_AI) to ≈0.75 (AI_SIMPLE and AI_CRISIS). The difference between NO_AI and AI_CRISIS is highly statistically significant (t-tests, p < 0.001) with a large effect size (Cohen’s d ≈ 2.13). Statistical significance was confirmed using both parametric (Welch t-test) and non-parametric (Mann–Whitney U) tests due to moderate right skewness in resilience distributions, with identical conclusions under both approaches.

These effects are decomposed graphically in Figure 3. Figure 3A displays the absolute resilience scores, confirming the systematic improvement from NO_AI to AI-based configurations. Figure 3B reports crisis throughput, showing that both AI variants sustain higher absolute performance during disruption than the non-AI case. Figure 3C highlights the sharp reduction in recovery time under AI assistance, with AI_CRISIS achieving near-instantaneous recovery to the operational threshold. Finally, Figure 3D shows the full distribution of absolute resilience across replications, indicating both a higher central tendency and reduced variance under AI support. AI_SIMPLE primarily maximizes throughput and achieves the highest crisis performance level, whereas AI_CRISIS achieves a particularly favorable trade-off between throughput and recovery speed, resulting in the highest absolute resilience score.

Relative to the centralized and distributed paradigms, absolute resilience values computed using the same formulation are substantially lower. Using the same absolute metric across all paradigms yields: Centralized (R_absolute = 0.266 ± 0.042), Distributed (R_absolute = 0.498 ± 0.051), Self-Organized NO_AI (R_absolute = 0.677 ± 0.039), Self-Organized AI_CRISIS (R_absolute = 0.749 ± 0.028). The resulting hierarchy is unambiguous: centralized < distributed < self-organized (NO_AI) < self-organized (AI).

4.6. Coalition Dynamics and Network-Level Adaptation

Coalition and network indicators provide additional insight into the mechanisms generating high resilience in the self-organized paradigm. Under the NO_AI configuration, self-organized systems form on average ≈5.5 coalitions per step during the baseline phase, with a moderate increase during the crisis as workers attempt to combine capacities under resource scarcity. Under AI_CRISIS, coalition formation exceeds 8 coalitions per step during baseline and remains elevated during and after the disruption.

Network structure exhibits clear paradigm-dependent differences. At a representative baseline time (e.g., t = 50), the centralized network shows very low density (≈0.02) and near-zero clustering, consistent with a star topology. The distributed network displays moderate density (≈0.18) and high clustering (≈0.65), reflecting departmental segmentation. The self-organized network exhibits intermediate density (≈0.12) and moderate clustering (≈0.42) characteristics of small-world topologies and dynamically reconfigures as collaboration partners change.

During the crisis, self-organized network density increases temporarily by approximately 18% (p < 0.001, paired t-test comparing t = 150–169 vs. t = 100–149), indicating intensified collaboration, and subsequently returns toward baseline values after recovery. In the centralized and distributed paradigms, network structure remains largely static (density changes < 3%, not significant), severely limiting topological adaptation to shocks.

The combined micro- and meso-level observations support a coherent interpretation of the macro-level resilience results. Frequent coalition formation and adaptive networking enable rapid redistribution of effort and flexible exploitation of heterogeneous competencies. Crisis-aware AI stabilizes these emergent processes through context-sensitive recommendations while preserving distributed autonomy and avoiding the introduction of a new centralized control layer.

5. Discussion

The simulation results clarify how organizational structure and AI assistance jointly shape resilience in an HCP production system under supply chain disruption. Centralized, distributed, and self-organized paradigms exhibit distinct performance profiles, and the integration of a crisis-aware AI assistant modifies the behavior of self-organized systems in a systematic way.

