Architecting the human-centric factory: analyzing the relationship between cognitive control units and operational antifragility in Industry 5.0

Manufacturing is moving beyond Industry 4.0's efficiency-oriented logic toward the human-centred, adaptive, and collaborative orientation of Industry 5.0. Human cognition, creativity, judgement, and collaborative intelligence are treated as strategic complements to intelligent technologies (Sheikh et al., 2024; Ghobakhloo et al., 2024). Competitiveness therefore depends not only on cost, speed, automation, and digital integration, but also on the capacity to learn and improve amid supply-chain disruption, geopolitical uncertainty, labour shortages, and demand volatility. This capacity reflects operational antifragility (OA): improvement through volatility rather than mere resistance or recovery (Taleb, 2012; Lou et al., 2024; Camarinha-Matos et al., 2024).

Despite extensive research on artificial intelligence, cyber-physical systems, robotics, digital twins, and automation, much of the Industry 4.0 literature remains technology-centred and implicitly equates greater automation with better performance (Villar et al., 2023; Keshmiry and Hassani, 2025; Qadeer et al., 2025; Hussain et al., 2025). Three connected gaps remain. First, resilience and adaptive capability are often conflated with OA, although resilience emphasizes recovery whereas antifragility denotes improvement because of volatility (Taleb, 2012). Second, human cognition is rarely modelled as the mechanism linking intelligent technologies and relational trust to OA, even though digitalization can intensify information overload, attention fragmentation, alert fatigue, and decision pressure (Saniuk et al., 2024). Third, studies seldom examine the governance arrangements that determine whether cognitive benefits become coordinated, system-level responses (Rožanec et al., 2023; Camarinha-Matos et al., 2024).

These gaps motivate three research questions.

Theoretically, the study integrates CLT and STS by positioning CLR as the cognitive mechanism connecting CCUs and HMMT with OA, and APPs as a sociotechnical governance mechanism that both supports OA and conditions the translation of cognitive efficiency into antifragile performance. Practically, it shows managers and system designers that antifragility requires more than intelligent systems or greater machine autonomy: it also requires cognitively manageable interfaces, calibrated trust, explicit decision rights, and clear escalation procedures. The following sections develop the hypotheses, describe the methods, report the results, and discuss their implications and limitations.

This study integrates CLT and STS to explain how human-centred design mechanisms relate to OA in I5MEs (Bucci et al., 2024; Saniuk et al., 2024). CLT holds that working memory is limited; performance deteriorates when task demands exceed that capacity (Sweller, 1988). This constraint is salient in advanced manufacturing, where operators monitor autonomous systems, interpret algorithmic outputs, and manage exceptions (Ghobakhloo et al., 2023a). CCUs can reduce extraneous load by prioritizing relevant information, while calibrated HMMT can reduce unnecessary monitoring; poorly designed interfaces or miscalibrated trust may instead increase load or encourage automation complacency (Villar et al., 2023; Lou et al., 2024). STS complements CLT by treating organizations as interdependent social and technical subsystems that require joint optimization (Emery and Trist, 1965; Cherns, 1976; Carayon et al., 2014). APPs operationalize this alignment by specifying human and machine decision rights, responsibilities, and escalation rules under uncertainty (Fernández-Miguel et al., 2025a).

OA must be distinguished from resilience and adaptive capability. Resilience concerns absorbing disruption and restoring acceptable functioning (Camarinha-Matos et al., 2024), whereas adaptive capability concerns sensing change and reconfiguring resources over time (Rožanec et al., 2023). An antifragile system improves because of disorder rather than merely surviving or adjusting to it (Taleb, 2012). CCUs and APPs are likewise narrower than general human-centred automation or ergonomics. CCUs coordinate information and exceptions at the human-machine interface; APPs govern decision authority and exception handling (Cherns, 1976). This distinction exposes a theoretical gap in technology-first accounts that treat resilience or adaptability as automatic consequences of technological sophistication (Villar et al., 2023). CLT suggests that technological and relational resources affect performance through cognitive demands, while STS suggests that these benefits depend on structural alignment. Evidence from other sociotechnical settings similarly shows that technology alone does not ensure adaptive performance; alignment among work structures, human cognition, and technical processes is decisive (Carayon et al., 2014). Combining CLT and STS therefore provides a mediation-moderation account of OA in which cognitive mechanisms explain how inputs operate and governance structures specify when those mechanisms yield system-level gains.

