A Data-Driven Evaluation Framework for Quantifying the Impact of Artificial Intelligence on Industrial Process Performance

Abstract

This study proposes a data-driven evaluation framework to quantify the impact of artificial intelligence (AI) on industrial process performance and enterprise value creation. The framework integrates enterprise value assessment based on the Feltham–Ohlson model with a multi-level performance evaluation framework that incorporates a hybrid Analytic Hierarchy Process (AHP) and Entropy Weight Method (EWM) for indicator weighting, together with Fuzzy Comprehensive Evaluation (FCE) for multi-dimensional aggregation. This integrated approach enables systematic analysis of AI-driven effects from the perspectives of intelligent investment input, operational governance environment, and process output performance. Using panel data from 3515 Chinese A-share listed firms (20,076 firm-year observations) during 2014–2022, a Process Performance Index (PI) is constructed to measure AI-enabled operational capability across resource allocation efficiency, coordination effectiveness, and production performance dimensions. Empirical results indicate that PI is positively associated with abnormal earnings and firm profitability, demonstrating that AI-enabled process capability contributes to sustained enterprise value growth. The findings further show increased digital technology investment intensity, knowledge-based human capital accumulation, and improved data governance conditions, accompanied by enhanced production and service performance. By explicitly integrating AHP–EWM weighting and FCE aggregation within the Feltham–Ohlson valuation structure, the proposed framework provides an interpretable quantitative mechanism linking AI adoption, operational capability development, and enterprise value creation. The results offer practical insights for evaluating intelligent transformation strategies in the context of Industry 5.0 and data-driven industrial development.

1. Introduction

The rapid advancement of information technologies, including big data, cloud computing, and the Internet of Things, has created the technical foundation for the large-scale deployment of artificial intelligence (AI) in industrial environments [1]. Rather than being limited to isolated digital services, AI is increasingly embedded in core production and operational processes, enabling data-driven monitoring, adaptive control, and intelligent decision-making across the entire lifecycle of industrial systems [2,3]. As a result, traditional production lines are evolving toward intelligent and interconnected process networks in which scheduling, resource allocation, and performance optimization can be continuously adjusted based on real-time information. This transformation reflects the broader paradigm shift from Industry 4.0 toward Industry 5.0, where intelligent automation is increasingly integrated with human-centered design, resilient supply chains, and sustainable production strategies [4,5]. Compared with Industry 4.0, which emphasizes cyber–physical integration and automation efficiency, Industry 5.0 highlights human–AI collaboration, environmental sustainability, and system robustness. This shift implies that evaluation frameworks must move beyond traditional productivity metrics and incorporate multidimensional indicators that reflect adaptability, resilience, and long-term value creation.

In modern manufacturing systems, AI-enabled robots and cyber–physical platforms support highly automated and precision-oriented operations. Examples include robotic assembly and disassembly systems, collaborative human–robot production lines, and multi-factory coordinated scheduling, which significantly improve throughput, reduce defect rates, and enhance resource utilization [6,7,8,9]. Beyond the shop floor, AI-enabled analytics facilitate predictive planning, dynamic task allocation, and intelligent logistics coordination, allowing enterprises to respond rapidly to disturbances and demand fluctuations. Through capabilities in pattern recognition, data fusion, and automated decision support, AI has become a key enabler of intelligent operation and performance-oriented optimization in industrial processes [10,11]. In particular, AI enhances resource allocation efficiency by enabling dynamic scheduling, predictive maintenance planning, and adaptive coordination across distributed production units. These improvements strengthen operational flexibility and total factor productivity, enabling enterprises to optimize the utilization of both physical and digital resources.

Despite these advantages, quantitatively measuring how AI affects industrial process performance remains challenging. Improvements produced by AI are multidimensional and may appear as increased investment in intelligent equipment, enhanced data governance environments, and improved production and service outputs. International organizations have developed standards and frameworks to guide AI governance and performance assessment. For example, ISO/IEC 42001 specifies requirements for AI management systems [12], the NIST AI Risk Management Framework provides structured lifecycle guidance for trustworthy AI deployment [13], and ISO 22400 defines standardized key performance indicators (KPIs) for manufacturing operations management [14]. These frameworks improve governance transparency and risk control but do not provide a direct quantitative mechanism for measuring how AI contributes to enterprise value creation. In particular, they focus on responsible AI deployment rather than evaluating the marginal economic benefits generated through AI-enabled process optimization.

To bridge this gap, many existing studies attempt to evaluate AI or digitalization using customized indicator systems or statistical learning models. Pandey et al. [15] combined data mining with indicator-based assessment to analyze enterprise digital adoption, while Yu [16] applied time-series analysis to evaluate regional digital economy development. Ali et al. [17] integrated multidimensional indicators with learning algorithms to analyze the sustainability impacts of AI. Existing approaches can generally be classified into three categories: indicator-based frameworks, data-driven statistical models, and hybrid evaluation approaches. Indicator-based frameworks provide interpretability but often rely on fixed weighting structures that lack adaptability to rapidly evolving AI-enabled operational environments. Data-driven models offer strong predictive performance but frequently operate as black boxes, limiting interpretability for managerial decision-making [18]. Hybrid approaches improve flexibility but often lack a clear theoretical linkage between operational indicators and enterprise value creation.

Although these approaches reveal useful correlations, they face several limitations when applied to industrial systems. Static indicator systems cannot easily adapt to rapidly evolving AI-enabled production technologies. Furthermore, strong heterogeneity across enterprises reduces the transferability of models calibrated at macro or regional levels. Many predictive models sacrifice interpretability, making it difficult to connect predicted outcomes with operational improvement mechanisms. Hybrid methods such as the neural-network-based framework of Dong et al. [19] improve modeling flexibility but often rely on simplifying assumptions that cannot fully capture nonlinear process dynamics. More importantly, most existing studies evaluate operational performance or financial outcomes separately, without explicitly modeling the transmission mechanism through which AI-enabled process optimization generates abnormal earnings and long-term enterprise value. This lack of theoretical linkage limits the ability of managers and policymakers to interpret how AI investment contributes to measurable economic performance improvements.

Therefore, there is a need for an evaluation framework that simultaneously links AI adoption to measurable value creation, captures multidimensional operational conditions, and maintains interpretability for engineering and managerial decision-making. In particular, connecting financial value evolution with process-level input, environment, and output indicators provides a holistic perspective on how AI-enabled optimization translates into tangible operational gains. Such a framework should integrate KPIs describing intelligent investment intensity, governance conditions, and operational output efficiency into a unified quantitative structure capable of explaining both process-level performance improvement and enterprise value enhancement.

To address these challenges, this paper proposes a data-driven evaluation framework for quantifying the impact of artificial intelligence on industrial process performance. The framework integrates enterprise value modeling based on the Feltham–Ohlson (F–O) theory [20,21] with a multi-level fuzzy comprehensive evaluation method. The F–O valuation framework provides a theoretically grounded mechanism linking abnormal earnings to firm value, allowing AI-enabled operational capability to be interpreted as value-relevant “other information” beyond traditional accounting variables [22]. By operationalizing this component through a structured indicator system weighted using hybrid AHP–Entropy methods [23], the proposed framework establishes an explicit analytical linkage between AI adoption, process optimization, and enterprise value creation.

By leveraging longitudinal enterprise data, the proposed framework provides an interpretable and systematic approach to evaluating how AI-enabled technologies influence operational efficiency and value generation in industrial systems.

