PRIME–INSPECT: A Socio-Technical Framework for Trustworthy Intelligent Automation and Real-Time Decision-Making in Industry 4.0

Abstract

Intelligent automation is a core component of Industry 4.0, enabling artificial intelligence (AI) systems to support or execute operational and managerial decisions in real time. In high-risk industrial environments such as mining and metallurgy, real-time decision-making improves efficiency but also raises critical challenges related to trust, explainability, human oversight, and institutional accountability. This study proposes PRIME–INSPECT, a two-layer socio-technical framework designed to support trustworthy AI-driven real-time decision-making. The PRIME (predict, regulate, interpret, mitigate, execute) layer formalizes the operational decision flow, embedding control mechanisms, uncertainty quantification, and explainability into the automation pipeline. The INSPECT (integrity, navigability, supervisory control, policy maturity, ethical compliance, collaboration, trust calibration) layer defines the organizational and governance conditions required for safe deployment. The framework is conceptually developed through a structured literature synthesis and supported by exploratory empirical grounding through stakeholder perceptions from IT and top management participants, alongside an illustrative industrial use case intended to demonstrate conceptual applicability rather than engineering performance validation. The findings highlight the importance of aligning operational AI processes with institutional safeguards to support calibrated trust and responsible automation. The empirical component is intended to provide conceptual and organizational grounding of framework dimensions rather than quantitative validation of predictive performance. PRIME–INSPECT provides a structured architecture for designing and governing AI-enabled real-time decision systems in high-risk industrial contexts.

1. Introduction

Digital transformation has entered a phase in which artificial intelligence (AI) is no longer merely an analytical support tool, but an active executor and coordinator of operational and managerial decisions. In contemporary real-time decision-making systems (RTDs), data collection, analysis, recommendation generation, and decision execution occur in sub-second latencies, often without the possibility of manual ex ante validation for each decision. While such systems promise speed, efficiency, and optimization, they simultaneously introduce new classes of organizational and operational risk, as errors may propagate faster than traditional control and governance mechanisms can respond [1,2].

These challenges are particularly pronounced in high-risk industrial systems, such as mining, metallurgy, and other process industries. In these environments, RTD relies on continuous streams of sensor data (SCADA/IoT), process analytics, predictive maintenance, and automated control of critical parameters. The consequences of incorrect AI-driven recommendations in such systems extend far beyond financial losses and may include safety incidents, environmental damage, regulatory non-compliance, and reputational harm. Consequently, operational continuity in these sectors depends not only on predictive accuracy but also on trustworthiness, explainability, human oversight, and institutional accountability [3].

The existing literature offers well-established models of technology acceptance, such as the Technology Acceptance Model (TAM) and the Unified Theory of Acceptance and Use of Technology (UTAUT), which emphasize perceived usefulness, ease of use, and behavioral intention [4,5]. More recent models extend this perspective by addressing ambivalence and resistance toward AI through acceptance–avoidance mechanisms [6]. However, these frameworks primarily explain whether users are willing to adopt AI technologies and provide limited guidance on how organizations should govern AI-driven decisions in real-time, particularly in environments where risk, safety, and accountability are central concerns [7].

Recent research increasingly emphasizes that trust in AI, model explainability (XAI), human-in-the-loop control, and AI governance mechanisms are critical enablers of responsible AI adoption [8,9,10]. In real-time decision-making systems, trust cannot be treated as a static psychological preference; rather, it represents a dynamic relationship shaped by transparency, the ability to intervene, clearly defined responsibilities, and the organization’s capacity to monitor and audit algorithmic behavior. Insufficient explainability and governance increase the risk of both algorithm aversion and automation bias—either rejecting AI after isolated errors or over-relying on algorithmic outputs despite uncertainty [11,12].

Despite these advances, existing research remains fragmented across several dimensions. Current approaches typically address isolated aspects of AI-driven decision-making—such as predictive accuracy, user acceptance, explainability, or governance—without providing an integrated operational architecture that connects real-time AI decision pipelines with institutional oversight mechanisms in high-risk industrial environments. As a result, organizations lack a structured framework that simultaneously addresses how AI-driven decisions are generated, constrained, interpreted, and executed, while ensuring that these processes remain aligned with governance, accountability, and trust requirements under real-time conditions. To address this gap, this paper proposes PRIME–INSPECT, an integrative two-layer socio-technical framework for trustworthy AI-driven real-time decision-making. In this framework, the PRIME layer represents the operational AI decision layer, encompassing the end-to-end process through which data-driven predictions are transformed into actionable decisions under real-time constraints. In contrast, the INSPECT layer represents the governance and institutional oversight layer, defining the organizational, ethical, and regulatory conditions under which such decisions can be deployed safely and responsibly.

Operationally, the PRIME layer—whose name encodes its five sequential stages of predict, regulate, interpret, mitigate, and execute—captures the end-to-end decision flow from data-driven prediction to constrained, auditable execution. The INSPECT layer, an acronym for integrity of data, navigability of explanations, supervisory control, policy and governance maturity, ethical compliance, collaboration between management and technical experts, and trust calibration, defines the organizational and institutional conditions under which these decisions can be deployed safely and responsibly. A distinctive contribution of the framework is the explicit formalization of trust calibration, emphasizing that neither insufficient trust nor blind reliance on AI is desirable; instead, optimal performance emerges from a context-dependent balance between automation and human oversight [9,13]. Within this operational layer, AI models play a central role—particularly in the predict and interpret stages—where machine-learning algorithms generate forecasts, detect anomalies, and provide explanatory outputs that support real-time human and automated decision-making.

PRIME–INSPECT is designed as a domain-agnostic framework applicable across industries. In this study, mining and metallurgy are used as illustrative prototypes of high-risk industrial systems in which explainability, governance, and calibrated trust are particularly critical. These sectors are characterized by intensive use of real-time data, high costs of failure, and stringent regulatory requirements, making them suitable environments for demonstrating the applicability of the proposed framework. The empirical component of the study is explicitly designed as an exploratory and illustrative grounding of the framework, rather than as a formal validation of its predictive or operational performance. Its purpose is to examine whether the core socio-technical dimensions underlying PRIME–INSPECT—such as trust, explainability, governance maturity, and human oversight—are recognizable and meaningful to key organizational stakeholders involved in AI-enabled real-time decision-making.

It is important to clarify that the purpose of this study is not to propose or validate a specific predictive algorithm, optimization method, or control-theoretic model. Rather, the objective is to introduce a socio-technical architecture that integrates operational AI decision flows with governance and oversight mechanisms in real-time industrial contexts. The empirical component is therefore positioned as exploratory grounding intended to assess the recognizability and organizational relevance of framework dimensions rather than to establish predictive superiority or formal engineering performance benchmarks.

The remainder of the paper is structured as follows. Section 2 reviews the relevant literature on AI adoption, trust, explainable AI, perceived risk, and AI governance, with particular emphasis on high-risk industrial contexts. Section 3 develops the theoretical foundations and formally introduces the PRIME–INSPECT framework. Section 4 illustrates the framework through representative processes in mining and metallurgy. Section 5 discusses theoretical and practical implications, limitations, and directions for future research.

2. Literature Review and Conceptual Foundations

2.1. AI-Enabled RTD in High-Risk Industrial Systems

AI in high-risk industrial environments is increasingly deployed through pipelines that integrate real-time data acquisition (e.g., SCADA/IoT), predictive modeling, and operational decision support. In mining, recent systematic reviews document the maturation of AI-driven predictive maintenance, metallurgical process control, and fault detection, including the use of digital twins and hybrid AI approaches to support near-real-time monitoring and asset management [3].

From an operational perspective, such deployments are embedded in socio-technical arrangements including monitoring dashboards, escalation protocols, operator verification, override mechanisms, and post hoc audits. This framing shifts the central question from “Can the model predict accurately?” to “Under what conditions can the model be trusted, supervised, and governed when decisions must be acted upon quickly?”—a requirement that becomes defining in safety- and regulation-sensitive industrial settings. Answering this question requires integrating insights from technology adoption research, trust theory, explainable AI, and governance frameworks, streams that the following subsections review in turn.

2.2. Adoption and Managerial Decision-Making Perspectives on AI

Research on adoption has traditionally relied on TAM and UTAUT to explain use intention and behavior, with perceived usefulness, ease of use, and social influence as key constructs. Yet managerial adoption of AI for decision-making introduces additional determinants related to accountability, uncertainty, and organizational readiness. Empirical work focused on managers shows that attitudes and behavioral intentions toward AI are shaped not only by performance expectations, but also by perceived control and risk in decision contexts [6].

Systematic reviews further emphasize that barriers to adopting AI-based decision systems commonly include limited transparency, lack of governance structures, and unclear allocation of responsibility—factors that adoption models capture only indirectly and that become particularly salient under real-time constraints [7].

These findings indicate that adoption intent alone is insufficient: understanding whether and how organizations can responsibly act on AI-generated decisions requires attention to trust, explainability, and oversight—addressed in the following subsections.

2.3. Trust, Reliance, and Calibration in Human–AI Decision-Making

Trust is consistently identified as a key determinant of effective human–AI collaboration; however, the literature distinguishes trust from reliance and highlights persistent miscalibration problems. Two dominant failure modes are widely reported: (i) algorithm aversion, where users reject algorithmic advice after observing errors, and (ii) automation bias, where users accept incorrect AI recommendations or fail to verify them adequately—both of which are damaging in RTD contexts where time pressure is high [11,12].