From an STS perspective, OA emerges when technical components support adaptive capacity under volatility. CCUs integrate decision support, real-time exception handling, and intelligent feedback to structure information flows and facilitate rapid sensemaking between people and autonomous systems (Villar et al., 2023). Accordingly, CCUs should help manufacturers reconfigure operations and learn from disruption.

Cognitively integrated manufacturing architectures can strengthen learning and adaptability after disturbances (Lou et al., 2024), while intelligent interfaces can accelerate reconfiguration and reduce coordination losses when conditions change (Sheikh et al., 2024). Human-in-the-loop arrangements may also outperform exclusively automated or human-controlled systems during recovery (Rožanec et al., 2023). Nevertheless, added coordination layers can increase cognitive load when they are poorly matched to human processing limits (Saniuk et al., 2024). The net association between CCUs and OA is therefore tested rather than assumed.

CLT explains why CCUs should be associated with CLR. When operational demands exceed working-memory capacity, performance declines (Sheikh et al., 2024). In I5MEs, operators must interpret complex data, monitor autonomous systems, and manage exceptions simultaneously. CCUs can reduce extraneous load by filtering irrelevant material, prioritizing urgent signals, and presenting feedback coherently (Bucci et al., 2024). Cognitively transparent human-AI interfaces can consequently reduce perceived workload and improve intervention accuracy during anomalies (Ghobakhloo et al., 2023a; Aldabousi et al., 2025; Shemais et al., 2024).

Evidence from manufacturing dashboards indicates that cognitive load depends strongly on information architecture rather than information volume alone (Fernández-Miguel et al., 2025b). By aligning technical complexity with human capacity, CCUs can reduce coordination strain, role ambiguity, and operator fatigue while sustaining decision quality during disruption (Villar et al., 2023; Sheikh et al., 2024).

STS identifies effective coordination between human and technological actors as a foundation of adaptive performance under uncertainty (Villar et al., 2023). HMMT, conceptualized as calibrated cognitive trust, enables appropriate reliance on automation, timely human intervention, and flexible human-machine coordination. Trust in semi-autonomous systems has been associated with faster coordination, fewer errors, and greater flexibility during abnormal events (Rožanec et al., 2023), and high-trust human-machine teams can outperform low-trust teams in dynamic production environments (Camarinha-Matos et al., 2024). Calibrated trust may also release cognitive resources for sensemaking by reducing redundant verification (Fernández-Miguel et al., 2025b). Excessive trust, however, can reduce vigilance and create automation complacency during novel disruptions (Villar et al., 2023). The association between HMMT and OA must therefore be established empirically.

CLT provides a direct rationale for the HMMT-CLR association. Low trust compels operators to monitor outputs continuously, justify algorithmic recommendations, and simulate alternative responses, increasing cognitive strain (Lou et al., 2024). Calibrated trust stabilizes expectations and reduces uncertainty about system behaviour. Human-automation research links such trust to lower perceived workload, less intensive monitoring, and greater decision confidence (Priyadarshini et al., 2024). Transparent and trustworthy AI-assisted planning similarly reduces cognitive burden (Merchán-Cruz et al., 2025). From an STS perspective, trust also aligns human and machine agency, reducing role ambiguity and cognitive conflict at the interface (Bucci et al., 2024).

CLT also explains the CLR-OA association. Lower extraneous load frees cognitive resources for problem-solving, learning, and adaptive decisions (Keshmiry and Hassani, 2025). In I5MEs, this capacity helps operators identify weak signals, correct errors, reconfigure operations, and convert disruption into learning (Lou et al., 2024; Saniuk et al., 2024). CLR therefore aligns technical complexity with human capacity and supports coordinated system improvement (Fernández-Miguel et al., 2025b).

STS proposes that individual learning produces system-level effects only when supporting structures are in place. APPs provide adaptable rules for task allocation, decision authority, and the limits of human and machine autonomy (Lou et al., 2024). Clear authority and escalation procedures turn individual cognitive clarity into coordinated responses and distribute learning across decentralized production systems (Saniuk et al., 2024; Ghobakhloo et al., 2023a). APPs should therefore strengthen the extent to which CLR is associated with OA by enabling timely decisions, reconfiguration, and learning during disruption.