5.1. Theoretical Implications

5.1.1. Organizational Structure and Resilience Mechanisms

The first implication concerns the hierarchy of decision paradigms. Across baseline, crisis and recovery phases, self-organized systems outperform centralized and distributed structures in both throughput and recovery speed. Centralized systems maintain relatively high proportional retention but operate at low absolute capacity and recover slowly because a single managerial bottleneck constrains allocation under resource availability [19]. Distributed systems partially reduce managerial bottlenecks but continue to face substantial coordination frictions at departmental interfaces [20,32].

Self-organized systems distribute decision-making across all workers, allowing parallel task–skill matching and flexible coalition formation for complex tasks [21,34]. Network adaptation reinforces these dynamics: small-world communication structures [45] and temporary increases in connectivity during disruption support rapid partner discovery and reallocation of effort. These findings support complexity-theoretic arguments [67] that resilience is derived less in a fixed hierarchy and more in adaptive, emergent coordination patterns.

The mechanisms underlying this structural advantage can be traced to specific micro-level behaviors. Self-organized systems exhibit fundamentally different coordination patterns compared to hierarchical alternatives. Rather than funneling all decisions through centralized gatekeepers (which creates processing bottlenecks under load) or confining coordination within departmental silos (which inhibits resource reallocation when local units face heterogeneous shocks), self-organized architectures enable parallel, skill-based task matching across the entire workforce. This eliminates single points of failure and allows rapid exploitation of distributed competencies. The spontaneous coalition formation mechanism further extends effective capacity: when individual workers encounter tasks exceeding their capabilities, they autonomously recruit complementary partners from their local network neighborhood, achieving collective capability that no individual possesses. This emergent problem-solving occurs without managerial approval delays or cross-departmental negotiation overhead, explaining the substantial throughput advantage observed even under stable conditions.

5.1.2. AI as Metacognitive Enhancement vs. Centralized Control

A second implication relates to the role of AI. In the present setting, AI functions as a metacognitive layer rather than a central controller [68]. Under stable conditions, AI recommendations increase throughput by improving task assignment and intensifying coalition formation. During disruption, a crisis-aware reward function shifts AI behavior toward conservative resource use, larger coalitions and queue control. As a result, AI-enhanced self-organized systems maintain the highest absolute crisis throughput and exhibit nearly instantaneous recovery.

However, the magnitude of the AI effect (+10.7% absolute resilience) remains smaller than the structural effect of self-organization (+151% throughput improvement vs. centralized systems). This suggests that organizational structure is the primary determinant of resilience; AI provides an additional improvement when embedded in a suitable design but does not compensate for the fragility of centralized hierarchies. This reinforces a socio-technical view [69] in which technology and structure must be aligned.

The distinction between crisis-aware (AI_CRISIS) and generic (AI_SIMPLE) AI architectures further illuminates this dynamic. While both configurations improve baseline performance (+21.6%), only crisis-aware adaptation achieves superior absolute resilience (+10.7% vs. +10.2% for AI_SIMPLE). This differential effect arises from the reward function shift: during resource scarcity, AI_CRISIS explicitly prioritizes throughput retention and volatility reduction over short-term efficiency maximization, stabilizing coalition formation rates and preventing the backlog explosions that degrade recovery trajectories. In contrast, AI_SIMPLE continues optimizing for peak throughput, inadvertently encouraging resource-intensive task selections that exacerbate scarcity constraints. This demonstrates that crisis-aware AI must not simply maintain static policies learned during stable periods, but actively reconfigure objectives when environmental conditions change.

5.1.3. Resilience Measurement Bias and Metric Selection

A third implication concerns resilience measurement. Traditional retention-based metrics favor centralized systems because proportional performance loss is modest, even though these systems deliver the lowest absolute output during disruption [14,29]. Self-organized systems with AI appear weaker by retention alone because they start from a higher baseline, yet they maintain substantially higher absolute throughput and recover faster. An absolute resilience metric that normalizes crisis performance to a fixed capacity and recovery to a fixed operational threshold [15,64] resolves this contradiction and yields rankings consistent with practical viability.