The main contributions of this study are summarized as follows:

• A unified data-driven evaluation framework is developed to quantify the impact of artificial intelligence on industrial process performance by integrating value-based financial modeling with a multi-level evaluation system using AHP-Entropy weighting and fuzzy comprehensive evaluation.
• A multidimensional indicator system is constructed based on Input–Environment–Output logic to describe interactions among intelligent investment, governance conditions, and operational output performance.
• The study establishes an explicit analytical linkage between AI-enabled operational capability and abnormal earnings within the Feltham–Ohlson valuation structure, providing theoretical support for evaluating intangible AI-driven assets.
• An empirical study based on longitudinal enterprise data validates the statistical reliability and practical effectiveness of the proposed framework.
• The resulting evaluation methodology provides interpretable decision support for intelligent resource allocation, operational optimization, and policy design in AI-enabled industrial systems.

2. Related Work

2.1. Indicator-Based Evaluation Frameworks

Indicator-based evaluation frameworks represent a foundational and widely adopted approach for assessing the value and impact of complex systems, including enterprise informatization, digital transformation, and AI-enabled industrial processes. These frameworks typically construct hierarchical indicator systems covering multiple dimensions such as technology investment intensity, innovation capability, operational efficiency, governance environment quality, and output performance. A composite performance score is then derived using structured weighting and aggregation methods, including expert scoring, Analytic Hierarchy Process (AHP), entropy weighting, or survey-based evaluation [23].

The main advantage of indicator-based frameworks lies in their strong interpretability and structural transparency. The multi-level indicator hierarchy enables decision makers to examine the contribution of different dimensions of enterprise capability, thereby supporting diagnostic analysis of strengths and weaknesses across technological, organizational, and environmental factors. For example, Kulyasov et al. [24] developed a multi-indicator framework to evaluate the ecological and economic effectiveness of decarbonization strategies in industrial enterprises, demonstrating how structured evaluation systems can be applied to complex multi-objective decision scenarios. Similarly, digital transformation maturity models employ structured indicator systems to quantify enterprise readiness for Industry 4.0 and Industry 5.0 adoption [25].

Despite these advantages, indicator-based frameworks face several important limitations when applied to AI-enabled industrial systems. First, they typically rely on expert-defined weights and relatively static index structures, which may not adapt effectively to rapidly evolving AI technologies and production configurations [25]. Second, most indicator-based approaches employ linear aggregation mechanisms, which are insufficient to capture nonlinear and coupled relationships among investment inputs, operational environments, and process outputs [26,27]. Third, standardized KPI systems clarify what should be measured but do not explain how AI-enabled improvements translate into economic value creation. In particular, isolating the marginal contribution of AI from concurrent investments in infrastructure, organizational capability, and digital transformation remains a major methodological challenge [28]. Consequently, traditional indicator-based evaluation approaches are limited in their ability to explain the transmission mechanism linking process-level improvements to enterprise value creation.

2.2. Data-Driven and Machine Learning-Based Assessment Models

With the increasing availability of enterprise-level operational and financial data, data-driven evaluation approaches based on statistical learning and machine learning have attracted growing attention. These approaches typically leverage panel data, econometric modeling, or predictive learning algorithms to estimate relationships between digital adoption, AI capability development, and performance outcomes. Compared with indicator-based frameworks, data-driven models provide a stronger capability for identifying hidden patterns and nonlinear relationships from large-scale datasets.

For example, Pandey et al. [15] combined data mining with structured evaluation indicators to assess digital adoption requirements of small and medium-sized enterprises. Yu [16] employed time-series econometric models to analyze regional digital economy development patterns and identify growth dynamics associated with technology adoption. Ali et al. [17] integrated machine learning algorithms with sustainability indicators to evaluate the contribution of AI-enabled technologies to circular economy development. Mikalef and Gupta [29,30] conceptualized AI capability as a measurable organizational resource influencing firm-level performance.

Although these approaches provide valuable empirical evidence, they often prioritize predictive accuracy rather than interpretability. Many machine learning models operate as black-box predictors, making it difficult to attribute observed performance improvements to specific operational factors such as intelligent investment intensity, governance structure optimization, or process-level efficiency gains. As emphasized by Rudin [18], interpretable models are particularly important in high-stakes decision-making contexts where managerial transparency and accountability are required. Furthermore, empirical findings derived from data-driven approaches often face comparability limitations due to heterogeneity in industrial sectors, enterprise size distributions, and technological maturity levels. Consequently, purely data-driven approaches often lack sufficient explanatory power to support strategic decision-making related to AI investment planning and resource allocation in industrial systems.

2.3. Hybrid Evaluation Approaches

To balance interpretability with analytical flexibility, hybrid evaluation approaches have been proposed that integrate structured indicator systems with computational intelligence or multi-criteria decision-making methods. These approaches typically combine statistical modeling, optimization algorithms, or machine learning techniques with hierarchical indicator structures to improve model adaptability and predictive capability.

For example, Luthra et al. [26] proposed a hybrid AHP–VIKOR framework for sustainable supplier selection under multi-criteria decision environments. Shen et al. [27] combined machine learning with innovation value chain indicators to evaluate enterprise green innovation performance. Such hybrid approaches improve methodological flexibility and can better capture nonlinear interactions among evaluation indicators. However, these models are typically designed to produce relative rankings or composite scores rather than to establish theoretically grounded relationships between operational performance metrics and enterprise financial value.

This limitation becomes particularly evident when considering classical financial valuation theories. The Feltham–Ohlson (F–O) model [20,21] demonstrates that enterprise value is determined not only by book value but also by the present value of expected abnormal earnings. Prior research indicates that intangible assets, innovation capability, and digital transformation strategy significantly influence abnormal earnings and long-term firm value [22,31]. This theoretical structure provides a rigorous analytical basis for integrating non-financial operational indicators into financial valuation models. AI-enabled operational capabilities, such as intelligent resource allocation, predictive maintenance, and adaptive production scheduling, can therefore be interpreted as economically relevant information influencing abnormal earnings generation.

Despite its theoretical advantages, the F–O model has rarely been integrated with structured operational evaluation frameworks in AI-enabled industrial contexts. Existing hybrid evaluation approaches primarily focus on methodological integration without establishing explicit theoretical linkages between process-level indicators and enterprise value creation. This limitation highlights the need for a unified framework capable of linking micro-level operational indicators with macro-level financial performance through an interpretable analytical structure.

The present study addresses this gap by integrating the F–O valuation model with a multi-level fuzzy comprehensive evaluation framework based on AHP–Entropy weighting [23]. Unlike purely statistical or composite scoring approaches, the proposed framework explicitly models how AI-enabled operational capability contributes to abnormal earnings through multidimensional indicators describing intelligent investment input, governance environment quality, and process output performance. This integration responds to calls in accounting and management literature to develop analytical frameworks linking non-financial performance indicators to enterprise value creation [32].

By establishing an interpretable linkage between AI-enabled operational optimization and financial valuation, the proposed framework provides a theoretically grounded approach for evaluating industrial process performance in the context of intelligent manufacturing and Industry 5.0 transformation. Compared with existing indicator-based, data-driven, and hybrid evaluation approaches, the proposed method improves interpretability, theoretical consistency, and decision relevance.

3. Construction of the Assessment Model

This section presents the construction of the proposed hybrid assessment model, a multi-stage analytical framework designed to quantitatively link AI-enabled operational capability with enterprise value creation. The objective of the framework is to establish a theoretically grounded and empirically measurable pathway through which artificial intelligence influences industrial process performance and subsequently contributes to abnormal earnings and firm value.

As illustrated in Figure 1, the framework integrates financial valuation theory with multi-dimensional operational performance evaluation. Specifically, AI-enabled process capability is represented through a Process Performance Index (PI), which is constructed from a structured indicator system capturing intelligent investment input, operational governance environment, and process output performance. The PI serves as a measurable representation of enterprise-level AI operational capability and is incorporated into the F–O valuation structure as value-relevant “other information” affecting abnormal earnings.