Recent evidence syntheses indicate that automation bias occurs across high-stakes domains and is influenced by cognitive workload, time pressure, and perceived AI system authority [14].

A key contemporary direction is trust calibration, defined as designing systems and workflows that support appropriate reliance based on model competence, uncertainty, and task criticality. Empirical evidence demonstrates that explanation types can significantly affect calibration outcomes [15], while transparency and feedback mechanisms may influence reliance accuracy and user experience [16]. These findings support treating calibration not as a “soft” psychological byproduct of adoption, but as an explicit design objective in socio-technical AI deployments. Foundational theoretical work by Jacovi et al. [17] offers a formal grounding for this perspective by distinguishing between “warranted” and “unwarranted” trust in AI systems. Their formalization of “contractual trust” wherein trust is understood as a user’s belief that an implicit or explicit behavioral contract with the AI model will hold provides a conceptual basis for treating trust calibration as a designable property rather than a psychological disposition. This framing directly supports the PRIME–INSPECT view that appropriate trust emerges from verifiable, contract-like alignments between AI behavior, operational constraints, and governance expectations.

In this context, explainability and human–AI interaction design emerge as critical mechanisms for achieving appropriate trust calibration, particularly in high-risk and time-sensitive decision environments.

2.4. Explainable AI as an Operational Requirement in Industrial Contexts

Explainable AI (XAI) is often motivated by transparency and compliance. However, in industrial decision settings, it also functions as an operational capability: explanations support operator verification, troubleshooting, escalation decisions, and post-incident learning. In predictive maintenance contexts, recent systematic reviews specifically emphasize human-in-the-loop XAI as a way to improve usefulness, trustworthiness, and adoption by aligning explanations with the decision needs of maintenance stakeholders and their real workflows [18].

However, the literature also indicates that “having explanations” is insufficient: explanation class, timing, actionability, and user interaction design determine whether XAI improves calibration and safer reliance [15].

Building on this operational perspective, recent advances in explainable artificial intelligence (XAI) emphasize the importance of transparency in building trust in AI-driven decision-making systems. Ribeiro et al. [19] introduced LIME (Local Interpretable Model-Agnostic Explanations), a widely adopted method that explains individual predictions of any classifier by approximating it locally with an interpretable model. Their work highlights that trust in AI systems depends not only on predictive performance but also on the ability to understand and interpret individual model decisions.

Complementing this perspective, Amershi et al. [20] propose a set of human-AI interaction guidelines that address the challenges arising from the probabilistic and often unpredictable behavior of AI-infused systems. Their work emphasizes the importance of transparency, user control, feedback mechanisms, and clear communication of system capabilities and limitations.

Together, these perspectives collectively reinforce the need for integrated approaches that combine explainability, human oversight, and governance mechanisms within operational decision pipelines. These insights directly inform the INSPECT layer of the proposed PRIME–INSPECT framework, which incorporates explainability, supervisory control, and trust calibration as core components of trustworthy AI-driven real-time decision-making. Yet ensuring that explanations actually influence operator behavior requires not only technical explainability mechanisms, but also formalized human oversight structures that define when and how operators can act on them—the focus of the following subsection.

2.5. Human Oversight: Supervision, Control Rights, and Teaming

A growing stream of work formalizes “human control” of AI systems and differentiates two main approaches: supervisory human control (humans monitor and can intervene) and human–machine teaming (roles and authority are distributed dynamically) [21].

For RTD in high-risk industrial systems, this literature implies that oversight must be operationalized through explicit control rights (e.g., stop/override), defined intervention points, escalation rules, and auditable decision pathways—supported by explainability and governance mechanisms rather than informal practice.

Yet formalizing these control rights requires institutional anchoring: clear policies, accountability structures, and governance maturity that extend beyond the technical system itself, as the following subsection addresses.

2.6. AI Governance and Risk Management Frameworks

AI governance research has moved from high-level principles toward structured governance practices. A recent comprehensive review synthesizes responsible AI governance as a combination of structural, relational, and procedural practices, linking governance maturity to antecedents and organizational outcomes [22].

Complementary computational perspectives on trustworthy AI have further operationalized its core dimensions. Liu et al. [23] provide a comprehensive appraisal of trustworthy AI from a computational perspective, identifying six critical dimensions—safety and robustness, fairness, explainability, privacy, accountability, and environmental well-being—and demonstrating that trustworthy AI cannot be reduced to algorithmic performance alone. Li et al. [24] extend this perspective by treating trustworthy AI as a lifecycle challenge: trustworthiness must be systematically established and assessed across data acquisition, model development, system deployment, and continuous monitoring—a view directly consistent with the governance requirements of high-risk RTD systems.

In parallel, risk management frameworks provide lifecycle-oriented guidance for identifying and managing AI risks. The NIST AI Risk Management Framework (AI RMF 1.0) formalizes trustworthy AI as a risk governance problem and emphasizes ongoing monitoring, accountability, and context-sensitive controls—features directly relevant for high-risk RTD deployments [25]. Complementary work on AI risk measurement further underscores the need to operationalize “trustworthy AI” as measurable, governable risk constructs [10]. Taken together, these governance frameworks establish what trustworthy AI should achieve but leave unspecified how operational decision pipelines should be structured to achieve it under real-time constraints—the integration gap that motivates the synthesis in Section 2.7 and the framework developed in Section 3.

2.7. Synthesis: The Integration Gap Motivating PRIME–INSPECT

Taken together, these research streams provide strong conceptual and empirical building blocks: adoption models explain intention and usage [6,7], trust and calibration studies explain appropriate reliance [14,15,16,17], XAI research explains how transparency can support human judgment [18,19,20], oversight research defines control modes and constraints [21], and governance frameworks structure organizational responsibilities and risk management [22,25]. In particular, prior research has extensively addressed explainability [19] and human–AI interaction design [20], yet these dimensions are typically treated independently rather than as integrated components of a unified socio-technical architecture. However, these contributions remain insufficiently connected to the operational logic of real-time AI decision-making, where prediction, interpretation, intervention, and execution must be tightly coupled under conditions of uncertainty and time pressure.

Three integration gaps are particularly salient.

• Micro-level process gap: There is limited formalization of the end-to-end RTD operational flow (from prediction to execution) with embedded risk controls;
• Macro-level institutional gap: Governance maturity, accountability, and collaboration are often discussed without explicit coupling to the operational pipeline;
• Human–AI coupling gap: Trust calibration is recognized as critical, but is seldom treated as a first-class design objective jointly supported by explainability, oversight, and governance mechanisms.

These gaps motivate the development of an integrative two-layer framework. In the next section, we address this need by formally developing PRIME–INSPECT, linking the micro-level flow of AI-supported RTD decisions (PRIME) to the macro-level institutional safeguards and enablers required for trustworthy deployment (INSPECT), with mining and metallurgy serving as illustrative prototypes of high-risk industrial systems.

3. The PRIME–INSPECT Framework: Architecture and Operationalization

PRIME–INSPECT is a two-layer socio-technical framework designed to address three integration gaps identified in the literature: the absence of a formalized end-to-end operational flow for AI-driven real-time decision-making; insufficient coupling between operational AI pipelines and institutional governance; and the lack of explicit trust calibration as a design objective. As illustrated in Figure 1, the framework links a micro-level operational layer (PRIME) with a macro-level governance layer (INSPECT), connected through a cross-layer trust calibration feedback mechanism.

3.1. Formal Representation of Sequential Decision Logic

The mathematical expressions introduced in this section are intended as formal representations of sequential socio-technical decision logic rather than as control-theoretic optimization models or differentiable computational operators. Their purpose is to provide conceptual precision and procedural traceability by explicitly specifying how operational stages interact within the PRIME layer.

Rather than modeling a continuous physical system through differential equations, the formalization captures structured transformations of decision states under real-time constraints. The equations therefore serve as workflow formalizations that support interpretability, auditability, and governance alignment across operational decision stages.

3.2. The PRIME Layer: Operational Decision Flow

The PRIME acronym—predict, regulate, interpret, mitigate, execute—describes the end-to-end lifecycle of an AI-supported real-time decision. Rather than prioritizing prediction accuracy alone, PRIME conceptualizes AI-driven decision-making as a controlled socio-technical process in which algorithmic outputs are continuously contextualized, constrained, and validated prior to execution.

Formally, the PRIME layer can be represented as a sequential decision function in which each stage transforms the system state under real-time constraints:

$$a\left(t\right)=(\mathrm{E}\mathrm{x}\mathrm{e}\mathrm{c}\mathrm{u}\mathrm{t}\mathrm{e}\circ \mathrm{M}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{e}\circ \mathrm{I}\mathrm{n}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{t}\circ \mathrm{R}\mathrm{e}\mathrm{g}\mathrm{u}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{e}\circ \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{t})\left(X\left(t\right);\hspace{0.17em}\theta \right)$$

(1)

where $a\left(t\right)$ is the system action at time $t$, $X(t)\in {\mathbb{R}}^d$ is the $d$-dimensional vector of real-time sensor readings, $\theta$ denotes the learned model parameters, and $\circ$ denotes sequential stage execution rather than strict mathematical composition: each stage augments a shared decision state $\mathit{S}\left(t\right)=(\widehat{y},\hspace{0.17em}\sigma ,\hspace{0.17em}v,\hspace{0.17em}\phi ,\hspace{0.17em}m)$ that is available to all subsequent stages.