The practical consequences of this measurement bias are non-trivial. Supplier selection processes that rely exclusively on proportional resilience metrics risk systematically favoring low-performing but “proportionally stable” organizations over high-performing systems that maintain greater absolute capacity during disruptions. Investment decisions based on relative retention may misallocate resources toward buffering low-capacity systems rather than enhancing the adaptive mechanisms of already-capable organizations. Performance incentive structures tied to proportional recovery can perversely encourage suppression of baseline performance to appear more resilient, generating moral hazard incompatible with operational efficiency objectives. The absolute resilience formulation eliminates these distortions by evaluating systems against external operational viability thresholds rather than idiosyncratic baselines.

5.1.4. Emergent Coordination Mechanisms

Finally, micro-level analysis identifies coalition dynamics and network adaptation as key mechanisms. Higher coalition rates, longer coalition lifetimes and greater skill complementarity are associated with better performance and resilience [44]. Adaptive network rewiring, particularly in self-organized systems, enables temporary increases in connectivity (+18% density during crisis, p < 0.001) without losing small-world structure [61]. Together, these mechanisms show how macro-resilience emerges from local interaction rules and evolving relational patterns.

5.2. Practical Implications for Industry 5.0

For incumbent organizations with established hierarchical structures, abrupt paradigm shifts risk operational disruption. A phased transition strategy mitigates risk while capturing resilience benefits (

Table 7. Phased implementation roadmap for transitioning from hierarchical to self-organized structures. Timeline, scope, and key interventions are specified for each phase, along with monitoring metrics and governance protocols. Organizations can adapt phase durations and scope based on size, industry constraints, and risk tolerance.

Phase	Timeline	Scope	Key Interventions	Monitoring & Governance
1. Pilot deployment	3–6 months	Single production cell (10–20 workers)	Transparent task visibility (digital dashboards) Bounded autonomy (5–10 task choices) Skill-matching platform for coalitions	Metrics: throughput, queue, coalitions, satisfaction Risk mitigation: managerial override available
2. Horizontal scaling	6–12 months	3–5 departments (50–100 workers)	Cross-departmental task visibility Reputation systems for cooperation tracking Crisis detection thresholds	Success criteria: >15% throughput gain Challenge: resolve cross-unit conflicts
3. Crisis resilience integration	12–18 months	Organization-wide (100+ workers)	Crisis-aware AI advisory activation Reward function calibration Worker training on AI interpretation	Infrastructure: digital twin, MES integration, wearables Governance: manual override protocols, quarterly audits
4. Continuous adaptation	18+ months	Institutionalization	Dynamic network analysis Periodic utility recalibration External disruption benchmarking	Sustainability: internal facilitators, autonomous maintenance, lessons learned documentation

).

In practice, the transition from efficiency to stability mode requires real-time monitoring of resource availability and backlog indicators. Organizations can implement threshold-based triggers (e.g., resource utilization > 80%, queue size > 2× normal) integrated into Manufacturing Execution Systems (MES) or digital twin platforms. When thresholds are breached, the AI system automatically shifts reward weights and broadcasts updated recommendations to workers via shop-floor displays or wearable devices. Human supervisors retain override authority to manually activate crisis mode based on qualitative signals (supplier communications, market volatility) not captured by quantitative metrics.

Industry-specific adaptations:

• Automotive assembly: Emphasize station-level autonomy within takt-time constraints; coalition formation for quality escalations
• Semiconductor fabrication: Balance autonomy with cleanroom protocols; crisis AI focused on yield preservation during material shortages
• Pharmaceutical manufacturing: Regulatory compliance boundaries within which self-organization operates; crisis mode for batch prioritization under API scarcity
• Custom machinery: Project-based coalition formation; AI assists in skill-task matching for one-off orders

AI appears most useful when implemented as an advisory system that provides context-sensitive recommendations rather than imposing decisions [11,23]. Crisis-aware AI that explicitly changes objectives when disruptions are detected, prioritizing stability and throughput preservation over short-term efficiency [41], can enhance the resilience of already flexible, self-organized organizations.