The proposed methodology unfolds in three sequential stages. First, the theoretical foundation of the model is established by adopting the F–O valuation framework. The F–O model provides a rigorous accounting-based structure linking firm value to book value, abnormal earnings, and value-relevant information beyond traditional financial variables. In the context of AI-enabled industrial systems, operational capability improvements arising from intelligent resource allocation, adaptive production control, and data-driven decision-making can be interpreted as intangible value drivers influencing future abnormal earnings. Therefore, the F–O model offers an appropriate theoretical basis for incorporating process-level AI performance indicators into financial valuation analysis.

Second, a multi-level indicator system is constructed to empirically characterize AI-enabled operational capability. The indicator hierarchy is designed according to an Input–Environment–Output structure, enabling systematic measurement of intelligent investment intensity, governance, and data environment quality, and process-level output effectiveness. To ensure both interpretability and objectivity, the relative importance of indicators is determined using a hybrid Analytic Hierarchy Process and Entropy Weight Method (AHP–EWM). The AHP component incorporates expert knowledge regarding the relative importance of indicators, while the entropy method captures objective variation in the observed data, thereby reducing subjectivity bias and improving robustness of the weighting structure.

Finally, Fuzzy Comprehensive Evaluation (FCE) is employed to aggregate the weighted multi-level indicators into a single composite metric, referred to as the Process PI. The FCE approach is particularly suitable for evaluating complex industrial systems characterized by uncertainty, heterogeneity, and nonlinear interactions among multiple performance dimensions. Through fuzzy membership functions and hierarchical aggregation, FCE enables the transformation of multi-source indicator information into a normalized performance index that reflects the overall effectiveness of AI-enabled industrial processes.

The resulting PI provides a quantitative representation of enterprise-level AI operational capability and is incorporated into the F–O valuation model as an explanatory variable for abnormal earnings. This integration establishes an explicit analytical linkage between AI-enabled process optimization and enterprise value creation, thereby bridging the gap between operational performance measurement and financial valuation theory.

3.1. Theoretical Foundation: Linking AI-Driven Operational Capabilities to Enterprise Value via the Feltham–Ohlson Model

With the continuous penetration of artificial intelligence technologies into industrial production systems, enterprise value has emerged as a comprehensive and integrative indicator for evaluating the effectiveness of AI-enabled operational transformation. AI technologies support a wide range of industrial applications, including intelligent manufacturing systems, automated scheduling mechanisms, predictive maintenance platforms, adaptive process control, and data-driven decision-support environments. These capabilities extend beyond traditional digital services and fundamentally reshape the operational logic of industrial production systems by enabling real-time coordination, adaptive optimization, and intelligent resource allocation.

In modern industrial environments, AI-enabled operational capability represents an important source of firm-specific competitive advantage. Through intelligent scheduling, predictive optimization, adaptive production planning, and enhanced data utilization, AI improves resource allocation efficiency and enhances operational resilience under uncertain conditions. These process-level improvements generate economic benefits that extend beyond traditional physical assets and are typically reflected through improved productivity, reduced operational cost, and enhanced system robustness. Consequently, AI-enabled operational capability constitutes an intangible strategic asset that contributes to long-term enterprise value creation.

From a process-operation perspective, AI improves production efficiency by optimizing task allocation, improving resource utilization, reducing operational uncertainty, and supporting data-informed managerial decision-making. AI-driven industrial applications also promote integration between manufacturing and service-oriented activities, including predictive maintenance, intelligent logistics coordination, and data-driven industrial service systems. These improvements enhance production continuity, increase throughput stability, and improve system-level coordination efficiency. As a result, evaluating enterprise value under AI-enabled operational transformation provides an economically meaningful basis for assessing the performance impact of artificial intelligence in industrial processes.

The selection of the F–O model as the theoretical foundation of this study is motivated by the need to capture the economic value of intangible AI-driven capabilities. Traditional Discounted Cash Flow (DCF) approaches typically treat firms as black-box generators of cash flows and provide limited insight into how operational improvements translate into financial performance. In contrast, the F–O model establishes a transparent analytical linkage between accounting variables, abnormal earnings, and firm value [20,21]. This structure makes the model particularly suitable for analyzing AI-enabled industrial systems, where performance improvements arise primarily from intangible sources such as algorithmic decision capability, intelligent process optimization, and data-driven operational coordination.

Within the F–O framework, enterprise value is determined by both book value and the expected future stream of abnormal earnings. Abnormal earnings represent the portion of firm income that exceeds the normal expected return on invested capital, thereby capturing firm-specific competitive advantages.

Total earnings are decomposed as follows:

$$x_t=x_t^n+x_t^a$$

(1)

where $x_t$ represents total earnings; $x_t^n$ denotes normal earnings; and $x_t^a$ denotes abnormal earnings.

Normal earnings are defined as the expected return on book value:

$$x_t^n=g_f{\varphi }_{t-1}$$

(2)

where ${\varphi }_{t-1}$ denotes the book value of net assets and $g_f$ represents the normal expected rate of return.

Abnormal earnings are expressed as

$$x_t^a=p_t-g_f{\varphi }_{t-1}$$

(3)

where $p_t$ represents observed earnings.

The abnormal earnings term $x_t^a$ is particularly important in the context of AI-enabled industrial transformation because it captures the financial premium generated by firm-specific operational advantages. These advantages include intelligent scheduling capability, predictive maintenance optimization, adaptive production control, and enhanced data utilization efficiency. Through these mechanisms, enterprises adopting AI technologies may achieve performance levels exceeding industry benchmarks, resulting in persistent abnormal earnings.

Under the residual income valuation structure, enterprise value is defined as

$${\Gamma }_t={\varphi }_t+\sum _{i=1}^{\infty }{\displaystyle \frac{E_t\left(x_{t+i}^a\right)}{{(1+g_f)}^i}}$$

(4)

This formulation indicates that enterprise value consists of two components: current book value and the discounted present value of expected future abnormal earnings.

In AI-enabled industrial environments, this formulation is particularly meaningful because long-term value creation is primarily driven by persistent abnormal earnings generated through intelligent operational capability rather than solely through physical capital accumulation. Therefore, the F–O model provides a theoretically grounded framework for evaluating the financial contribution of AI-enabled process optimization.

The dynamic evolution of abnormal earnings follows the linear information dynamics structure:

$$x_{t+1}^a=\upsilon x_t^a+H_t+{\epsilon }_{1,t+1}$$

(5)

where $\upsilon$ represents the persistence parameter of abnormal earnings, and $H_t$ represents value-relevant information beyond traditional accounting variables.

The variable $H_t$ plays a crucial theoretical role because it captures non-financial information influencing future firm performance. In the context of AI-enabled industrial systems, $H_t$ can be interpreted as the operational capability generated through intelligent investment, data governance infrastructure, and process optimization effectiveness. These multidimensional capabilities represent the mechanism through which AI adoption influences abnormal earnings generation.

Therefore, the Process PI serves as an empirical operationalization of $H_t$. The PI aggregates multidimensional indicators describing intelligent investment input, operational governance environment quality, and process output effectiveness, thereby providing a measurable representation of AI-enabled operational capability. Through this structure, process-level performance evaluation is directly incorporated into the financial valuation framework.

