Human-centric Automation – A Functional, Dynamic and Transactional Analysis

Industry 5.0 promises to put humans back at the center of manufacturing. But it's far more complicated than that. By contributing writer: Sambit Ghosh, PhD, from HSN Solutions.

Abstract

The article presents a novel analytical framework for human-centric automation, addressing the critical intersection of technological disruption and human work in the era of Industry 5.0 (a shift beyond pure efficiency toward sustainability and worker well-being). Structuring our analysis through three core modalities—Functions, Dynamics, and Transactions—we interrogate the ontology of the human-in-the-loop not merely as a production component, but as a cognitive agent seeking to minimize decision-driven feedback variance. We synthesize developments in task-based macroeconomics, cognitive science, and network theory to argue resilient industrial systems must preserve the variance bound of human agency. We conclude that realizing UN SDG 8 (Decent Work and Economic Growth) and UN SDG 9 (Industry, Innovation, and Infrastructure) require transforming automation from a profit-driven implementation to an enterprise-wide human agency support and scaffolding for human skill improvement.

Introduction

We live in an era of polycrisis, defined by precariousness and epistemic uncertainty [1]. While this is affecting all aspects of human life, the manufacturing sector is particularly at the forefront as it often acts as a central point of multiple human and environmental factors. Hence, the global manufacturing landscape is undergoing a profound structural transformation, driven by disruptive technologies that are not merely reshaping markets but fundamentally questioning the ontology of human labor. Against this backdrop, the United Nations Sustainable Development Goals—specifically SDG 8 (Decent Work and Economic Growth) and SDG 9 (Industry, Innovation, and Infrastructure)—provides the critical normative anchor for our analysis. They compel us to ask: How do we build resilient industrial systems that do not just survive technological disruption but actively enable and secure the dignity and agency of the human worker?

While Artificial Intelligence (AI) and its vast technological machinery have democratized access to tools through open-source innovation and flexible pricing, this rapid diffusion often masks a deeper friction: the churn of digital transformation is leaving behind a wake of polarization, a widening gap between tech-absorbers who leverage AI for augmentation and resource-constrained entities where automation risks “stunting” human potential [2]. The social transaction of work—the intimate rhythm of production that connects an individual to their community—is increasingly mediated by algorithmic opacity, challenging the historical distinction between meaningful work and mere metabolic labor.

This paper interrogates these challenges through a novel human-centric automation lens, structured around three modalities: Functions, Dynamics, and Transactions (FDT).

• Functions: Examining how human skills are codified as functional capabilities that define the human’s role in the loop.

• Dynamics: Utilizing Active Inference and Network Theory to explain the human agent as an active participant seeking to minimize uncertainty by aligning internal mental models with external reality.

• Transactions: Analyzing the economic and informational exchanges where biological health, cognitive load, and ethics act as non-negotiable constraints.

Theoretical Framework

Cognitive Science and Skill Codification

Cognitive science has evolved into vast subdomains and a number of theories explaining human cognition. Transitioning from the fundamentals of memory, perception and action, advanced functioning like decision-making, problem-solving, and expertise have also been studied extensively. However, classical cognitive models often fail to account for the context-dependence of human decision-making under stress. Quantum Cognition suggests human agency is quantum probabilistic; for instance, the order in which an operator receives a safety alarm versus a production target fundamentally alters their decision state [3].

The function a HiTL performs in the workplace has also evolved. Integrating these realities into an economic framework requires a precise language of tasks and skills. Acemoglu and Zilibotti [4] highlight how technologies and tasks designed for high-skilled labor in the Global North often lead to productivity stagnation in the Global South due to skill mismatch. This mismatch creates a skill-bias friction. AI offers a reversal of this dynamic: by designing AI as a dynamic scaffold, we can lower the entry barrier for complex tasks, turning production goals into learning opportunities. By using outcome-based competency frameworks, we transform the human-in-the-loop (HiTL) into an agent with specific functional properties, enabling clear pathways for career progression.