The measurement results indicate that organizations should monitor absolute crisis throughput and recovery to fixed thresholds, complemented by micro-indicators such as coalition rates, workload balance and network connectivity [70]. These metrics provide a more realistic picture of operational viability than retention alone and can serve as early warning and diagnostic tools.

5.3. Limitations and Future Research

The model adopts a stylized task environment, a fixed organization size and a single disruption pattern, and assumes high information visibility typical of mature digital twin infrastructures [10]. These simplifications support clear comparison across paradigms but limit their applicability to all industrial settings, particularly continuous process industries or highly regulated domains.

The disruption model focuses exclusively on supply-side capacity shocks (resource reduction) rather than demand-side shocks (task arrival rate increases) or hybrid scenarios. This choice reflects the supply chain disruptions observed during COVID-19 and semiconductor shortages [4,5], where material availability was the primary constraint. However, alternative disruption types—such as sudden demand surges, delivery delays, or cascading failures—may induce different organizational responses. For instance, demand-side shocks might favor centralized prioritization over distributed task selection, potentially altering the relative advantages of self-organized structures. Future research should systematically compare organizational resilience across multiple disruption typologies to establish boundary conditions for the reported findings.

The AI assistant relies on tabular RL with hand-crafted state aggregation and reward shaping [58,59], which may not scale to very large or highly heterogeneous organizations. The rapid emergence of Large Language Models (LLMs) and generative AI in manufacturing contexts suggests promising extensions. Hybrid architectures combining LLM-based semantic reasoning (e.g., interpreting unstructured supplier communications, detecting subtle crisis signals from qualitative data) with RL-driven tactical coordination (numerical state-action optimization) could enhance both crisis detection accuracy and recommendation acceptance rates. However, such integration introduces computational overhead, interpretability challenges, and trust calibration complexities that require careful investigation.

Sensitivity Analysis and Robustness

To assess the robustness of findings to parameter choices, several sensitivity analyses were conducted:

• Skill distribution: Worker skill range [0.2, 0.7] was extended to [0.2, 0.9] to include “master” level workers (10% of workforce with S_i > 0.8). Centralized throughput increased modestly (4.55 → 5.12 tasks/step, +12.5%) as executives could leverage high-skill workers, but self-organized systems maintained superiority (9.28 → 9.61 tasks/step, +3.6%), confirming that structural bottlenecks rather than skill ceilings drive centralized underperformance. ANOVA comparing paradigms remained highly significant (p < 0.001) under both skill distributions.
• Utility function weights: Alternative weight configurations were tested: urgency-prioritizing (w = 0.1/0.3/0.2/0.4), equal weights (w = 0.25 each), and feasibility-heavy (w = 0.2/0.3/0.4/0.1). Self-organized systems maintained superior throughput across all configurations (range: 8.9–9.7 tasks/step), though urgency-prioritizing workers induced +18% workload volatility (standard deviation increased from 0.08 to 0.15). Paradigm rankings remained unchanged (p < 0.001), with Spearman rank correlation ρ > 0.96 across all weight schemes.
• Centralized manager capacity: Executive processing limits were varied from 10 to 50 tasks/step. Centralized throughput scaled sublinearly: 10 tasks/step → 4.2 tasks/step, 15 → 4.55, 30 → 6.2, 50 → 7.8. Performance asymptoted near distributed levels (6.55 tasks/step) at 50 tasks/step but remained below self-organized (9.28 tasks/step), confirming that centralized underperformance stems from concentration of authority (structural bottleneck) rather than specific numerical processing limits. This finding validates the 15-task limit as representative of bounded rationality constraints without artificially handicapping centralized structures.
• Resilience metric exponents: Reversing throughput-recovery weights to (0.4 throughput, 0.6 recovery) or equalizing (0.5 each) altered absolute resilience scores by <8% but preserved rankings (self-organized > distributed > centralized, correlation ρ > 0.94) and AI superiority (AI_CRISIS > NO_AI, p < 0.001 under all formulations). This robustness indicates that the qualitative advantage of self-organized structures and crisis-aware AI does not depend on specific exponent calibration.
• Disruption severity: Varying resource reduction from 20% to 80% capacity loss produced proportional throughput drops but identical paradigm rankings. At 20% reduction, self-organized systems retained 85% throughput vs. 78% for distributed and 88% for centralized; at 80% reduction, retention values were 45%, 38%, and 52%, respectively. The absolute resilience hierarchy remained unchanged across all severity levels (p < 0.001, ANOVA). Coalition formation rates increased monotonically with disruption severity in self-organized systems (from 5.5 at 20% to 12.3 at 80%), demonstrating adaptive intensification of collaboration under stress.
• Organizational scale: Simulations with 50 agents (35 workers, 12 managers, 3 AI) and 200 agents (140 workers, 50 managers, 10 AI) produced qualitatively identical findings. Throughput scaled approximately linearly with workforce size (self-organized: 4.6, 9.3, 18.5 tasks/step for N = 50, 100, 200), while paradigm performance ratios remained stable (self-organized/centralized ≈ 2.0–2.1 across scales). This suggests that results generalize across realistic manufacturing cell sizes (10–250 workers) commonly observed in Industry 4.0/5.0 settings.