To further improve interpretability, abnormal earnings can be expressed through observable operational indicators derived from the DuPont decomposition structure:

$$ROE=ROA\times EM$$

(6)

$$ROA=MOS\times \kappa$$

(7)

$${\theta }_t={\displaystyle \frac{S_t(MO{S}_t{\kappa }_{t}E{M}_{t-1}-r)}{{\kappa }_{t}E{M}_{t-1}}}$$

(8)

This decomposition enables abnormal earnings to be interpreted using observable operational indicators such as profit margin, asset utilization efficiency, and financial leverage. These indicators reflect operational improvements resulting from intelligent scheduling strategies, adaptive capacity utilization, predictive maintenance planning, and data-driven process optimization. Consequently, the F–O model establishes a transparent analytical linkage between AI-enabled operational capability and enterprise value creation.

The importance of incorporating non-financial performance indicators into valuation models has been widely recognized in accounting and finance literature, particularly as intangible assets increasingly dominate firm value [22,31]. AI-enabled operational capability represents a new category of intangible asset that influences enterprise value through improved process efficiency, enhanced coordination capability, and adaptive decision-making mechanisms. By integrating the Process Performance Index into the F–O valuation structure, this study provides a theoretically consistent framework linking AI adoption, operational performance improvement, and financial value creation.

Therefore, the Feltham–Ohlson model provides a rigorous theoretical bridge linking AI-enabled operational capability with enterprise value, forming the financial foundation of the proposed hybrid evaluation framework.

3.2. Construction and Weighting of Multi-Level Indicator System Based on Input–Environment–Output Logic

To empirically operationalize the “other value-relevant information” term ($H_t$) introduced in the F–O model, this study constructs a structured multi-level evaluation indicator system (

Table 1. Evaluation indicator system for AI-enabled operational capability.

Primary Dimension	Secondary Indicator	Code	Proxy Variable (Definition)
Intelligent Investment	R&D Intensity	C1	R&D expenditure/total revenue
	Human Capital	C2	proportion of R&D personnel
	Digital Intangible Assets	C3	digital intangible assets/total intangible assets
	Government Support	C4	government subsidies/total assets
Operational & Institutional Environment	Operating Cost Control	C5	operating expense ratio
	Asset Management Efficiency	C6	total asset turnover
	Inventory Management Level	C7	inventory/total assets
	Customer Concentration	C8	HHI customer index
	Supplier Concentration	C9	HHI supplier index
	Governance Supervision Quality	C10	proportion of independent directors
	Ownership Concentration	C11	largest shareholder ownership ratio
Process Output	Technological Innovation Output	C12	number of invention patent applications
	Profitability	C13	gross profit margin
	Enterprise Growth	C14	revenue growth rate

) capable of quantifying AI-enabled operational capability. The $H_t$ term captures economically relevant information that is not fully reflected in traditional accounting variables. In AI-enabled industrial environments, such information includes intelligent investment intensity, digital infrastructure maturity, governance efficiency, and process-level productivity improvement. These operational characteristics constitute the underlying mechanism through which artificial intelligence contributes to abnormal earnings generation and enterprise value creation.

Therefore, the indicator system developed in this study serves as the empirical representation of AI-enabled operational capability within the F–O valuation framework. By translating process-level improvements into measurable indicators, the proposed system enables integration of non-financial operational information into enterprise value analysis.

3.2.1. Construction of the Evaluation Indicator System

AI-enabled industrial performance improvement can be interpreted as a dynamic interaction process among intelligent investment, operational environment conditions, and process output effectiveness. From the perspectives of economic growth theory and industrial organization theory, artificial intelligence contributes to enterprise productivity improvement through three primary mechanisms:

Intelligent factor investment enhances enterprise production capability through the development of digital infrastructure, deployment of algorithmic decision-support systems, and cultivation of skilled technical personnel with expertise in artificial intelligence and data analytics. Operational environment optimization improves coordination efficiency by strengthening governance structures, enhancing supply chain integration, and establishing institutional mechanisms that support stable and secure implementation of AI-enabled industrial processes. Process output improvement reflects observable gains in productivity, innovation capability, and financial performance, resulting from enhanced resource allocation efficiency and adaptive process optimization enabled by artificial intelligence technologies.

These mechanisms jointly describe how AI adoption affects industrial production systems through resource allocation efficiency improvement, operational coordination enhancement, and continuous process optimization.

Based on this theoretical logic, this study constructs a three-level evaluation structure following the Input–Environment–Output (I–E–O) analytical framework. Compared with macro-level digital economy indicator systems frequently used in regional studies, the proposed framework focuses on firm-level measurable indicators that directly capture enterprise operational capability and financial performance. This improves cross-firm comparability and reduces measurement bias introduced by regional economic heterogeneity or industry-level aggregation effects.

The indicator selection process integrates theoretical insights from intelligent manufacturing literature, enterprise digital transformation studies, and financial performance evaluation research. Candidate indicators were initially collected from prior empirical studies on AI capability evaluation, industrial digitalization assessment, and innovation performance measurement. The final selection was refined through expert consultation and data availability screening to ensure relevance, interpretability, and empirical measurability. The resulting indicator system provides a structured representation of AI-enabled operational capability that is compatible with financial valuation modeling.

The final evaluation indicator system consists of

• Three first-level dimensions describing the major mechanisms through which AI influences industrial performance;
• Fourteen second-level indicators capturing key operational characteristics of AI-enabled industrial systems;
• Measurable firm-level proxy variables derived from financial statements, patent databases, and corporate governance disclosures.

The three primary dimensions are defined as follows:

(1) Intelligent Investment (B1). This dimension reflects the foundational resource commitment required to develop AI-enabled operational capability. AI technologies rely heavily on knowledge accumulation, data infrastructure construction, and skilled technical personnel. Therefore, intelligent investment captures the intensity of technological input and digital-capability accumulation within the enterprise. Specifically, R&D Intensity (C1) measures financial investment in technological development, Human Capital (C2) reflects the proportion of skilled technical personnel engaged in innovation activities, Digital Intangible Assets (C3) represent accumulated digital knowledge resources such as software and data assets, and Government Support (C4) captures policy-driven external resource acquisition supporting AI innovation activities.

(2) Operational and Institutional Environment (B2). This dimension describes the internal and external conditions under which intelligent investments are transformed into operational performance improvements. AI technologies typically generate value through improved coordination efficiency and optimized resource allocation. Therefore, the operational environment includes both internal efficiency indicators and external governance conditions. Operating Cost Control (C5) reflects enterprise cost management capability, Asset Management Efficiency (C6) captures asset utilization effectiveness, and Inventory Management Level (C7) reflects supply chain coordination efficiency. Market Structure indicators include Customer Concentration (C8) and Supplier Concentration (C9), which measure the stability and resilience of enterprise supply chain relationships. Governance Supervision Quality (C10) evaluates the effectiveness of corporate governance mechanisms, while Ownership Concentration (C11) reflects decision-making stability and strategic consistency.

(3) Process Output (B3). This dimension evaluates the observable economic outcomes generated by AI-enabled operational optimization. Process output indicators measure how effectively intelligent investment and operational coordination improvements are translated into tangible enterprise performance gains. Technological Innovation Output (C12) captures innovation capability through patent applications, Gross Profit Margin (C13) reflects operational efficiency and cost optimization outcomes, and Enterprise Growth Rate (C14) measures market expansion capability and dynamic competitiveness improvement.

These three dimensions jointly represent the multidimensional mechanism through which AI adoption contributes to abnormal earnings generation. Intelligent investment provides the necessary technological foundation; operational environment indicators capture coordination and governance efficiency; and process output indicators reflect realized productivity improvements and market performance. Together, these indicators form the empirical basis for constructing the Process PI, which serves as the measurable proxy for AI-enabled operational capability in the F–O valuation framework.