Multi-Scale Networks and Active Inference

Within this skill-codified ecosystem, reality is defined by the coupling of an external physical network of production with the internal, evolving mental network models held by individual human agents. This multi-scale interaction is analyzed using graph signal processing to isolate specific informational or emotional signals as the network evolves through discrete transactional exchanges of energy and data. We ground the dynamics of this interaction in the principle of Active Inference. In traditional Active Inference models, agents minimize surprise (or variational free energy) by updating their internal probabilistic models to match sensory inputs. Additionally, they update their internal predictive models to update actions on the environment. In our framework, this generative model is upgraded to a dynamic network topology. The human agent seeks to minimize the divergence between their internal network (mental model) and the external network (physical reality). This minimization occurs through two pathways: perception (updating the internal network to better match reality) and action (manipulating the external network to better align with the internal model) [5]. The interactions occur via a Markov Blanket—a set of states that separates the worker’s cognitive processes from external reality. A healthy HiTL system ensures this blanket is permeable enough for effective action but robust enough to prevent cognitive overload and burnout.

The Task-Based Economic Network

Finally, the manufacturing environment is a shared task network [6] where skill is a dynamic tool for navigation. High skill levels correspond to a greater capacity to predict network dynamics and resolve uncertainty. The worker uses codified skills to navigate the active inference cycle, maintaining stability within the shared human-machine task network.

The Nature of Work

To understand human-automation interaction, we must distinguish between fundamental activities. Labor is the cyclical activity required for biological survival, while Work is teleological, intended to create a durable world of artifacts [7]. Friction arises when industrialization forces the body’s natural rhythm to align with the relentless, non-natural sequence of machines. Historically, industrial transitions have often relied on coercive structures or led to the polarization of labor markets. However, social democratic transactions—such as cultural activities and open debate—ensure that human perceptions are preserved in a human-centric dialectic.

As systems become more automated, the human operator’s role often becomes more difficult, not easier [8]. Humans are removed from routine control but expected to intervene instantly during critical failures. Technology often introduces hidden cognitive tasks—mental operators required to bridge the gap between the system’s logic and the real-world task. This is a failure of functional alignment: optimizing machine speed while ignoring the human’s cognitive Active Inference loop.

Decent and Meaningful Work

The pursuit of automation must be linked to the psychological reality of the worker. Decent work—safe conditions, adequate time, and organizational values—is the antecedent to meaningful work [9]. Precarious work strips the meaning-making capacity from the worker. Hence, when algorithms dictate tasks without context, the worker cannot perceive their contribution to the greater good, leading to alienation. Moreover, the cognitive load of monitoring complex systems can lead to mental fatigue as debilitating as physical injury, as evident from Industry 4.0 driven advanced decision-making. Consequently, the transition to Industry 5.0 prioritizes a human-centric, resilient framework where technology is adaptable to the diversity of the worker [10].

HiTL AI

HiTL is defined by who controls the learning process: Active Learning (system queries human), Interactive Machine Learning (shared control), or Machine Teaching (human curates knowledge) [11]. The black box of algorithms must be bridged by Explainable AI (XAI), shifting from technical descriptions to answering human-centered questions like "How do I change this?" [12]. The field is moving toward Human-Autonomy Teaming, as the cobot applications in manufacturing are indicating. Success relies on Shared Mental Models, where the human models the AI’s capabilities and the AI computationally models the human’s state [13]. Here, trust is the central currency. Interestingly, operators often prioritize functional reliability over the identity of the partner (Human vs. AI), but trust remains fluid and relational [14]. Design must pivot to a Human-Centered Human-AI Collaboration that prioritizes well-being over efficiency. Additionally, we need to guard against the moral buffer of automation, where distance from decision-making leads to complacency or emotional detachment [15].

Uncertainty

A fundamental risk in modern manufacturing is the conflation of automation (self-acting mechanisms) with autonomy (the human capacity for moral law). True innovation is a virtue-based social practice that AI cannot replicate [16]. Algorithmic management can reduce workers to appendages of the machine, blocking intrinsic social goods. Optimal human-machine teams must actually maximize entropy production—encouraging conflict and debate—to process uncertainty effectively and ensure the adjustment of the Markov blankets. If an AI suppresses edge-case signals to present a clean solution, it makes the system brittle to uncertain events.