APPs may also contribute directly to OA. By defining autonomy levels, decision rights, and escalation procedures, they facilitate rapid reconfiguration and shared agency during disruption (Rožanec et al., 2023; Lou et al., 2024). Rule-based, modular, and decentralized production arrangements support experimentation and learning under uncertainty, while clear autonomy boundaries reduce coordination failures in human-robot collaboration (Keshmiry and Hassani, 2025; Villar et al., 2023; Camarinha-Matos et al., 2024). APPs thus reduce vulnerabilities associated with over-centralization and unstructured automation.

CLT positions CLR as the mechanism linking CCUs to OA. Although CCUs can improve information processing and decision support, poorly designed systems may overload operators and weaken adaptive performance (Bucci et al., 2024; Camarinha-Matos et al., 2024). Their benefits therefore depend on presenting information in a cognitively accessible form, especially during anomalies (Ghobakhloo et al., 2023a; Keshmiry and Hassani, 2025). By aligning technical complexity with human decision capacity, CLR enables CCUs to support learning and improvement through disruption (Lou et al., 2024).

CLR should likewise mediate the HMMT-OA relationship. Mistrust or miscalibrated trust increases uncertainty, monitoring, and verification, weakening adaptive decisions under disruption (Keshmiry and Hassani, 2025; Villar et al., 2023). Calibrated trust reduces workload and mental fatigue, supports cognitive stability (Lou et al., 2024), and coordinates human and technical subsystems (Bucci et al., 2024). HMMT should therefore relate to OA through the cognitive processes required for adaptive learning and coordinated action (Saniuk et al., 2024).

Following Figure 1, which shows the conceptual framework:

Figure 1 presents the proposed research model.

The study population comprised large manufacturing firms in India, Pakistan, Bangladesh, Sri Lanka, and Nepal. Regional manufacturing directories and industry associations identified a sampling frame of approximately 820 firms with more than 200 employees and advanced production technologies. This setting is appropriate because manufacturers across the region are adopting AI-supported decision systems, human-robot collaboration, and cyber-physical production while facing volatile markets, labour-intensive operations, and infrastructure constraints (Fernández-Miguel et al., 2025a). The organization was the unit of analysis. Firms were selected by a computer-generated simple random sequence, and one qualified key informant from senior management, operations, production systems, or digital transformation represented each participating organization. Six experts (three senior practitioners with more than ten years of Industry 5.0 implementation experience and three manufacturing or operations scholars) pretested the questionnaire for relevance and clarity. Of 520 questionnaires distributed, 215 usable responses were received after follow-up reminders and selected site visits, yielding a usable response rate of 41.35%. Table 1 reports the sample profile. Respondents occupied senior management (28.37%), operations management (26.98%), production/system management (25.58%), and digital-transformation management (19.07%) roles; 84.19% had at least five years of experience and 77.68% held postgraduate qualifications. These characteristics support the key-informant design.

The sample of 215 also exceeded the PLS-SEM ten-times heuristic. The most complex endogenous specification contained five incoming effects, implying a heuristic minimum of 50 observations; the achieved sample was more than four times that value. It also provided a substantial basis for bootstrapping the mediation and moderation paths, although the estimates remain subject to the limitations of a single-informant cross-sectional design (Hair et al., 2014).

The questionnaire adapted established measures of cognitive systems, human-machine interaction, and advanced manufacturing management. All items used a five-point Likert scale from strongly disagree (1) to strongly agree (5). CCUs were measured with 18 items covering adaptive decision support, information structuring, and exception handling (Sheikh et al., 2024; Lou et al., 2024; Villar et al., 2023). HMMT comprised six items on reliability, predictability, interpretability, and trust in system recommendations (Priyadarshini et al., 2024; Villar et al., 2023). CLR used nine items on mental effort, information overload, and decision clarity (Bucci et al., 2024); APPs used seven items on task ownership, autonomy boundaries, escalation, and flexible authority (Camarinha-Matos et al., 2024); and OA used six items on learning, adaptation, and performance improvement following disruption (Sheikh et al., 2024). Appendix A provides the complete instrument and sources. Table A1 in the Appendix A shows the measurement instrument.