Predict is the core analytical stage in which machine-learning models generate forecasts, classifications, or anomaly scores from real-time data streams. Depending on the task, applicable approaches include supervised models (e.g., Random Forest, Gradient Boosting), time-series models (e.g., LSTM), and unsupervised anomaly detection (e.g., Isolation Forest, Autoencoders). Outputs are explicitly probabilistic—quantifying uncertainty is treated as a prerequisite for downstream decision control, not an optional feature.

$$\widehat{y}\left(t\right)=f\left(X\left(t\right);\hspace{0.17em}\theta \right), \sigma \left(t\right)=g\left(X\left(t\right);\hspace{0.17em}\theta \right)$$

(2)

where $\widehat{y}\left(t\right)$ is the point prediction (e.g., remaining useful life, anomaly score), $\sigma \left(t\right)$ is the associated uncertainty estimate, $f$ is the predictive model (e.g., LSTM, XGBoost, Random Forest), and $g$ is the uncertainty quantification function (e.g., MC Dropout, ensemble variance, quantile regression).

Regulate embeds formal constraints that govern how predictive outputs may be used. These include operational limits, safety thresholds, physics-based rules, and regulatory requirements. By evaluating AI outputs against predefined boundaries before they influence system behavior, this stage prevents unsafe or non-compliant actions—including the suppression of false alarms triggered by sensor faults rather than genuine process deviations.

$$v(t)=\{\begin{array}{ll}\mathrm{pass}& \mathrm{if}\forall \hspace{0.17em}i\in \{1,\dots ,k\}:c_i(\widehat{y},\hspace{0.17em}\sigma ,\hspace{0.17em}X)\le {\tau }_i\\ \mathrm{flag}& \mathrm{if}\exists \hspace{0.17em}i:{\tau }_i<c_i(\widehat{y},\hspace{0.17em}\sigma ,\hspace{0.17em}X)\le {\tau }_i^{\mathrm{*}}\vee \sigma \left(t\right)>{\sigma }_{\mathrm{m}\mathrm{a}\mathrm{x}}\\ \mathrm{block}& \mathrm{if}\exists \hspace{0.17em}i:c_i(\widehat{y},\hspace{0.17em}\sigma ,\hspace{0.17em}X)>{\tau }_i^{\mathrm{*}}\vee \neg \hspace{0.17em}\mathrm{P}\mathrm{l}\mathrm{a}\mathrm{u}\mathrm{s}\mathrm{i}\mathrm{b}\mathrm{l}\mathrm{e}\left(X\left(t\right)\right)\end{array}$$

(3)

where $v\left(t\right)\in \{\mathrm{pass},\hspace{0.17em}\mathrm{flag},\hspace{0.17em}\mathrm{block}\}$ is the regulatory decision, $C=\{c_1,\dots ,c_k\}$ is the constraint set, ${\tau }_i$ and ${\tau }_i^{\mathrm{*}}$ are the soft and hard thresholds for the $i$-th constraint, ${\sigma }_{\mathrm{m}\mathrm{a}\mathrm{x}}$ is the maximum tolerable prediction uncertainty, and $\mathrm{P}\mathrm{l}\mathrm{a}\mathrm{u}\mathrm{s}\mathrm{i}\mathrm{b}\mathrm{l}\mathrm{e}\left(X\right(t\left)\right)$ is a physics-based consistency check. The conditions are evaluated in priority order: block takes precedence over flag, which takes precedence over pass. That is, if any constraint exceeds a hard threshold or physical plausibility fails, the system blocks regardless of other constraint states.

Interpret operationalizes explainability as a real-time requirement, not a post hoc transparency add-on. Model outputs are accompanied by explanations—such as feature importance scores or local model-agnostic explanations—enabling operators to assess the rationale, confidence, and limitations of a recommendation before acting. This stage directly supports the trust calibration mechanism described in Section 3.4.

$$\phi \left(t\right)=\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{l}\mathrm{a}\mathrm{i}\mathrm{n}\left(f,\hspace{0.17em}X\left(t\right),\hspace{0.17em}\widehat{y}\left(t\right)\right)\in {\mathbb{R}}^d$$

(4)

where $\phi \left(t\right)$ is the feature attribution vector, ${\phi }_j(t)$ quantifies the relevance of the $j$-th feature to the prediction $\widehat{y}(t)$, and $\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{l}\mathrm{a}\mathrm{i}\mathrm{n}$ denotes any post hoc XAI method, e.g., SHAP [26], LIME [19], or Integrated Gradients. When $\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{l}\mathrm{a}\mathrm{i}\mathrm{n}$ is instantiated as SHAP, the attributions satisfy the Shapley efficiency axiom ($\sum {\phi }_j=\widehat{y}-\mathbb{E}[\widehat{y}]$), guaranteeing completeness. The framework does not prescribe a specific method; the pipeline-level requirement is that every prediction be accompanied by a feature-level explanation vector.

Mitigate provides structured risk management when predictions fall outside acceptable confidence levels or violate operational constraints. Mitigation strategies include deferring decisions to human supervisors, applying conservative fallback rules, triggering alternative decision paths, or initiating additional data validation. This stage operationalizes human-in-the-loop control under elevated-risk conditions.

The integration of automation with human-in-the-loop control has been formally studied in adjacent domains; for instance, Baldi et al. [27] demonstrate a switched self-tuning approach in building automation that dynamically adjusts control strategies based on occupancy and thermal conditions—a paradigm analogous to the mitigate stage’s conditional routing between automated and human-led responses.

$$m(t)=\{\begin{array}{ll}\mathrm{proceed}& \mathrm{if}v\left(t\right)=\mathrm{pass}\wedge \sigma \left(t\right)\le {\sigma }_{\mathrm{l}\mathrm{o}\mathrm{w}}\\ \mathrm{adjust}& \mathrm{if}v\left(t\right)=\mathrm{pass}\wedge \sigma \left(t\right)>{\sigma }_{\mathrm{l}\mathrm{o}\mathrm{w}}\\ \mathrm{escalate}& \mathrm{if}v\left(t\right)=\mathrm{flag}\\ \mathrm{halt}& \mathrm{if}v\left(t\right)=\mathrm{block}\end{array}$$

(5)

where $m\left(t\right)\in \{\mathrm{proceed},\hspace{0.17em}\mathrm{adjust},\hspace{0.17em}\mathrm{escalate},\hspace{0.17em}\mathrm{halt}\}$ is the mitigation decision, and ${\sigma }_{\mathrm{l}\mathrm{o}\mathrm{w}}$ is the uncertainty threshold below which fully automated execution is acceptable. Note that ${\sigma }_{\mathrm{l}\mathrm{o}\mathrm{w}}\le {\sigma }_{\mathrm{m}\mathrm{a}\mathrm{x}}$ by design: ${\sigma }_{\mathrm{l}\mathrm{o}\mathrm{w}}$ separates confident predictions (suitable for full automation) from uncertain-but-acceptable ones, whereas ${\sigma }_{\mathrm{m}\mathrm{a}\mathrm{x}}$ (Equation (3)) marks the boundary beyond which predictions are flagged as unreliable. While the mitigation decision $m\left(t\right)$ depends only on $v\left(t\right)$ and $\sigma \left(t\right)$, the explanation vector $\phi \left(t\right)$ is forwarded as contextual information when $m\left(t\right)\in \{\mathrm{adjust},\hspace{0.17em}\mathrm{escalate},\hspace{0.17em}\mathrm{halt}\}$, enabling the human operator to understand why the system recommends a particular action.

Execute is the final enactment of the decision, whether through automated control, semi-automated assistance, or human-led action. Execution is contingent on successful completion of all preceding stages, ensuring that real-time decisions remain compliant with safety, governance, and accountability requirements.

$$h(t)=\{\begin{array}{ll}\mathrm{automated}& \mathrm{if}m\left(t\right)=\mathrm{proceed}\\ \mathrm{semi}\text{-}\mathrm{automated}& \mathrm{if}m\left(t\right)=\mathrm{adjust}\\ \mathrm{human}\text{-}\mathrm{led}& \mathrm{if}m\left(t\right)\in \{\mathrm{escalate},\hspace{0.17em}\mathrm{halt}\}\end{array}$$

(6)

where $h\left(t\right)\in \{\mathrm{automated},\hspace{0.17em}\mathrm{semi}\text{-}\mathrm{automated},\hspace{0.17em}\mathrm{human}\text{-}\mathrm{led}\}$ is the execution mode, determined by the mitigation outcome $m\left(t\right)$, and $a\left(t\right)$ denotes the final system action under the selected mode.

Equations (1)–(6) should therefore be interpreted as formal abstractions of socio-technical workflow logic rather than as mathematical optimization or stability formulations. Their primary role is to improve conceptual transparency and provide a structured representation of decision sequencing within high-risk AI-enabled operational environments.

Together, these equations formalize a single decision cycle at time $t$. In continuous operation, execute closes the loop: the system action $a\left(t\right)$ alters the physical process, producing updated sensor readings $X(t+1)$ that re-enter the predict stage, thereby establishing the monitoring cycle depicted in Figure 1 (dashed feedback arrow).

The PRIME layer thus addresses the micro-level process gap by providing a structured, auditable operational workflow. It encompasses both technical and social elements: while predict and regulate are primarily technical, interpret, mitigate, and execute involve human judgment, organizational authority, and accountability, making PRIME socio-technical by design rather than by label.