For the AI teammate, this necessitates robust Uncertainty Quantification (UQ). Liu [17] distinguishes between aleatoric uncertainty (inherent noise) and epistemic uncertainty (lack of knowledge), arguing AI must communicate both to be a reliable partner. However, Vashurin [18] identifies a critical “alignment gap”: an AI model can be mathematically calibrated (its confidence score matches its accuracy) yet still fail to improve human decision-making if the uncertainty is not visualized intuitively. Thus, the challenge is not merely algorithmic but sociotechnical—metrics must measure “decision deferral rates” (when the AI correctly yields to the human) rather than just raw accuracy.

Applications in Manufacturing

In real scenarios, the application of FDT reveals manufacturing is a microcosm of broader societal exchanges. Extrapolating from current industrial trends, we observe Industry 4.0 is effectively a high-velocity transaction market. Every sensor reading, every work order, and every human intervention is a bid in a market for efficiency. However, as in financial markets, high-frequency trading (or in this case, high-frequency task allocation) can destabilize the market participants.

The entire manufacturing ecosystem operates as an information transaction network. Human skills are the functional units that determine the type of transactions. Micro, Small and Medium Enterprises (MSME) operate under extreme constraints (financial, resource, logistics, human-skill, etc.) and overwhelmingly lack automation. While the larger industrial entities act as technology absorbers, the MSMEs are increasingly being left behind. Technology absorption inherently includes access to enterprise-wide high-skilled HiTLs. This widening gap not only affects the MSME owners and promoters in their own active inference loop, but the workers are also stuck in a manufacturing reality that is increasingly being pressurized to meet stringent regulations. From a human perspective, their active inference loop operates under extreme conditions—they do not have the scope to minimize uncertainty when it comes to family planning, raising children with decent education, living with dignity, and accessing good healthcare. As a result, because they are unable to minimize uncertainty, they have to adjust their biological functioning, leading to health hazards and low-skilled jobs. Not to mention their constant fears of being replaced by automation and subsequent lack of trust in using new technologies. Moreover, the MSME owners cannot risk investing in better technology (hence enhancing productivity) or afford to train their workers (upskilling). Hence, these MSMEs are stuck in a vicious cycle. Often, conversations with such MSME owners and workers lead to an uncomfortable conclusion:they are completely aware of this situation and still unable to do anything.

The transition to advanced automation is not a uniform upgrade but is constrained by the historical distribution of MSMEs, their varying levels of digital maturity, and the local availability of education and skills. Successful integration of HiTL systems requires more than just technology; it requires a maintenance network—support systems for service, operations, and technical training. Without these, the mentality of tech adoption remains defensive, and the transition leads to stunting rather than augmentation. Often, one has to spend a significant amount of time with the owners and workers to understand them closely, both culturally and functionally, and provide customized solutions they trust will work immediately. This is challenging, as even borrowing a few hours of their time for a few days is expensive to them and is highly uncertain in its outcome. A significant fraction of MSMEs are pre-digital, hence a discussion of data or AI has to start from the point at which they are currently operating.

It is in such scenarios the FDT framework provides a foundation to understand and map the variety of technical and human-centric attributes that dynamically affect the MSME one is engaging with. Broadly, the functional mode provides the immediate solutions that will benefit the enterprise and hence the technological (hardware and software) requirements. The dynamic mode allows one to plan a scaffold of skill transitions for not only the workers but also the owners and promoters. The transaction mode governs the economic models, return-of-investment, and profit forecasts. In all modes, AI plays a critical role in developing and deploying the solutions.

Automation Design for Decent and Meaningful Work

The FDT framework points toward a unified design philosophy for automation. To achieve decent work, automation must be designed as a scaffold for competence, not a substitute for it. Drawing from the literature on Human Performance Modeling and System Control [19], effective automation design must adhere to the principle of Observability and Directability. The internal state of the automation must be visible to the user (reducing epistemic uncertainty), and the automation must be responsive to the user’s redirection (preserving agency). When automation is opaque (black box), it breaks the dynamics loop; the operator cannot minimize surprise because they cannot predict the machine’s behavior. Navigating the skill scaffolding is therefore hindered.

Furthermore, we propose meaningful work in automation is achieved through cognitive coupling. Rather than automating the thinking and leaving the human with the monitoring (which leads to vigilance decrement), systems should automate the routine calculation while engaging the human in high-level constraint setting and anomaly resolution. This preserves the function of the human as a strategic problem solver and ensures the codified human skills are periodically active.