The selected indicators emphasize measurability, interpretability, and consistency with financial valuation theory. All indicators are observable at the firm level and can be obtained from enterprise financial reports, patent databases, and corporate governance disclosures. This ensures replicability and improves comparability across enterprises operating in heterogeneous industrial environments.

3.2.2. Determination of Indicator Weights: Hybrid AHP–Entropy Method

Determining appropriate indicator weights is a critical step in constructing a reliable multi-indicator evaluation framework. A single weighting strategy often introduces methodological limitations. Purely subjective weighting approaches, such as expert scoring or Analytic Hierarchy Process (AHP), may reflect domain knowledge but are potentially affected by expert bias. Conversely, purely objective methods such as the Entropy Weight Method (EWM) rely exclusively on statistical dispersion characteristics of data and may neglect the theoretical importance of indicators related to AI-enabled operational capability. To address these limitations, this study adopts a hybrid weighting approach integrating AHP and EWM to balance theoretical interpretability and empirical robustness.

The hybrid weighting strategy ensures that indicator importance reflects both domain expertise and observable data variation. By combining subjective knowledge regarding the operational significance of AI capability factors with objective information derived from large-scale enterprise data, the resulting weight vector improves reliability, reduces bias, and enhances interpretability of the Process Performance Index (PI). Similar hybrid weighting approaches have been widely adopted in multi-criteria decision analysis literature to improve robustness of evaluation systems [33,34,35].

3.2.3. Subjective Weights Based on AHP

The Analytic Hierarchy Process (AHP) is employed to derive subjective weights reflecting theoretical and practical importance of each indicator in explaining AI-enabled operational capability. AHP is particularly suitable for complex decision environments where multiple criteria influence evaluation outcomes.

To ensure reliability of expert judgment, this study convened a structured expert panel consisting of 15 specialists with interdisciplinary expertise in industrial AI applications, intelligent manufacturing systems, and enterprise performance evaluation. The panel included:

• Five senior executives from manufacturing enterprises with practical experience in AI-enabled digital transformation;
• Five academic researchers specializing in artificial intelligence, digital economy, and industrial engineering;
• Five data scientists with experience in enterprise AI deployment and intelligent decision-support systems.

Experts performed pairwise comparisons of the 14 secondary indicators according to their perceived importance in influencing AI-enabled operational capability and enterprise performance improvement. Pairwise comparison matrices were constructed following standard AHP procedures.

To ensure logical consistency of expert judgments, each comparison matrix was evaluated using the consistency ratio (CR):

$$CR={\displaystyle \frac{CI}{RI}}$$

(9)

where $CI$ denotes the consistency index, and $RI$ represents the random consistency index. Following established AHP criteria, matrices satisfying $CR<0.1$ were considered acceptable, ensuring logical coherence of expert evaluations.

This process ensures that subjective weights incorporate domain knowledge regarding AI capability development, industrial operational mechanisms, and enterprise value creation logic.

The resulting subjective weight vector is expressed as

$$W^{AHP}=(w_1^{AHP},w_2^{AHP},\dots ,w_{14}^{AHP})$$

(10)

Objective Weights Based on the Entropy Weight Method

While AHP captures the theoretical importance of indicators, it does not account for statistical variability in empirical observations. Therefore, the Entropy Weight Method (EWM) is introduced to derive objective weights based on the dispersion characteristics of indicator data.

The entropy weighting procedure is applied to the full panel dataset consisting of 20,076 firm-year observations from 3515 listed industrial enterprises over the period 2018–2022. The large sample size improves the statistical reliability of indicator variability measurement and reduces sensitivity to firm-specific noise.

The entropy value for each indicator is calculated as

$$e_j=-k\sum _{i=1}^n{p}_{ij}ln\left(p_{ij}\right)$$

(11)

where $p_{ij}$ represents the normalized value of indicator j for observation i, and k is a normalization constant.

The entropy weight is determined as

$$w_j^{EWM}={\displaystyle \frac{1-e_j}{{\sum }_{j=1}^m(1-e_j)}}$$

(12)

Indicators with greater variability across enterprises contain more discriminative information and therefore receive larger weights. This ensures that indicators contributing stronger differentiation in AI-enabled operational capability have greater influence on the overall evaluation score.

The resulting objective weight vector is expressed as

$$W^{EWM}=(w_1^{EWM},w_2^{EWM},\dots ,w_{14}^{EWM})$$

(13)

Comprehensive Weight Integration

To integrate subjective expert knowledge with objective data characteristics, the final indicator weight is determined using a linear combination:

$$w_j=\alpha w_j^{AHP}+(1-\alpha )w_j^{EWM}$$

(14)

where

• $w_j^{AHP}$ denotes subjective weight derived from expert evaluation;
• $w_j^{EWM}$ denotes objective entropy weight derived from data dispersion;
• $\alpha$ represents the balance coefficient.

In this study, $\alpha =0.5$ is selected to assign equal importance to theoretical knowledge and empirical information. This balanced weighting strategy reduces bias associated with purely subjective evaluation while preserving the theoretical interpretability of AI capability indicators.

The final weight vector is expressed as

$$W=(w_1,w_2,\dots ,w_{14})$$

(15)

By integrating expert knowledge with empirical variability information, the hybrid AHP–EWM approach improves robustness, interpretability, and stability of the evaluation system. The resulting weights provide a theoretically grounded and data-supported foundation for constructing the fuzzy comprehensive evaluation model.

3.2.4. Performance Scoring via Automated Fuzzy Comprehensive Evaluation

With the comprehensive weight vector $W=(w_1,w_2,\dots ,w_{14})$ determined in Section 3.2.2, the Fuzzy Comprehensive Evaluation (FCE) method is employed to aggregate multi-dimensional indicators into a single quantitative performance score, referred to as the Process Performance Index (PI). The PI serves as the empirical representation of AI-enabled operational capability and corresponds to the value-relevant information term $H_t$ in the Feltham–Ohlson valuation framework.

Compared with conventional linear aggregation methods, FCE provides a flexible evaluation mechanism capable of handling ambiguity and gradual transitions between performance levels. AI-enabled operational capability often exhibits continuous and nonlinear characteristics, where an enterprise may partially satisfy criteria associated with multiple qualitative performance levels simultaneously. For example, an enterprise may exhibit operational characteristics that are partially “Good” and partially “Excellent” in terms of innovation output or resource allocation efficiency. FCE captures such intermediate states through membership functions, thereby providing a more realistic representation of enterprise performance conditions.

To ensure scalability and objectivity of evaluation across the large panel dataset (20,076 firm-year observations), the fuzzy evaluation procedure is fully automated and data-driven, avoiding reliance on manual expert scoring for membership assignment. This improves reproducibility and reduces subjectivity bias in the evaluation process.

Step 1: Definition of Factor Set and Evaluation Set

The factor set consists of the 14 secondary-level indicators defined in Section 3.2.1:

$$U=\{C_1,C_2,\dots ,C_{14}\}$$

(16)

Each element in the factor set corresponds to one measurable aspect of AI-enabled operational capability.

The evaluation set defines qualitative performance grades used to interpret indicator performance levels:

$$V=\{v_1,v_2,v_3,v_4\}=\{Excellent, Good, Medium, Poor\}$$

(17)

The four-grade structure is adopted to balance interpretability and differentiation capability. It allows sufficient granularity to distinguish enterprise performance levels while maintaining interpretability for managerial decision-making. It is important to note that the four performance grades do not represent an additional hierarchical indicator level, but rather linguistic evaluation categories used in fuzzy reasoning.