These sensitivity analyses demonstrate that the reported findings reflect structural and behavioral regularities rather than fine-tuned parameter calibration. Paradigm rankings, AI benefits, and absolute resilience advantages persist across reasonable parameter perturbations, supporting the external validity of conclusions.

Future research could extend the model with heterogeneous task types (e.g., divisible vs. indivisible, sequential vs. parallel dependencies), multiple interacting disruptions (compound shocks), empirically calibrated parameters from specific industries, and alternative AI architectures beyond tabular Q-learning [17,64,71]. Multi-organizational settings such as supply networks [72] also represent a natural extension, allowing investigation of how organizational-level resilience propagates across interconnected systems and whether network-level interventions (e.g., supplier diversification, inventory pooling) interact with intra-organizational decision paradigms.

6. Conclusions

This study developed and analyzed an ABM of organizational resilience in an HCP production system, comparing centralized, distributed and self-organized decision paradigms under supply chain disruption, with and without a crisis-aware AI assistant in the self-organized case.

The results show that self-organized structures are consistently superior to centralized and distributed configurations in terms of baseline throughput, crisis performance and recovery speed. Centralized systems suffer from managerial bottlenecks and structural rigidity, while distributed systems remain limited by locally confined decision-making and weak cross-unit coordination. Self-organized systems, by contrast, leverage autonomous task selection, coalition formation and adaptive network structures to reallocate effort and maintain capacity under stress.

The integration of a crisis-aware AI advisor further improves the resilience of self-organized systems. By shifting its objective from efficiency to robustness when resource scarcity is detected, the AI assistant encourages conservative resource strategies and larger coalitions, enabling higher absolute throughput during disruption and near-instantaneous recovery after resource restoration. At the same time, the results suggest that AI is most effective when it augments an already resilient structural design rather than attempting to compensate for shock-sensitive hierarchical structures.

A key methodological contribution is the use of an absolute resilience metric that normalizes crisis performance to a fixed capacity and recovery to an operational threshold. This metric avoids the high-performer penalty inherent in purely relative measures and yields resilience rankings that reflect actual productive capacity maintained during and after disruption.

Overall, the findings support the view that resilient Industry 5.0 organizations are likely to combine self-organized collective dynamics, adaptive communication networks and context-aware AI advisory layers, evaluated using absolute, capacity-oriented resilience measures. Such socio-technical configurations are better positioned to cope with the volatility, uncertainty and complexity that characterize modern HCPS.