3.3. The INSPECT Layer: Governance and Institutional Conditions

The INSPECT layer defines the organizational and institutional conditions under which the PRIME operational flow can be deployed safely, legitimately, and sustainably. The acronym stands for integrity of data, navigability of explanations, supervisory control, policy and governance maturity, ethical compliance, collaboration between management and technical experts, and trust calibration. The term “INSPECT” is deliberate: just as industrial inspection regimes provide systematic oversight of physical processes, this layer provides continuous institutional oversight of AI-driven decision processes, verifying that they remain safe, accountable, and aligned with organizational expectations. In this sense, the INSPECT layer can be interpreted as a set of governance constraints applied across the PRIME pipeline, ensuring that each operational stage remains aligned with institutional, regulatory, and ethical requirements.

Integrity refers to the quality, provenance, and reliability of data and models. It encompasses data governance practices, validation procedures, and mechanisms for detecting data drift or model degradation over time.

Navigability captures the cognitive accessibility of AI outputs and explanations to human stakeholders. It goes beyond formal explainability: an explanation that is technically correct but operationally opaque fails this criterion. Navigability ensures that outputs support real decision-making rather than abstract transparency.

Supervisory control formalizes human authority over AI-driven decisions through clearly defined intervention rights, override mechanisms, escalation protocols, and accountability structures, particularly under abnormal or high-risk conditions.

Policy and governance maturity reflects the existence and integration of organizational policies governing AI use, including risk management frameworks, compliance procedures, and continuous monitoring practices.

Ethical compliance ensures alignment with ethical principles, safety standards, and sector-specific regulatory requirements, with particular emphasis on harm prevention and fairness in high-stakes decision contexts.

Collaboration emphasizes coordinated interaction between top management, IT professionals, and domain experts to ensure that strategic objectives, technical capabilities, and operational realities remain aligned throughout the AI lifecycle.

Trust calibration is addressed in detail in Section 3.4 as a cross-layer coupling mechanism.

3.4. Trust Calibration as a Cross-Layer Coupling Mechanism

A distinctive feature of PRIME–INSPECT is the treatment of trust calibration not as a passive psychological outcome, but as an explicit design objective jointly shaped by both layers. Trust calibration is influenced simultaneously by micro-level decision characteristics—prediction uncertainty, explanation quality, and mitigation triggers—and by macro-level institutional conditions—governance maturity, supervisory authority, and organizational norms (Figure 1).

Appropriate trust emerges when operators can understand AI behavior (interpret and navigability), intervene when necessary (mitigate and supervisory control), and observe consistent alignment between AI actions, organizational policies, and safety outcomes (regulate, policy maturity, and ethical compliance). When these conditions are absent, two failure modes arise: automation bias, where operators over-rely on AI outputs despite uncertainty; and algorithm aversion, where operators reject AI recommendations following isolated errors. PRIME–INSPECT addresses both by embedding calibration mechanisms at every stage of the operational flow and anchoring them in institutional oversight. Thus, trust calibration is not treated as an emergent or purely psychological phenomenon, but as a controllable system characteristic shaped jointly by algorithmic behavior, explanation mechanisms, and governance structures.

A related question concerns the explainability of PRIME–INSPECT as a framework itself, rather than of the AI models it governs. PRIME–INSPECT achieves framework-level explainability through three properties: structural transparency, in that every decision stage and governance condition is explicitly named, defined, and functionally coupled; procedural auditability, in that Equations (1)–(6) provide a complete formal trace of how each input state X(t) is transformed into an action a(t) through identifiable, inspectable stages; and accountability assignment, in that each stage of the PRIME pipeline maps to specific INSPECT conditions that define who is responsible for oversight at that point. In this sense, PRIME–INSPECT is not a black box, its decision logic is fully visible, traceable, and attributable, which is a prerequisite for organizational trust in any governance framework.

To illustrate the operational implications of trust calibration, it is useful to consider a simplified feedback loop scenario reflecting real-time decision-making conditions in industrial environments. Consider a typical decision cycle in which the AI system predicts an elevated risk of failure in a monitored component and assigns a confidence level to the prediction. The system simultaneously provides an interpretable explanation (e.g., feature attribution or anomaly visualization) to support operator understanding.

Upon receiving the alert, the human operator may either accept the recommendation, adjust it, or override it based on contextual knowledge, field inspection, or additional signals not captured by the model. The selected action, along with the system prediction and contextual conditions, is recorded in an audit log.

Following execution, the actual outcome (e.g., whether a failure occurred, was prevented, or did not materialize) is compared with the initial prediction. If the prediction proves accurate, the system’s credibility is reinforced, supporting confidence reinforcement and supervised threshold adjustment where appropriate. Conversely, if the prediction results in a false positive or false negative, calibration mechanisms are triggered to adjust model thresholds, refine explanation outputs, or update operator guidance.

In cases where systematic patterns emerge—such as frequent human overrides of correct system recommendations or repeated false alarms—the system may initiate a structured review process, including model retraining, feature reassessment, or adjustment of decision policies. In this way, trust calibration operates as a continuous feedback loop linking prediction performance, human judgment, and governance oversight.

Table 1. Illustrative trust calibration feedback loop.

Trigger	Human Action	Audit Outcome	Calibration Response
High-risk prediction with high confidence	Accept recommendation	Failure prevented	Reinforce confidence; consider threshold adjustment subject to supervisory review
High-risk prediction (false alarm)	Override decision	No failure occurs	Adjust thresholds, refine model sensitivity
Low-risk prediction but failure occurs	No intervention	Failure detected	Increase model sensitivity, update features
Frequent overrides of correct alerts	Override despite accurate prediction	Pattern detected in logs	Initiate review, retraining, or policy adjustment

summarizes representative trust calibration feedback scenarios and corresponding calibration responses within the PRIME–INSPECT framework.

4. Empirical Grounding of the PRIME–INSPECT Framework

4.1. Research Design and Objectives

The empirical component of this study is designed to provide empirical grounding for the PRIME–INSPECT framework rather than to test formal causal hypotheses. Its primary objective is to assess the relevance, internal coherence, and practical recognizability of the core dimensions underlying the proposed socio-technical architecture in real organizational settings where AI-enabled real-time decision-making is emerging.

Specifically, the empirical analysis examines how key organizational stakeholders perceive and evaluate dimensions related to trust, explainability, human oversight, governance maturity, risk, and accountability—constructs that are theoretically central to PRIME–INSPECT. The purpose of this assessment is not to validate the framework statistically, but to demonstrate that its components reflect meaningful and observable concerns in practice.

Accordingly, the research design does not aim to test causal relationships between variables, nor does it employ hypothesis-driven modeling techniques. Instead, it adopts an exploratory and descriptive approach that supports the conceptual development of the framework by illustrating how its dimensions manifest across different organizational roles.

A dual-sample design was adopted to reflect the two-layer structure of PRIME–INSPECT. Responses from Top Management Team (TMT) members capture strategic, governance-oriented, and institutional perspectives aligned with the INSPECT layer. In contrast, responses from IT and technical managers reflect operational, system-level, and implementation-oriented perspectives aligned with the PRIME layer. This design enables the analysis to explore complementarities and potential misalignments between strategic intent and operational execution, which are central to understanding the deployment of trustworthy AI in high-risk industrial contexts.

The empirical component is intentionally exploratory in nature. It does not aim to validate the framework through predictive benchmarking, causal inference, or experimental comparison. Instead, it is designed to examine whether the socio-technical dimensions embedded within PRIME–INSPECT correspond to recognizable and meaningful organizational concerns among stakeholders involved in AI-enabled decision-making.

4.2. Survey Instruments and Data Collection

Two structured survey instruments were developed and administered to capture the perspectives of distinct organizational stakeholder groups involved in AI-enabled decision-making. Both instruments were designed to align with the conceptual dimensions of the PRIME–INSPECT framework while remaining accessible to respondents with different roles and responsibilities. To improve structural coherence, the two instruments are presented here as complementary components of a single measurement design rather than as isolated subsections.

All survey items were measured using a four-point Likert-type scale ranging from strong disagreement to strong agreement. The use of an even-numbered scale was intentional, aiming to reduce neutral or ambivalent responses and to encourage respondents to express directional perceptions. Likert-type items were treated as approximately interval-scaled for the purpose of descriptive analysis, a practice commonly adopted in exploratory and organizational research [28]. The four-point scale was particularly suitable for the study context, as neutral positions may obscure underlying perceptions regarding trust, governance, and accountability in AI-enabled decision-making.

From the strategic and institutional perspective, the TMT instrument targeted senior managers and executives involved in governance, accountability, and organizational oversight. It focused on perceptions of AI governance and policy maturity, organizational risk associated with AI-driven decisions, trust in AI-supported decision-making, responsibility allocation, and ethical and regulatory considerations. This design allowed respondents to assess organizational readiness and governance capacity without requiring detailed technical knowledge of AI systems.

From the operational and technical perspective, the IT instrument targeted managers and specialists responsible for the design, implementation, operation, or supervision of AI-enabled systems. It captured perceptions related to AI operationalization and system reliability, the explainability and interpretability of AI outputs, the availability and effectiveness of human oversight mechanisms, intervention and escalation points in AI-driven workflows, and confidence in real-time data and model performance. In this way, the IT survey directly reflects the operational logic of the PRIME layer, capturing how AI-supported decisions are generated, interpreted, constrained, and acted upon in practice.