Step 2: Construction of Single-Factor Fuzzy Evaluation Matrix

The fuzzy evaluation matrix R captures the membership degree of each indicator with respect to each evaluation grade:

$$R=\left[\begin{array}{cccc}{r}_{11}& r_{12}& r_{13}& r_{14}\\ r_{21}& r_{22}& r_{23}& r_{24}\\ ⋮& ⋮& ⋮& ⋮\\ r_{14,1}& r_{14,2}& r_{14,3}& r_{14,4}\end{array}\right]$$

(18)

where $r_{ij}$ represents the membership degree of indicator $C_i$ to evaluation grade $v_j$.

To ensure objectivity and scalability across heterogeneous enterprises, membership functions are constructed using statistical distribution characteristics of each indicator across the full panel dataset. Specifically, trapezoidal membership functions are employed, and threshold parameters are determined endogenously based on quartile statistics of each indicator’s empirical distribution:

• 25th percentile defines the boundary between Poor and Medium;
• 50th percentile defines the boundary between Medium and Good;
• 75th percentile defines the boundary between Good and Excellent.

This quartile-based membership construction ensures that evaluation criteria reflect the relative performance distribution of enterprises within the sample rather than arbitrary expert-defined thresholds. Consequently, the evaluation framework remains adaptive to industry heterogeneity and improves comparability across firms operating under different production scales and technological maturity levels.

The resulting fuzzy evaluation matrix R has dimension $14\times 4$, representing the membership degree of each indicator to the four evaluation grades.

Step 3: Comprehensive Fuzzy Evaluation and Construction of PI

The comprehensive fuzzy evaluation vector B is obtained through fuzzy synthesis of the hybrid weight vector W and the fuzzy evaluation matrix R:

$$B=W·R$$

(19)

$$B=(b_1,b_2,b_3,b_4)$$

(20)

where

• $b_1$ represents membership degree to the Excellent level;
• $b_2$ represents membership degree to the Good level;
• $b_3$ represents membership degree to the Medium level;
• $b_4$ represents membership degree to the Poor level;

and

$$\sum _{j=1}^4{b}_j=1$$

(21)

According to the principle of maximum membership degree, the qualitative performance level of each firm-year observation can be determined by identifying the largest component in vector B. However, to enable quantitative econometric analysis and facilitate comparison across enterprises and time periods, the linguistic evaluation results are further transformed into a continuous numerical index.

The Process Performance Index (PI) is defined as

$$PI=B·S$$

(22)

where the scoring vector is defined as

$$S={(4,3,2,1)}^T$$

(23)

Through this transformation, qualitative fuzzy evaluation results are mapped into a continuous quantitative scale. A higher PI value indicates stronger AI-enabled operational capability, reflecting higher intelligent investment effectiveness, stronger operational coordination efficiency, and superior process output performance.

Compared with simple weighted summation approaches, the FCE-based PI index captures nonlinear interactions among indicators and provides smoother differentiation between enterprise performance levels. This improves the stability of performance measurement and enhances the interpretability of AI capability evaluation results.

The resulting Process Performance Index serves as the key explanatory variable representing AI-enabled operational capability in the subsequent empirical analysis. By integrating hybrid weighting and fuzzy aggregation mechanisms, the proposed evaluation framework provides a theoretically consistent and empirically robust.

3.3. Linking Process Performance to Enterprise Value

The final step of the proposed framework establishes an explicit analytical linkage between the process-level performance evaluation and enterprise value creation.

Abnormal earnings $x_t^a$ represent the portion of enterprise income exceeding the expected normal return on invested capital. Abnormal earnings, therefore, capture firm-specific competitive advantages generated through superior operational capability, technological innovation, and efficient resource allocation. In AI-enabled industrial environments, these advantages arise from intelligent decision systems, predictive process optimization, adaptive scheduling mechanisms, and enhanced data utilization capability.

The Process Performance Index $P{I}_t$, derived from the fuzzy comprehensive evaluation of multidimensional indicators, serves as a quantitative representation of the “other value-relevant information” term introduced in the F–O model. Specifically, PI captures the joint influence of intelligent investment intensity, operational coordination efficiency, governance effectiveness, and process output performance. These factors represent the operational mechanism through which artificial intelligence contributes to productivity improvement and sustained abnormal earnings.

To empirically examine the relationship between AI-enabled operational capability and financial performance, abnormal earnings are modeled as a function of the Process Performance Index:

$$x_t^a={\beta }_0+{\beta }_{1}P{I}_t+{\epsilon }_t$$

(24)

where $x_t^a$ denotes abnormal earnings calculated under the F–O valuation framework; $P{I}_t$ represents the Process Performance Index of firm i in period t; ${\beta }_0$ is a constant term; ${\beta }_1$ measures the marginal contribution of AI-enabled operational capability to abnormal earnings; and ${\epsilon }_t$ denotes the stochastic disturbance term.

The coefficient ${\beta }_1$ provides a direct quantitative interpretation of the economic value generated by AI-enabled operational improvement. A positive and statistically significant ${\beta }_1$ indicates that improvements in intelligent investment effectiveness, operational environment quality, and process output performance are associated with increased abnormal earnings. This empirical relationship validates the theoretical assumption that AI-enabled operational capability constitutes an important driver of firm value creation.

From a managerial perspective, an increase in PI implies improved coordination efficiency across production resources, enhanced decision quality through data-driven analytics, and stronger innovation output capability. These improvements translate into cost reduction, productivity enhancement, and increased market competitiveness, ultimately contributing to sustained abnormal earnings growth.

Therefore, the proposed framework establishes a logically consistent pathway linking artificial intelligence adoption to enterprise value creation through the following mechanism:

$$\mathrm{AI}\mathrm{Adoption}\to \mathrm{Operational}\mathrm{Capability}\to P{I}_t\to x_t^a\to \mathrm{Enterprise}\mathrm{Value}$$

(25)

By integrating process-level performance evaluation with financial valuation theory, the proposed framework provides both theoretical interpretability and empirical testability. This linkage ensures that the Process Performance Index is not merely a composite indicator, but a value-relevant measure capable of explaining variations in enterprise performance under AI-enabled industrial transformation.

The integrated framework, therefore, enables systematic evaluation of AI-driven operational improvement and its contribution to enterprise value, providing a quantitative basis for analyzing the economic impact of artificial intelligence adoption in industrial processes.

4. Empirical Analysis and Results

This section empirically validates the theoretical framework developed in Section 3. The objective is to examine whether artificial intelligence-enabled operational capability, quantified by the Process Performance Index (PI), contributes to enterprise value creation. By linking AI-enabled process performance to abnormal earnings within the Feltham–Ohlson (F-O) valuation framework, this section provides empirical evidence demonstrating that operational excellence derived from intelligent investment and process optimization generates measurable financial value.

The empirical evaluation proceeds in several stages. First, the research design is introduced, including sample construction, variable definitions, and econometric specification. Second, enterprise value evolution is illustrated using representative firms in order to provide economic intuition consistent with the Feltham–Ohlson valuation logic. Third, descriptive statistics and diagnostic tests are conducted to examine statistical validity and structural reliability of the constructed indicator system. Fourth, baseline regression analysis evaluates the direct relationship between process performance and firm profitability. Fifth, robustness tests verify the stability of empirical results under alternative specifications. Sixth, endogeneity concerns are addressed using instrumental variable estimation. Seventh, mechanism analysis investigates internal efficiency and external information channels through which AI-enabled operational capability creates value. Finally, heterogeneity analysis examines whether the value relevance of process capability differs across ownership structures and corporate life-cycle stages.