Because exogenous and endogenous variables were reported by the same respondents at one time, common method bias (CMB) was possible (Ghobakhloo et al., 2023b; Bucci et al., 2024). Procedural safeguards were incorporated into the survey design, and Harman's single-factor test was used as a supplementary diagnostic (Podsakoff et al., 1997). The first unrotated factor explained 42.87% of the variance, below the conventional 50% warning threshold. This result suggests that no single factor dominated the covariance; however, Harman's test cannot exclude CMB, and this residual risk is acknowledged in the limitations.

Table 1 summarizes the participating firms and respondents by firm size, management structure, operating history, position, experience, education, and gender, thereby documenting the context and suitability of the key informants.

The 215 respondents represented experienced decision makers: 63.72% of participating firms had operated for more than 20 years, 84.19% of respondents had at least five years of professional experience, and 77.68% held postgraduate qualifications. The sample was 72.56% male and 27.44% female (Table 1).

The hypotheses were tested in SmartPLS 3.2.8 using PLS-SEM, which is appropriate for prediction-oriented models containing mediation and moderation and is less restrictive regarding multivariate normality than covariance-based SEM (Götz et al., 2010; Hair et al., 2014). The reflective measurement model was assessed for indicator reliability, internal consistency, convergent validity, and discriminant validity. Standardized loadings ranged from 0.623 to 0.961, exceeding the 0.50 minimum. Composite reliability values were satisfactory, and all average variance extracted values exceeded 0.50 (Hair et al., 2014). Discriminant validity was evaluated using the heterotrait-monotrait ratio (HTMT), applying thresholds of 0.85 for conceptually distinct constructs and 0.90 for closely related constructs (Henseler et al., 2015). All reported HTMT values met the applicable criterion. Variance inflation factors were below 5, indicating no problematic collinearity (Table 2).

Path significance was evaluated by bootstrapping 2,000 subsamples (Table 3). CCUs and HMMT were positively associated with CLR, supporting H2 and H4, and CLR was positively associated with OA, supporting H5. APPs were positively associated with OA and positively moderated the CLR-OA relationship, supporting H7 and H6, respectively. The direct CCUs-OA and HMMT-OA paths were not significant; thus, H1 and H3 were not supported. The indirect effects were significant, and variance accounted for was 85.14% for CCUs → CLR → OA and 83.96% for HMMT → CLR → OA (Table 4). Because both values exceed 80%, the results support full mediation and therefore H8 and H9.

The model explained 42.6% of the variance in CLR (R² = 0.426; adjusted R² = 0.418) and 61.2% of the variance in OA (R² = 0.612; adjusted R² = 0.605), as reported in Table 5. The standardized root mean square residual was 0.061, indicating acceptable fit for the PLS-SEM model (Table 6; Hair et al., 2014). Together with the significant indirect effects in Table 4, these results confirm full, rather than partial, mediation.

Predictive relevance was medium for OA (Q² = 0.322) and CLR (Q² = 0.267), based on the 0.02, 0.15, and 0.35 benchmarks for small, medium, and large values (Cohen et al., 2013). Effect sizes (Table 7) show a large APPs-OA effect (f² = 0.459), a medium CLR-OA effect (f² = 0.168), medium CCUs-CLR (f² = 0.341) and HMMT-CLR (f² = 0.276) effects, and negligible-to-small direct CCUs-OA and HMMT-OA effects (Cohen, 1988). Figure 2 illustrates the positive interaction: the CLR-OA relationship is stronger when APPs are high.

The findings support a human-centred, cognition-based explanation of OA in I5MEs. Addressing RQ1 and RQ2, CCUs and HMMT were positively associated with CLR but had no significant direct associations with OA; CLR fully mediated both relationships. Intelligent coordination and calibrated trust therefore appear to contribute to antifragility by reducing cognitive load rather than by acting as independent performance drivers. CCUs structure complex information and prioritize signals, while calibrated trust stabilizes expectations and reduces unnecessary monitoring. This interpretation accords with evidence that cognitive control systems and calibrated trust reduce workload and improve attentional allocation (Lou et al., 2024; Villar et al., 2023). It also extends prior research by showing that, after CLR is included, the direct CCUs-OA and HMMT-OA paths become negligible. Consistent with CLT (Sweller, 1988), cognitive capacity is not an ancillary ergonomic concern but the mechanism through which technological and relational resources are associated with antifragile performance.