4.3. Sample Characteristics

The empirical analysis draws on responses from two distinct samples corresponding to the TMT and IT survey instruments. The samples are analyzed separately and comparatively, reflecting their different organizational roles and analytical relevance. Respondents operate within industrial and organizational contexts where AI-enabled analytics, automation, or decision-support systems are present or under active consideration, with experience ranging from early-stage experimentation to more advanced operational use.

The empirical dataset consists of two respondent groups aligned with the dual-layer architecture of the PRIME–INSPECT framework. The TMT sample included 85 senior managers and executives responsible for governance, oversight, strategic planning, and organizational decision-making. The IT sample included 151 technical managers, IT specialists, engineers, and system professionals involved in AI-enabled operational systems and digital infrastructure.

Respondents were recruited through professional networks and organizational outreach across sectors in which AI-supported analytics, automation, or decision-support technologies were already implemented or under consideration. The participant profiles varied in organizational role, years of experience, level of AI familiarity, and degree of involvement in technology-driven decision-making environments.

The TMT sample was dominated by respondents from technology, finance, telecommunications, and retail sectors, with organizations ranging from small enterprises to large corporations employing more than 1000 people. Most TMT respondents reported moderate-to-high familiarity with AI-related concepts, reflecting strategic exposure rather than purely technical expertise.

The IT sample consisted primarily of professionals with substantial technical experience. A large proportion of respondents reported more than ten years of experience in IT-related roles, while most participants indicated familiarity with AI technologies such as machine learning, analytics platforms, natural language processing, and automation tools. Approximately two-thirds of IT respondents had previously participated in AI-related projects or organizational initiatives.

Table 2. Sample characteristics of TMT and IT respondents.

Variable	TMT Sample	IT Sample
Sample size (N)	85	151
Primary respondent profile	Senior managers, directors, executives	IT managers, engineers, specialists
Dominant sectors	Technology, Finance, Telecommunications, Retail	Cross-sector technical roles
Organizational size	1–50 to >1000 employees	Mixed organizational environments
Years of professional experience	Senior managerial experience	Majority >10 years
AI familiarity	Mostly moderate to high	Mostly moderate to advanced
Exposure to AI initiatives	Strategic/governance perspective	Operational/implementation perspective
Participation in AI-related projects	Indirect or governance-level	~65% directly involved

summarizes the demographic and organizational characteristics of both respondent groups.

4.4. Operationalization of PRIME–INSPECT Dimensions

To enable empirical grounding of the PRIME–INSPECT framework, the theoretical dimensions introduced in Section 3 were operationalized through constructs captured by the two survey instruments. Rather than translating the framework into a set of testable causal hypotheses, this operationalization establishes a transparent and conceptually consistent link between abstract framework components and observable stakeholder perceptions.

Table 3. Mapping of survey constructs to PRIME–INSPECT framework dimensions.

PRIME–INSPECT Component	Layer	Survey Source	Indicative Survey Constructs
Predict	PRIME (Operational)	IT survey	Perceived reliability of AI predictions; confidence in model outputs; adequacy of real-time data inputs
Regulate	PRIME (Operational)	IT survey	Awareness of operational constraints; perceived adequacy of rule-based limits and safety thresholds
Interpret	PRIME (Operational)	IT survey	Importance of explainability for decision-making; clarity of AI-generated recommendations; usefulness of explanations
Mitigate	PRIME (Operational)	IT survey	Availability of human intervention points; perceived effectiveness of override and escalation mechanisms
Execute	PRIME (Operational)	IT survey	Scope of automation in decision execution; balance between automated and human-led actions
Integrity	INSPECT (Governance)	TMT survey	Data quality assurance; trust in data sources; confidence in model robustness
Navigability (XAI)	INSPECT (Governance)	IT survey	Cognitive accessibility of explanations; alignment between explanations and operational needs
Supervisory Control	INSPECT (Governance)	Both	Clarity of control rights; responsibility allocation; perceived ability to intervene in AI-driven decisions
Policy and Governance Maturity	INSPECT (Governance)	TMT survey	Existence of formal AI policies; clarity of accountability; risk management practices
Ethical Compliance	INSPECT (Governance)	TMT survey	Perceived alignment with safety standards, ethical responsibility, and regulatory compliance
Collaboration	INSPECT (Governance)	Both	Quality of coordination between management and IT; shared understanding of AI risks and capabilities
Trust Calibration	Cross-layer	Both	Appropriate reliance on AI outputs; avoidance of blind trust or algorithm aversion; confidence adjusted to context and risk

presents the mapping between the survey constructs and the corresponding PRIME–INSPECT components. This mapping serves two purposes. First, it demonstrates that the proposed framework is grounded in empirically observable dimensions that are meaningful to key organizational stakeholders. Second, it provides a structured basis for organizing the descriptive analysis of survey results, ensuring that empirical findings are interpreted in a manner consistent with the framework’s conceptual architecture.

Importantly, the mapping does not imply directional causality or statistical dependency between components. Instead, it reflects functional correspondence between perceived organizational and operational characteristics and the conceptual building blocks of PRIME–INSPECT. The mapped dimensions provide the analytical structure for the descriptive results that follow, making explicit how PRIME-related operational perceptions and INSPECT-related governance perceptions jointly ground the framework in stakeholder experience.

The following subsections present descriptive results for the major segments of the PRIME–INSPECT framework in turn. Rather than testing causal relationships, the analysis highlights whether stakeholder perceptions recognize and support the conceptual distinction between the operational PRIME layer and the governance-oriented INSPECT layer.

4.5. Descriptive Results Mapped to PRIME–INSPECT

4.5.1. PRIME-Related Operational Perceptions (IT Sample)

The IT and technical management survey provides insight into how operational stakeholders perceive AI-enabled real-time decision-making processes within the PRIME layer of the framework. Overall, the results indicate a moderately positive orientation toward the use of AI in operational contexts, accompanied by non-negligible variation in perceptions across respondents. Given the exploratory purpose of the empirical component, the analysis relies on descriptive statistics (means and standard deviations) rather than inferential techniques. This approach is consistent with prior research that empirically grounds conceptual or socio-technical frameworks, in which the objective is to assess the salience, coherence, and perceived relevance of theoretical dimensions rather than to test causal relationships or population-level effects. Standard deviations are reported to reflect the degree of agreement and heterogeneity in stakeholder perceptions, supporting nuanced interpretation without implying statistical significance.

The use of descriptive statistics is consistent with the exploratory purpose of the study. Because the empirical component is intended to provide conceptual grounding rather than hypothesis testing, the emphasis is placed on identifying recognizable perception patterns across stakeholder groups rather than establishing causal inference. Descriptive analysis therefore functions as a bridge between abstract framework dimensions and observable organizational perspectives.

Among the highest-rated items, respondents strongly emphasized the importance of IT involvement in the strategic planning of AI initiatives (mean = 3.73, SD = 0.52). While this result does not indicate the extent to which such involvement is currently realized in practice, it highlights a shared perception among IT stakeholders that strategic participation is a critical prerequisite for effective and trustworthy AI deployment.

A generally positive stance toward AI adoption is further reflected in respondents’ attitudes toward the use of AI tools in IT processes (mean = 3.51, SD = 0.62) and their overall positive orientation toward AI-enabled work practices (mean = 3.54, SD = 0.66). These findings suggest broad acceptance of AI as an operational support mechanism, while the associated standard deviations indicate some heterogeneity in experience and readiness across organizations.

With respect to decision speed and performance, respondents reported that AI-supported decisions are made faster than those relying solely on traditional approaches (mean = 3.44, SD = 0.70) and that AI contributes to improved business outcomes (mean = 3.36, SD = 0.73). Similarly, perceptions of operational efficiency gains attributable to AI—such as automation and enhanced monitoring—were rated moderately positively (mean = 3.37, SD = 0.74). Together, these results align with the predict and execute stages of the PRIME flow, indicating perceived operational benefits without implying uniform or guaranteed performance improvements.

Beyond efficiency considerations, several items relate to interpretability, oversight, and responsible use. Respondents noted that IT leadership support for AI development and deployment is present but not uniform (mean = 3.32, SD = 0.76) and that data protection and security considerations are accounted for during AI system development (mean = 3.31, SD = 0.80). The relatively higher dispersion associated with these items suggests varying levels of organizational maturity with respect to explainability, governance integration, and risk-aware operational practices.

To assess whether the survey scales formed internally coherent measures of the intended PRIME–INSPECT dimensions, internal consistency was evaluated using Cronbach’s Alpha (α). High α values indicate that respondents who rated one item in a construct highly tended to rate other items in the same construct similarly—confirming that each construct was perceived as a meaningful and distinct organizational concern rather than a collection of unrelated items. All constructs demonstrated high internal consistency, with α values ranging from 0.82 to 0.93, exceeding the recommended threshold of 0.70 [29]. In the IT sample, operational performance (α = 0.82), oversight and control (α = 0.86), explainability (α = 0.86), and governance alignment (α = 0.85) all formed reliable scales. In the TMT sample, strategic value (α = 0.84), governance maturity (α = 0.93), and ethics and compliance (α = 0.83) demonstrated comparable consistency. Inter-construct correlations further supported the coherence of the framework: operational performance and oversight and control were strongly positively correlated (r = 0.68, p < 0.01), while explainability showed a moderate negative correlation with other operational dimensions (r ≈ −0.25, p < 0.05), a pattern interpreted in Section 5.4 as reflecting heightened critical awareness among technically experienced respondents rather than a simple performance–interpretability trade-off. In the TMT sample, strategic value and governance maturity were strongly correlated (r = 0.69, p < 0.01), indicating that executives view governance as an enabler of value creation.