4.1. Research Design

4.1.1. Data and Sample Selection

Our sample construction begins with all Chinese A-share listed companies observed over the period 2014–2022. The A-share market provides an appropriate empirical context, as firms exhibit substantial variation in the intensity of artificial intelligence adoption, digital transformation maturity, and operational modernization across industries.

Following standard practices in empirical corporate finance research, several filtering procedures are applied to ensure data quality and comparability:

• Firms belonging to the financial industry are excluded due to their distinct accounting standards and regulatory structure;
• Firms designated as ST or *ST are removed because their abnormal financial conditions may distort performance measurement;
• Observations with missing values for key variables are excluded to ensure econometric validity.

After applying these screening criteria, the final unbalanced panel dataset contains 20,076 firm-year observations representing 3515 unique firms over the nine-year observation period.

To mitigate the influence of extreme observations, all continuous variables are winsorized at the 1st and 99th percentiles. Winsorization preserves distributional characteristics while reducing the impact of outliers and measurement noise.

The large-scale panel dataset captures heterogeneous AI adoption intensity across industries, including intelligent manufacturing, industrial automation, electronics production, digital industrial platforms, and software-enabled production services. The increased cross-sectional and temporal variability improves identification of the structural relationship between AI-enabled operational capability and enterprise value creation.

All data are obtained from the Guotai An (CSMAR) database, which is widely used in empirical accounting and finance research and ensures data reliability and comparability.

4.1.2. Variable Definitions and Model Specification

To operationalize the Feltham–Ohlson-based theoretical framework developed in Section 3, firm performance and firm value are selected as the key outcome variables.

The primary dependent variable is Return on Equity (ROE), an accounting-based measure of profitability that reflects the firm’s ability to generate earnings beyond the opportunity cost of capital. ROE captures abnormal profitability associated with superior operational capability and therefore serves as an observable proxy for abnormal earnings in the F-O framework.

For robustness analysis, Tobin’s Q (TobinQ) is employed as an alternative dependent variable. Tobin’s Q is a forward-looking market-based performance measure reflecting investors’ valuation of both tangible assets and intangible capabilities such as digital infrastructure, process intelligence, and organizational learning capacity.

The core explanatory variable is the Process Performance Index (PI). The PI is a composite indicator synthesizing fourteen granular indicators across three dimensions:

• Intelligent investment input;
• Operational governance and digital environment;
• Process output performance.

Indicator weights are determined using the hybrid AHP–Entropy weighting approach, ensuring that the PI reflects both expert knowledge and empirical variability in observed data.

To isolate the effect of PI and mitigate omitted variable bias, the regression model incorporates a standard set of firm-level control variables widely used in corporate finance research:

• Firm Size (Size), measured as the natural logarithm of total assets;
• Financial leverage (Lev);
• Operating cash flow (Cashflow);
• CEO-Chairman duality indicator (Dual);
• Fixed asset intensity (FIXED), measured as the ratio of fixed assets to total assets;
• Firm age (FirmAge), measured as the natural logarithm of years since establishment.

To test the central hypothesis that superior process performance contributes to enterprise value creation, we employ a two-way fixed effects panel regression model:

$$RO{E}_{it}={\beta }_0+{\beta }_{1}P{I}_{it}+\sum _j{\gamma }_{j}Control{s}_{it,j}+{\mu }_i+{\lambda }_t+{\epsilon }_{it}$$

(26)

where i denotes firm, and t denotes year.

Firm fixed effects (${\mu }_i$) control for time-invariant unobserved heterogeneity such as managerial capability, organizational culture, and long-term technological competence. Year fixed effects (${\lambda }_t$) capture macroeconomic shocks and policy changes affecting all firms simultaneously.

The coefficient of primary interest is ${\beta }_1$, which measures the marginal effect of AI-enabled process performance on firm profitability. Based on the theoretical framework linking PI to abnormal earnings, we expect ${\beta }_1$ to be significantly positive.

All standard errors are clustered at the firm level to account for serial correlation in residuals within firms over time.

4.2. Enterprise Value Illustration Based on the Feltham–Ohlson Model

Although the main empirical estimation relies on a large-scale panel dataset, three representative enterprises are retained for visualization purposes in order to illustrate heterogeneous enterprise value evolution trajectories under AI-enabled transformation. The selection of these firms is intended to be illustrative, covering different industry contexts and AI adoption stages to ensure their representativeness. Specifically, one firm is a leader in intelligent manufacturing with mature AI integration, another is from the industrial automation sector and is in a rapid growth phase, and the third is an innovator in smart logistics. Together, they represent mature, growth, and emerging states of AI enablement, respectively, providing a concrete microcosm for our subsequent large-sample analysis.

Figure 2 shows consistent upward trends in return on equity and abnormal earnings. The results illustrate how AI-enabled process capability improves resource allocation efficiency and supports sustained enterprise value creation consistent with the Feltham–Ohlson abnormal earnings mechanism.

4.3. Descriptive Statistics and Preliminary Analysis

Table 2. Descriptive statistics.

Variable	Obs	Mean	Std. Dev	Min	Max
ROE	20,050	0.067	0.137	−0.962	0.414
PI	20,076	0.115	0.021	0.029	0.363
Size	20,076	22.053	1.179	19.630	26.452
Lev	20,076	0.383	0.191	0.051	0.927
Cashflow	20,076	0.052	0.067	−0.172	0.266
Dual	20,076	0.342	0.474	0.000	1.000
FIXED	20,076	0.214	0.131	0.002	0.721
FirmAge	20,076	2.947	0.296	1.946	3.611

reports descriptive statistics for the main variables used in the empirical analysis.

The mean ROE for the sample is 6.7%, indicating moderate profitability across firms. The Process Performance Index has a mean of 0.115 and a standard deviation of 0.021, suggesting substantial variation in AI-enabled operational capability across firm-year observations.

The considerable dispersion in PI indicates heterogeneous adoption levels of artificial intelligence technologies and digital process optimization strategies, providing sufficient variation for econometric identification of the relationship between process capability and firm value.

To provide an initial assessment of potential multicollinearity issues, variance inflation factors (VIFs) are computed for all explanatory variables. The results are reported in

Table 3. Variance Inflation Factor (VIF) Test.

Variable	VIF	1/VIF
PI	1.06	0.94
Size	1.37	0.73
Lev	1.42	0.70
Cashflow	1.12	0.89
Dual	1.05	0.95
FIXED	1.11	0.90
FirmAge	1.06	0.94

.

The maximum VIF value is 1.42 and the mean VIF is 1.18, both well below the commonly accepted threshold of 10. These results indicate that multicollinearity is not a significant concern in the regression model and that estimated coefficients are unlikely to be distorted by linear dependence among explanatory variables.

Figure 3 shows that the Durbin–Watson statistic is concentrated around 2, indicating weak serial correlation in the regression residuals. This result supports the validity of the panel regression specification and suggests that the fixed-effects structure appropriately captures time-dependent variation.

Figure 4 shows that the Cronbach’s alpha coefficients all exceed 0.9, indicating strong internal consistency among indicators used to construct the Process Performance Index.

In the figure, the yellow dots represent the alpha values for individual firms, while the orange line indicates the average threshold across all firms. This result confirms that the multi-dimensional indicator system reliably captures AI-enabled operational capability across firms.

4.4. Baseline Regression Results

Table 4. Baseline Regression: Impact of Process Performance on Firm Profitability.