Addressing RQ3, APPs were positively associated with OA and strengthened the CLR-OA relationship. This pattern is consistent with the STS principle of joint optimization (Trist and Bamforth, 1951): APPs align social and technical components by specifying decision rights, escalation routes, and task allocation between humans and machines. Cognitive clarity is therefore more strongly associated with OA when operators also have flexible guidance and clear authority to act. The large APPs-OA effect and medium CLR-OA effect reinforce this interpretation, whereas the direct effects of CCUs and HMMT on OA were negligible. In practical terms, formal governance protocols appear to be the strongest individual lever in the model, followed by interface and workflow design that reduces cognitive load. Technology and trust investments contribute primarily by improving that cognitive pathway. The findings thus locate antifragility in the alignment of cognition, calibrated trust, and structured agency rather than in automation intensity alone.

The study makes four theoretical contributions. First, it integrates CLT and STS in a single account of OA, identifying CLR as the mechanism linking CCUs and HMMT to antifragility. Second, it conceptualizes APPs as both an antecedent and a boundary condition, extending STS by treating the allocation of agency as an active governance process. Third, it sharpens the distinction between antifragility, resilience, and adaptability by showing that improvement through disruption depends on aligned cognitive, relational, and structural mechanisms. Fourth, evidence from South Asian manufacturers extends Industry 5.0 research beyond its predominant Western and East Asian settings, while leaving cross-context generalizability for future testing.

Managers should design Industry 5.0 systems around cognitive usability rather than automation intensity alone. Decision-support interfaces should filter irrelevant information, prioritize critical signals, and explain recommendations. Calibrated trust requires operator training, transparent feedback, and explicit communication of system limits. APPs should formalize task ownership, autonomy boundaries, and escalation routes instead of leaving them as implicit consequences of technology deployment. Technology investment should therefore be paired with workflow simplification, interface design, and workforce development so that operators can interpret outputs and exercise authorized judgement. Regulators and industry associations can support these practices through standards for cognitive workload, interface transparency, explainable AI, and ethical governance of autonomous systems. Such measures can improve productivity and operational antifragility while protecting worker well-being and supporting an inclusive transition to human-centred manufacturing.

This study developed and tested an integrated model of OA in I5MEs. In response to RQ1 and RQ2, CCUs and HMMT were not directly associated with OA; instead, CLR fully mediated both relationships. In response to RQ3, APPs were positively associated with OA and strengthened the CLR-OA relationship. Antifragile manufacturing therefore appears to arise from the alignment of cognitive control, calibrated trust, cognitive efficiency, and structured agency rather than from automation intensity alone. The study advances Industry 5.0 theory by integrating cognitive and sociotechnical mechanisms in a mediation-moderation model, and it directs practice toward cognitively manageable interfaces, trust calibration, explicit governance protocols, and workforce development.

Several limitations qualify these findings. First, the cross-sectional design supports associations, not causal inference, and cannot capture the co-evolution of human-machine relationships; longitudinal studies, smart-factory logs, and digital-twin experiments could examine these dynamics. Second, the sample was limited to large manufacturers in five South Asian countries. Replication in small and medium-sized firms, services, and other regions is needed. Third, the study relied on single-informant managerial perceptions. Although procedural safeguards and a statistical diagnostic were applied, common method and social-desirability biases cannot be ruled out; future studies should combine surveys with production logs, downtime, quality outcomes, and real-time workload measures. Finally, research could test leadership, culture, workforce skills, and environmental volatility as additional boundary conditions and examine antifragility at individual and team levels. These extensions would refine the conditions under which technology, cognition, and sociotechnical governance jointly support human-centred antifragility.

The authors are thankful to the Deanship of Graduate Studies and Scientific Research at the University of Bisha for supporting this work through the Fast-Track Research Support Program.

The supplementary material for this article can be found online