Taken together, the IT sample results indicate that the PRIME operational stages—particularly prediction, execution, and efficiency-oriented use—are broadly recognized as valuable. At the same time, elements related to interpretation, mitigation, and structured oversight exhibit greater variability. This pattern mirrors the internal structure of the PRIME layer itself: technical optimism is strongest where AI delivers visible operational value, but weaker where explainability, escalation logic, and human intervention must be institutionalized in day-to-day practice.

4.5.2. INSPECT-Related Governance Perceptions (TMT Sample)

The Top Management Team (TMT) survey captures governance-oriented and strategic perceptions related to the INSPECT layer of the PRIME–INSPECT framework. Overall, the results suggest a generally optimistic but cautious managerial orientation toward AI-enabled decision-making, accompanied by notable variation in perceived governance maturity and organizational preparedness.

The highest-rated item in the TMT sample relates to the importance of assessing and mitigating bias in AI systems (mean = 3.45, SD = 0.68). This finding indicates a strong managerial awareness of ethical and fairness-related risks associated with AI-driven decisions. However, the related dispersion suggests that while concern is widespread, perceptions of organizational capability to address such risks vary across respondents.

A broadly positive stance toward AI adoption is further reflected in respondents’ evaluations of AI use in managerial work as a desirable practice (mean = 3.40, SD = 0.80) and their general positive attitude toward the use of AI tools in decision-making (mean = 3.38, SD = 0.74). In addition, respondents reported a moderate intention to actively support the expanded use of AI within their organizations over the coming period (mean = 3.35, SD = 0.81). Together, these findings suggest openness to AI integration at the strategic level, while also indicating heterogeneity in commitment and readiness.

Governance and policy-related perceptions exhibit similar patterns. Respondents moderately agreed that ethical, safety, and responsibility considerations are systematically addressed during the development and deployment of AI solutions (mean = 3.34, SD = 0.85). The relatively high standard deviation for this item indicates uneven institutionalization of governance practices, with some organizations reporting more structured approaches than others.

With respect to performance and decision speed, TMT respondents perceived that AI has the potential to enhance organizational competitiveness (mean = 3.33, SD = 0.79) and that AI-supported decision-making reduces the time required for key managerial decisions (mean = 3.27, SD = 0.90). Perceptions of actual time savings achieved through AI implementation were somewhat more modest (mean = 3.21, SD = 0.87), suggesting that expected benefits may not yet be fully realized across all organizational contexts.

Taken together, the TMT results indicate that INSPECT-related dimensions—particularly ethical awareness, strategic openness, and perceived competitive potential—are widely recognized, while governance maturity and consistent operationalization remain uneven. This pattern reinforces the logic of the INSPECT layer: trustworthiness at the strategic level depends not only on favorable attitudes toward AI, but also on the presence of formalized governance mechanisms, supervisory control, and clear accountability structures.

4.6. Alignment and Divergence Between TMT and IT Perspectives

Comparing the perceptions of IT and Top Management Team (TMT) respondents reveals both areas of alignment and notable divergence in how AI-enabled real-time decision-making is understood and evaluated across organizational levels. These patterns provide empirical insight into the socio-technical dynamics that PRIME–INSPECT is designed to address.

4.6.1. Areas of Alignment

Across both samples, respondents express a broadly positive orientation toward the use of AI in organizational decision-making. IT respondents emphasize the perceived operational benefits of AI in terms of efficiency, decision speed, and support for daily work processes, while TMT respondents highlight AI’s strategic potential, including competitiveness and improved managerial decision-making. This shared recognition of AI’s value suggests a common baseline of acceptance across organizational levels.

Both groups also converge on the importance of responsible AI deployment. IT respondents report awareness of data protection and security considerations during system development, while TMT respondents strongly emphasize the need to assess and mitigate bias and ethical risks. Although these concerns are articulated from different perspectives, together they point to a shared understanding that AI-enabled decisions require safeguards beyond technical performance alone.

4.6.2. Areas of Divergence

Despite this general alignment, several divergences emerge that are particularly relevant for the PRIME–INSPECT framework. IT respondents place strong emphasis on the importance of operational involvement in AI initiatives—most notably the perceived need for IT participation in strategic AI planning. In contrast, TMT responses focus more heavily on high-level intentions, ethical considerations, and anticipated benefits, with less consistent evidence of fully institutionalized governance practices.

This divergence suggests a potential gap between operational readiness and institutional maturity. While IT stakeholders appear prepared to engage with AI systems at the operational level, including considerations of explainability, intervention, and system reliability, managerial perceptions indicate variability in the formalization of policies, accountability structures, and governance mechanisms required to support these operational practices at scale.

Differences also emerge in perceptions of realized versus expected benefits. IT respondents report moderate operational gains associated with AI use, whereas TMT respondents tend to emphasize AI’s potential impact on competitiveness and decision speed. This contrast indicates that strategic expectations may, in some cases, outpace current operational implementation.

4.6.3. Implications for Trust Calibration and Socio-Technical Integration

The observed alignment and divergence patterns have direct implications for trust calibration. High-level managerial optimism, when not matched by consistent governance structures and operational feedback, may increase the risk of over-reliance on AI systems. Conversely, strong operational engagement without clear strategic guidance and accountability may lead to fragmented or ad hoc deployment practices.

These findings empirically support the central premise of the PRIME–INSPECT framework: trustworthy AI-driven real-time decision-making requires coordinated integration between operational decision flows (PRIME) and institutional governance conditions (INSPECT). Trust calibration emerges not as an abstract psychological construct, but as an organizational outcome shaped by the interaction between strategic intent, governance maturity, and operational control.

4.7. Illustrative Application in Metallurgy

The purpose of this illustrative application is not to present post hoc engineering validation or a performance benchmark based on a specific proprietary dataset. Rather, it serves to demonstrate how the PRIME–INSPECT framework can be instantiated within a representative high-risk process environment using industry-standard process parameters, typical operating variables, and safety constraints reported in metallurgical practice.

To provide additional grounding for the illustrative scenario, it is useful to highlight typical process parameters commonly monitored in blast furnace operations. In the context of tuyere condition monitoring and predictive maintenance, representative variables include cooling-water outlet temperature, inlet–outlet temperature differential, cooling-water flow rate, and pressure drop across the cooling system. Additional relevant indicators may include hot blast temperature and pressure, as well as localized thermal anomalies detected through infrared or embedded sensor measurements.

These parameters are widely used in metallurgical practice to assess thermal stress, cooling efficiency, and early signs of structural degradation in tuyere systems. In this study, they serve as illustrative examples of the types of signals that would feed into the PRIME decision workflow, rather than as inputs derived from a specific proprietary dataset. These variables are not used as empirical inputs in this study; rather, they are included to demonstrate how the PRIME decision logic could be instantiated using representative process signals commonly monitored in blast furnace operations.

To further ground the framework, the application focuses on real-time decision-making in blast furnace operations, where continuous processes, high energy intensity, and strict safety constraints require timely and reliable decisions based on SCADA and industrial IoT data streams. By using a scenario grounded in the physical realities of blast furnace operations, such as predictive maintenance for cooling nozzles (tuyeres), this section functions as a structured conceptual demonstration. Its goal is to translate the framework’s decision logic into a plausible operational context, showing how socio-technical governance mechanisms and formal decision stages (Equations (1)–(6)) interact with the safety and reliability requirements of Industry 4.0 systems.

Consider the specific example of predictive maintenance for blast furnace tuyeres (cooling nozzles), as illustrated in Figure 2. In a representative setup, the monitoring stream may include multiple thermocouple or thermographic measurement points per tuyere zone, cooling-water inlet and outlet temperatures, pressure and flow indicators, and contextual process variables such as hot-blast pressure and oxygen enrichment. These signals are typically sampled at sub-minute frequency and aggregated into short rolling windows for anomaly detection and Remaining Useful Life (RUL) estimation. Tuyere failure due to burn-through poses a critical safety risk because water ingress into molten iron may trigger severe incidents and therefore requires immediate, precise action.

The diagram illustrates how institutional safeguards (INSPECT, bottom layer) directly support specific operational decision stages (PRIME, top layer), with trust calibration serving as a feedback mechanism based on audit logs and near-miss review.

• PRIME Layer (Operational Flow):

• Predict: As shown in Figure 2, the AI system forecasts the Remaining Useful Life (RUL) of specific tuyeres based on temporal thermal patterns and correlated process variables, rather than merely flagging the current temperature level.
• Regulate: The prediction is constrained by physics-based and safety rules. For example, implausible temperature jumps, conflicting sensor readings, or deviations exceeding predefined process-safety margins are filtered before any recommendation is escalated into action, consistent with the regulation logic formalized in Equation (3).
• Interpret: The system provides operators with a visual heat map or ranked feature explanation that contextualizes the alert and clarifies whether the risk is associated with a localized hotspot, cooling instability, or a broader overheating trend, in line with the interpretability layer formalized in Equation (4).
• Mitigate: Recognizing the high-risk context, the system can trigger a controlled mitigation protocol, such as reducing blast intensity, increasing inspection priority, or switching to a more conservative operating mode, while awaiting human confirmation, consistent with the mitigation logic formalized in Equation (5).
• Execute: The final execution step—such as scheduling replacement, temporarily reducing load, or ordering shutdown of the affected unit—remains a human-led or explicitly authorized action, supported by the system’s preparatory adjustments and logged recommendations, consistent with the execution mapping formalized in Equation (6).