Finally, at the transactional level, design must incorporate HiTL-by-Design. This means the system’s objective function should include constraints on human factors like fatigue and stress. An optimization algorithm that maximizes production throughput but pushes the human operator into a high-risk fatigue state must be mathematically penalized within the control logic itself. By embedding physiological variance bounds into the machine’s code, we operationalize the protection of the worker, transforming the abstract right to safety into a hard, systemic constraint.

The complete FDT framework also enables SDG 9 and Industry 5.0. By ensuring the enterprise-wide Humans-in-The-Loop decent and meaningful work, pathways for innovation, sustainability, and industrial stability are generated within the enterprise through enterprise-wide active inference decision-making.

---

References

1. World Economic Forum (2025). Global Risks Report.

2. Furendal, M. and Jebari, K., 2023. The future of work: augmentation or stunting?. Philosophy & Technology, 36(2), p.36.

3. Pothos, E.M. and Busemeyer, J.R., 2022. Quantum cognition. Annual review of psychology, 73(1), pp.749-778.

4. Acemoglu, D. and Zilibotti, F., 2001. Productivity differences. The Quarterly Journal of Economics, 116(2), pp.563-606.

5. Parr, T., Pezzulo, G. and Friston, K.J., 2022. Active inference: the free energy principle in mind, brain, and behavior. MIT Press.

6. Acemoglu, D., 2025. The simple macroeconomics of AI. Economic Policy, 40(121), pp.13-58.

7. Arendt, H., 1958. The human condition. University of Chicago Press.

8. Bainbridge, L., 1983. Ironies of automation. In Analysis, design and evaluation of man–machine systems (pp. 129-135). Pergamon.

9. Duffy, R.D., Blustein, D.L., Diemer, M.A. and Autin, K.L., 2016. The psychology of working theory. Journal of counseling psychology, 63(2), p.127.

10. Schulte, P.A., Fischer, F.M., Iavicoli, I., de Anchieta Messias, I., Fisher, D.L., Popkin, S.M., Martinez, M.C., Cotrim, T. and Oakman, J., 2024. The Challenge of New Forms of Work, Innovative Technologies, and Aging on Decent Work: Opportunities for Occupational Safety and Health. La Medicina del Lavoro, 115(5), p.e2024037.

11. Mosqueira-Rey, E., Hernández-Pereira, E., Alonso-Ríos, D., Bobes-Bascarán, J. and Fernández-Leal, Á., 2023. Human-in-the-loop machine learning: a state of the art. Artificial Intelligence Review, 56(4), pp.3005-3054.

12. Liao, Q.V. and Varshney, K.R., 2021. Human-centered explainable ai (xai): From algorithms to user experiences. arXiv preprint arXiv:2110.10790.

13. Andrews, R.W., Lilly, J.M., Srivastava, D. and Feigh, K.M., 2023. The role of shared mental models in human-AI teams: a theoretical review. Theoretical Issues in Ergonomics Science, 24(2), pp.129-175.

14. Chiou, E.K. and Lee, J.D., 2023. Trusting automation: Designing for responsivity and resilience. Human factors, 65(1), pp.137-165.

15. Eich, A., Klichowicz, A. and Bocklisch, F., 2023. How automation level influences moral decisions of humans collaborating with industrial robots in different scenarios. Frontiers in psychology, 14, p.1107306.

16. Chiodo, S., 2022. Human autonomy, technological automation (and reverse). AI & society, 37(1), pp.39-48.

17. Liu, X., Chen, T., Da, L., Chen, C., Lin, Z. and Wei, H., 2025, August. Uncertainty quantification and confidence calibration in large language models: A survey. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2 (pp. 6107-6117).

18. Vashurin, R., Fadeeva, E., Vazhentsev, A., Rvanova, L., Vasilev, D., Tsvigun, A., Petrakov, S., Xing, R., Sadallah, A., Grishchenkov, K. and Panchenko, A., 2025. Benchmarking uncertainty quantification methods for large language models with lm-polygraph. Transactions of the Association for Computational Linguistics, 13, pp.220-248.

19. Sheridan, T. B., 2002. Humans and automation: System design and research issues. John Wiley & Sons, Inc.; Human Factors and Ergonomics Society.