• INSPECT Layer (Governance and Safeguards):

• Integrity and Navigability: This requires that training data include historically verified burn-through events or near-failures, that sensor calibration routines are documented, and that alerts are displayed in formats familiar to operators and furnace supervisors.
• Supervisory Control: As depicted in the diagram, the Shift Manager or responsible process engineer retains explicit override authority to dismiss or defer a recommendation if field inspection, auxiliary measurements, or contextual production constraints contradict the model output.
• Trust Calibration: A feedback loop ensures that both true alarms and near-miss predictions are reviewed against subsequent outcomes. This dynamic calibration helps to prevent automation bias and supports an appropriate balance between reliance and verification under real operating conditions.

Ultimately, this application illustrates that effective AI deployment in metallurgy requires more than high predictive accuracy alone; it demands a coupled socio-technical architecture. By integrating the PRIME layer’s operational speed with the INSPECT layer’s institutional guardrails, the framework translates abstract governance principles into concrete operational safeguards. In this sense, the industrial example is retained not as a substitute for future validation on proprietary plant logs, but as a structured demonstration of how trustworthy AI-enabled real-time decision-making can be operationalized in high-risk process environments.

Future work should evaluate the framework using longitudinal industrial datasets, including real-time monitoring logs, anomaly detection histories, intervention records, and safety outcomes, to assess how PRIME–INSPECT performs under real operational conditions.

5. Discussion

This section synthesizes the theoretical foundations reviewed in Section 2 with the empirical findings presented in Section 4 to articulate the contributions, implications, and limitations of the PRIME–INSPECT framework. The contribution of PRIME–INSPECT lies not in proposing a new predictive algorithm or mathematical control system, but in integrating fragmented research streams into a unified socio-technical architecture. Its value emerges from linking operational AI decision logic with institutional safeguards, thereby addressing an integration gap that remains underdeveloped in the existing Industry 4.0 governance literature. Rather than treating the empirical results as confirmatory evidence, the discussion interprets them as illustrative signals that inform the relevance, coherence, and practical plausibility of the proposed socio-technical architecture for AI-enabled real-time decision-making in high-risk industrial systems.

5.1. Positioning PRIME–INSPECT Within the Existing Literature

The literature on AI adoption, trust, explainability, human oversight, and governance has developed along essentially parallel streams. Adoption-oriented models such as TAM and UTAUT explain behavioral intention and technology acceptance but provide limited guidance on how AI-driven decisions should be governed once embedded in organizational processes, particularly under real-time constraints. Trust and reliance research highlights risks of algorithm aversion and automation bias, yet it often treats trust as a psychological outcome rather than as an organizational design variable. Explainable AI research demonstrates the importance of transparency, but frequently abstracts explanations from the operational and institutional contexts in which they are used. Governance and risk management frameworks articulate high-level principles and controls, but rarely specify how these mechanisms interact with operational decision flows.

PRIME–INSPECT responds to these fragmentation patterns by integrating insights across this literature into a unified socio-technical framework. The PRIME layer formalizes the micro-level operational flow of AI-supported real-time decisions. In contrast, the INSPECT layer captures the macro-level institutional conditions under which such decisions can be deployed responsibly. In doing so, the framework moves beyond acceptance- or principle-based perspectives and explicitly links operational execution, human oversight, explainability, governance maturity, and trust calibration within a single architecture.

As shown in

Table 4. Positioning PRIME–INSPECT relative to existing AI governance instruments.

Dimension	NIST AI RMF [25]	EU AI Act [30]	ISO/IEC 42001 [31]	OECD AI [32]	PRIME–INSPECT
Operational pipeline	Not specified	Not specified	Not specified	Not specified	Equations (1)–(6)
Risk classification	MAP function	Article 6	Clause 6.1 planning	Not specified	Equation (3)
Explainability req.	Narrative	Article 13	Annex A controls	Transparency and explainability principles	Equation (4)
Human oversight	Narrative	Article 14	Narrative	Human-centered values and human oversight	Equation (5)
Trust calibration	Not addressed	Not addressed	Not addressed	Not addressed	Section 3.4
Real-time decision loop	Not specified	Not specified	Not specified	Not specified	Equation (1)

, existing AI governance instruments provide essential normative and regulatory foundations but do not specify how these principles translate into operational decision flows under real-time constraints. The NIST AI RMF addresses risk governance through its MAP, MEASURE, and MANAGE functions, yet it leaves the internal structure of AI decision pipelines unspecified. The EU AI Act imposes transparency and human oversight requirements through Articles 13 and 14, but defines these as compliance obligations rather than as operational design constraints integrated into the decision cycle. ISO/IEC 42001 provides a management system standard with Annex A control objectives related to explainability, risk management, and organizational governance, yet these remain at the process level without formalizing how predictions, constraints, and interventions interact in real time. The OECD AI Recommendation articulates principles of transparency and accountability but does not provide operational or technical specifications.

PRIME–INSPECT complements these instruments by providing the operational architecture that bridges the gap between governance requirements and decision execution. Its two distinctive contributions—formal trust calibration as a cross-layer design objective (Section 3.4) and the PRIME pipeline formalization (Equations (1)–(6))—address dimensions that remain unspecified across all four frameworks. Notably, while existing instruments address transparency, oversight, and risk governance, they do not explicitly operationalize trust calibration as a real-time feedback mechanism that links prediction uncertainty, human intervention, audit outcomes, and subsequent adjustment of decision thresholds within a single auditable decision pipeline.

In addition to comparisons with regulatory and governance frameworks, it is important to position PRIME–INSPECT relative to relevant academic models addressing AI adoption and human–AI trust.

From the adoption perspective, the Integrated AI Acceptance–Avoidance Model (IAAAM) explains managerial intentions to adopt or avoid AI systems by integrating enabling factors such as perceived usefulness and facilitating conditions with inhibiting factors including perceived risk and threat [6]. While this model provides a comprehensive behavioral explanation of AI adoption, it primarily operates at the level of perception and intention, without specifying how AI-driven decisions are operationalized, controlled, and executed in real-time environments.

From the trust perspective, recent research conceptualizes trust as a dynamic and evolving process rather than a static property. The CHAI-T framework models trust in human–AI collaboration as a function of antecedents, interaction processes, and outcomes, emphasizing continuous feedback and adaptation over time [33]. Similarly, relational approaches to trust in automation highlight that trust emerges through ongoing interaction between human and AI agents, shaped by system responsiveness, context, and feedback mechanisms [34]. Additional studies further emphasize the importance of trust calibration, where human reliance must be aligned with system capabilities through explainability and feedback mechanisms [15].

However, despite these advances, existing models primarily address either adoption behavior or trust dynamics in isolation. They do not provide a structured integration of trust calibration mechanisms with operational decision logic, explainability constraints, and institutional governance requirements in high-risk, real-time environments.

In contrast, PRIME–INSPECT extends these approaches by integrating behavioral (adoption), relational (trust), and operational (decision execution) dimensions into a unified socio-technical architecture. It explicitly embeds trust calibration within the decision pipeline, linking predictive outputs, explainability, human oversight, and governance controls across real-time decision stages.

The empirical findings support this integrative positioning. IT respondents emphasize operational efficiency, prediction reliability, and the importance of technical involvement, while TMT respondents highlight ethical concerns, strategic intent, and governance considerations. The observed alignment and divergence between these perspectives precisely reflect the coordination challenges identified in the literature and underscore the need for an explicit framework that connects the operational and institutional dimensions rather than treating them in isolation. Furthermore, the metallurgical application presented in Section 4.7 demonstrates that this integration is not merely theoretical; it provides a necessary functional architecture for managing physical risks, such as equipment burn-through, which cannot be addressed by algorithmic accuracy alone.

Table 5. Comparison of PRIME–INSPECT with selected academic frameworks.

Dimension	IAAAM [6]	CHAI-T [33]	Responsivity/Relational Trust [34]	PRIME–INSPECT
Primary focus	AI adoption (acceptance vs. avoidance)	Dynamic trust management in human–AI collaboration	Trust formation through interaction and system behavior	Integrated socio-technical architecture for AI-driven decision-making
Analytical level	Individual/organizational perception	Interaction process (human–AI collaboration)	Behavioral and relational dynamics	Multi-level (operational + governance)
Operational decision pipeline	Not specified	Not specified	Not specified	Structurally defined decision workflow
Real-time decision support	Not addressed	Partially addressed	Context-dependent	Explicitly supported
Trust conceptualization	Implicit (as adoption factor)	Dynamic and process-based	Emergent and interaction-based	Explicitly formalized as calibration mechanism
Trust calibration mechanisms	Not specified	Feedback-based trust adaptation	Experience-based trust adjustment	Embedded cross-layer calibration (Section 3.4)
Explainability integration	Not central	Indirect (via interaction)	Indirect	Explicitly integrated within the decision workflow
Human oversight	Not specified	Implicit (collaboration)	Emphasized	Explicitly embedded within operational decision stages
Governance integration	Not specified	Not specified	Not specified	Explicit (INSPECT layer)
Applicability to high-risk real-time systems	Limited	Partial	Partial	Illustratively demonstrated in high-risk industrial contexts

summarizes the positioning of PRIME–INSPECT relative to selected academic frameworks addressing AI adoption, trust, and human–AI collaboration.

5.2. Theoretical Implications

From a theoretical perspective, PRIME–INSPECT contributes to the literature in three key ways.

First, it advances the conceptualization of AI-enabled real-time decision-making by explicitly modeling the end-to-end operational decision flow. The PRIME sequence—predict, regulate, interpret, mitigate, execute—provides a structured representation of how AI outputs are transformed into organizational actions under time pressure, extending prior work that typically focuses on isolated stages such as prediction accuracy or explanation quality. Crucially, the inclusion of the “regulate” stage formally integrates domain-specific constraints—such as physicochemical laws in metallurgy—directly into the AI decision loop, bridging the gap between data-driven machine learning and engineering-based process control.

Second, the framework reframes trust calibration as a socio-technical design outcome rather than a passive psychological state. By embedding trust calibration across both operational (PRIME) and institutional (INSPECT) layers, PRIME–INSPECT aligns with emerging views that appropriate reliance on AI depends on transparency, controllability, accountability, and governance maturity. This perspective extends trust research by positioning calibration as an organizational capability that can be designed, monitored, and adjusted.

Third, PRIME–INSPECT contributes to governance theory by operationalizing abstract governance principles within a concrete decision architecture. Rather than treating governance as an external constraint, the INSPECT layer integrates governance mechanisms—such as supervisory control, ethical compliance, and policy maturity—directly into the conditions that enable or constrain operational AI decision-making. This integration addresses a gap in the literature between high-level governance frameworks and the day-to-day use of AI systems.

5.3. Practical Implications

For practitioners, the PRIME–INSPECT framework offers a diagnostic and design tool for deploying AI in high-risk industrial contexts.

At the operational level, PRIME highlights that trustworthy AI deployment requires more than accurate models. Organizations must explicitly define regulation mechanisms, interpretation practices, intervention points, and execution boundaries. The empirical findings suggest that while operational stakeholders recognize these needs, their implementation remains uneven, underscoring the importance of a structured design approach over ad hoc practices. As illustrated by the blast furnace case study, defining clear “supervisory control” protocols (e.g., override authority during thermal anomalies) is a prerequisite for preventing automation bias and ensuring operator safety in critical scenarios.

At the managerial and governance levels, INSPECT provides a checklist of institutional enablers that must accompany the use of operational AI. The variability observed in TMT perceptions regarding governance maturity and ethical practices indicates that strategic intent alone is insufficient. Clear accountability structures, formalized policies, and cross-functional collaboration mechanisms are required to translate AI’s potential into reliable, controllable decision systems.

Importantly, PRIME–INSPECT encourages organizations to address the alignment between IT and management perspectives explicitly. Misalignment such as high strategic optimism without corresponding operational safeguards or strong technical engagement without governance clarity may increase organizational risk. The framework thus supports more informed dialog between technical and managerial stakeholders, particularly in safety- and regulation-sensitive industries such as mining and metallurgy.

5.4. Limitations

Several limitations of this study should be acknowledged to contextualize the findings and define the scope of the contributions.

First, the empirical component relies on a descriptive analysis of survey data and does not aim to test causal relationships between variables. While this approach is appropriate for the exploratory objective of grounding the framework’s dimensions in observable stakeholder perceptions, it limits the ability to infer the strength or directionality of statistical associations between constructs.

Second, the survey samples were drawn from a cross-section of organizations and are not statistically representative of the mining or metallurgy sectors specifically. Consequently, the empirical findings should be interpreted as validating the theoretical coherence and relevance of the PRIME–INSPECT constructs among key decision-makers, rather than as a generalizable behavioral analysis of the metallurgical industry.

Third, the reliance on self-reported perceptions introduces potential biases related to social desirability or optimism, particularly among TMT respondents regarding governance maturity. Although these perceptions are central to understanding the “human factor” in trust and governance, they may not fully reflect actual operational practices or objective system performance.

Finally, while the application to blast furnace predictive maintenance (Section 4.7) demonstrates the framework’s logical coherence and domain suitability, it represents a conceptual instantiation based on representative industrial processes rather than a longitudinal field study. Future research is required to validate the framework’s efficacy using real-time operational data from active metallurgical plants to assess its impact on long-term safety and efficiency metrics.

Additionally, the negative correlation observed between the explainability construct and other operational dimensions in the IT sample (r ≈ −0.25) could not be fully explained within the scope of this exploratory study. While a plausible interpretation was offered, linking this pattern to heightened critical awareness among technically experienced respondents, this finding warrants dedicated investigation in future research, potentially through qualitative methods that can probe the mechanisms underlying divergent perceptions of explainability among operational AI users.

Additional limitations relate to the illustrative nature of the industrial case and the exploratory scope of the empirical component. The framework has not yet been validated using longitudinal industrial datasets or experimentally benchmarked against alternative operational architectures. Accordingly, the findings should be interpreted as conceptual and organizational grounding rather than formal engineering validation.

5.5. Future Research Directions

Future research can extend this work in several directions. Quantitative studies could operationalize PRIME–INSPECT dimensions in larger samples and examine their interrelationships using inferential methods, thereby testing specific propositions derived from the framework. Longitudinal designs could explore how trust calibration and governance maturity evolve as AI systems move from pilot phases to full operational deployment.

Comparative studies across industries with differing risk profiles could assess how the relative importance of the PRIME and INSPECT dimensions varies across contexts. In addition, qualitative research such as case studies or ethnographic approaches could provide deeper insight into how PRIME–INSPECT is enacted in practice and how organizational actors negotiate control, accountability, and reliance in real-time decision environments.

Future research could also integrate PRIME–INSPECT with emerging regulatory and standards-based frameworks to explore how organizational architectures for trustworthy AI interact with external compliance requirements and evolving legal obligations.

Building on this, future research could further investigate how PRIME–INSPECT performs under real industrial deployment conditions using sensor-level monitoring data, intervention histories, and process safety outcomes. Such work would allow the framework to be evaluated not only conceptually but also operationally.

In addition, future research could more explicitly examine localization challenges in small-language and script-specific environments. Particular attention may be given to Cyrillic-script contexts in the Western Balkans and Bulgaria, where language resources, script variation, and AI model performance may introduce additional governance and interpretability considerations.

6. Conclusions

This paper addressed the growing challenge of deploying artificial intelligence as an active component of real-time decision-making in high-risk industrial systems. The framework should be interpreted as a socio-technical conceptual architecture supported by exploratory empirical grounding rather than as a formally validated industrial control model. As AI increasingly moves beyond analytical support toward autonomous or semi-autonomous operational roles, traditional governance mechanisms struggle to keep pace with the sub-second latencies and potential safety consequences of algorithmic decisions. In response, the study introduced PRIME–INSPECT, a socio-technical framework designed to ensure trustworthy AI performance by tightly coupling operational decision flows with institutional governance safeguards.

The paper’s core contribution lies in articulating a two-layer architecture that explicitly links micro-level operational processes with macro-level organizational oversight. The PRIME layer formalizes the end-to-end flow of AI-supported decisions, spanning prediction, regulation, interpretation, mitigation, and execution, while the INSPECT layer defines the governance maturity, data integrity, and supervisory conditions required for responsible deployment. By explicitly incorporating trust calibration as a cross-layer coupling mechanism, the framework moves beyond binary notions of automation versus human control, emphasizing instead a context-sensitive balance essential for safety-critical environments.

The empirical component provided illustrative grounding for the framework by examining the perceptions of key stakeholders. The findings revealed a “perception gap” between strategic intent (TMT) and operational readiness (IT), underscoring the necessity of a structured architecture to bridge this divide. Furthermore, the illustrative application to blast furnace predictive maintenance demonstrated the framework’s practical utility in a metallurgical context. It showed how abstract governance principles such as “supervisory control” can be translated into concrete operational mechanisms, such as physics-based regulation of sensor data and explicit override protocols, to prevent catastrophic equipment failure.

The revisions introduced in this study further clarify three key aspects of the framework’s contribution. First, PRIME–INSPECT extends existing academic models of AI adoption and trust by integrating behavioral acceptance, human oversight, governance structures, and real-time decision execution into a unified socio-technical architecture. Second, the illustrative blast furnace tuyere application demonstrates how the conceptual framework can be meaningfully connected to representative industrial process parameters, thereby grounding the model in physically realistic operational contexts. Third, the explicit operationalization of trust calibration as a feedback loop—linking prediction outcomes, human intervention, and audit-based adjustment—shows that trust is not treated as an abstract construct, but as a measurable and governable mechanism embedded within the decision workflow.

From a practical standpoint, PRIME–INSPECT provides organizations with a diagnostic tool for AI readiness. It encourages practitioners in the mining and metallurgy sectors to move beyond model-centric evaluations and to consider how explainability, human oversight, and accountability mechanisms jointly determine the safety of AI-driven decisions. In these high-risk industries, where algorithmic errors can lead to physical hazards and regulatory non-compliance, the proposed framework provides the necessary “safety interlocks” between digital predictions and physical assets.

In conclusion, PRIME–INSPECT provides a coherent conceptual foundation for designing trustworthy AI systems in the Industry 4.0 era. By bridging the fragmented literature and validating the approach through both stakeholder perceptions and domain-specific engineering scenarios, the framework contributes to the ongoing evolution of responsible AI, ensuring that operational efficiency does not compromise safety and accountability.