	(1) ROE	(2) ROE
PI	1.913 ***	1.624 ***
	(0.065)	(0.060)
Size		0.066 ***
		(0.003)
Lev		−0.401 ***
		(0.009)
Cashflow		0.370 ***
		(0.015)
Dual		0.000
		(0.003)
FIXED		−0.216 ***
		(0.013)
FirmAge		−0.084 ***
		(0.022)
_cons	−0.154 ***	−1.155 ***
	(0.007)	(0.083)

Notes: Standard errors in parentheses. *** $p<0.01$.

presents the results of the baseline regression analysis examining the relationship between the Process Performance Index (PI) and firm profitability (ROE). Column (1) reports the bivariate regression including only PI as the explanatory variable, while Column (2) reports the full specification including all control variables and two-way fixed effects.

The coefficient on PI remains positive and highly statistically significant across both specifications. In the preferred specification including full controls (Column 2), the coefficient estimate of 1.624 indicates that improvements in AI-enabled process capability are strongly associated with enhanced firm profitability.

From an economic magnitude perspective, a one-standard-deviation increase in PI (0.021) corresponds to an increase in ROE of approximately 3.41 percentage points ($1.624\times 0.021$), holding other variables constant. This effect size is economically meaningful, indicating that investments in intelligent process capability produce substantial financial returns.

This finding provides direct empirical support for the theoretical mechanism proposed in Section 3.2.4, confirming that superior process performance translates into abnormal earnings improvements within the Feltham–Ohlson valuation framework. The result suggests that AI-enabled operational capability constitutes a measurable source of competitive advantage and enterprise value creation.

4.5. Further Robustness Tests

To ensure the stability and reliability of the baseline findings, several robustness tests are conducted. These tests are designed to address potential concerns regarding alternative performance measures, sample selection, unobserved heterogeneity, and the predictive power of the proposed index.

First, Tobin’s Q is employed as an alternative forward-looking performance measure capturing market expectations of firm value. The results in

Table 5. Robustness tests.

	(1) TobinQ	(2) ROE	(3) ROE	(4) F.ROE
PI	3.373 ***	1.515 ***	1.627 ***	1.011 ***
	(0.527)	(0.084)	(0.060)	(0.078)
Size	−0.428 ***	0.062 ***	0.066 ***	−0.031 ***
Lev	0.086	−0.402 ***	−0.400 ***	−0.028 **
Cashflow	1.253 ***	0.297 ***	0.369 ***	0.201 ***
Dual	−0.121 ***	−0.000	0.000	0.004
FIXED	0.502 ***	−0.230 ***	−0.218 ***	−0.105 ***
FirmAge	1.826 ***	−0.076 **	−0.082 ***	−0.089 ***
_cons	5.674 ***	−1.065 ***	−1.159 ***	0.914 ***

Notes: Standard errors in parentheses. *** $p<0.01$, ** $p<0.05$.

, Column (1), show a positive and statistically significant coefficient on PI (3.373, $p<0.01$), indicating that capital markets recognize the value relevance of AI-enabled process capability.

Second, the sample is restricted to the pre-pandemic period (2014–2019) in order to mitigate potential macroeconomic disturbances associated with COVID-19. The coefficient for PI in Column (2) remains significantly positive, confirming that the baseline findings are not driven by pandemic-related shocks.

Third, Industry × Year fixed effects are introduced to control for time-varying industry-specific heterogeneity. The estimated coefficient remains highly stable in Column (3) (1.627, $p<0.01$), demonstrating robustness to unobserved industry-level variation.

Finally, the predictive capability of the proposed index is examined using future profitability (F.ROE) as the dependent variable. The significant positive coefficient in Column (4) suggests that improvements in AI-enabled operational capability generate persistent financial benefits rather than merely short-term accounting effects.

Collectively, these results provide strong evidence that the positive relationship between process performance and enterprise value is robust across alternative specifications and measurement approaches.

4.6. Addressing Endogeneity: Instrumental Variable Approach

Although fixed effects control for time-invariant heterogeneity, reverse causality, and omitted time-varying factors may still bias estimation results. For example, firms with higher profitability may have greater resources to invest in intelligent operational systems.

To address endogeneity concerns, we employ a two-stage least squares (2SLS) instrumental variable approach using the leave-one-out industry mean PI as an instrument for firm-level PI. The results are shown in

Table 6. Instrumental variable (2SLS) regression results.

	(1) First Stage	(2) Second Stage
IV	0.796 ***
	(0.089)
PI		1.751 **
		(0.869)
Size	0.004 ***	0.066 ***
Lev	0.002	−0.401 ***
Cashflow	0.022 ***	0.367 ***
Dual	0.001 *	0.000
FIXED	−0.016 ***	−0.214 ***
FirmAge	−0.011 ***	−0.082 ***

Notes: Standard errors in parentheses. *** $p<0.01$, ** $p<0.05$, * $p<0.1$.

.

The first-stage regression indicates that the instrumental variable strongly predicts PI ($\beta =0.796$, $p<0.01$). The Kleibergen–Paap rk Wald F-statistic of 78.105 exceeds conventional thresholds for weak instruments, confirming instrument relevance and strength.

The second-stage coefficient on instrumented PI remains positive and statistically significant ($\beta =1.751$, $p<0.05$). The slightly larger magnitude compared with the baseline estimate suggests that ordinary least squares estimation may underestimate the economic effect due to attenuation bias or measurement error.

These findings strengthen the causal interpretation of the relationship between AI-enabled process capability and enterprise value creation.

4.7. Empirical Channel Tests

To better understand the transmission mechanism linking AI-enabled process capability to enterprise value, we examine two theoretically grounded channels using a mixed-methods approach. Specifically, we investigate:

(1) Internal operational efficiency improvement, illustrated through representative case studies (Figure 5);

(2) External information signaling effects, tested via the quantitative analysis in

Table 7. Mechanism.

	ROE	TFP_LP	ROE	Analyst	ROE
PI	1.624 ***	6.168 ***	1.060 ***	5.359 ***	1.524 ***
	(0.060)	(0.138)	(0.062)	(0.494)	(0.059)
Size	0.066 ***	0.540 ***	0.015 ***	0.810 ***	0.051 ***
Lev	−0.401 ***	0.003	−0.402 ***	−0.899 ***	−0.384 ***
Cashflow	0.370 ***	0.889 ***	0.285 ***	0.759 ***	0.356 ***
Dual	0.000	−0.007	0.001	−0.019	0.000
FIXED	−0.216 ***	−1.052 ***	−0.117 ***	−0.765 ***	−0.202 ***
FirmAge	−0.084 ***	0.206 ***	−0.103 ***	−0.547 ***	−0.073 ***
TFP_LP			0.095 ***
Analyst					0.019 ***
_cons	−1.155 ***	−4.519 ***	−0.735 ***	−14.812 ***	−0.876 ***

Notes: Standard errors in parentheses. *** $p<0.01$.

.

Figure 5 presents descriptive evidence illustrating improvements in AI-enabled industrial output performance, using three dashed lines: blue for Firm A, red for Firm B, and red-purple for Firm C. The observed trend indicates continuous enhancement in digital production capability and intelligent service integration among firms adopting AI technologies.

Figure 5 visually demonstrates the improvement trend in AI-enabled industrial output performance. To explore the driving mechanisms behind this improvement, we further test the mediating effects in Table 7.

These mechanisms are motivated by the Resource-Based View, Dynamic Capability Theory, and Signaling Theory, which emphasize the strategic role of technological capability in generating sustainable competitive advantage.

Operational efficiency channel:

The Process Performance Index exhibits a significantly positive association with total factor productivity (TFP_LP), indicating that AI-enabled process capability improves resource allocation efficiency and production coordination. When TFP_LP is included in the regression, the coefficient of PI decreases but remains statistically significant, indicating partial mediation effects.

External monitoring channel:

PI is positively associated with analyst coverage, suggesting that superior operational capability functions as a credible signal to capital markets. Increased analyst attention improves the information environment and reduces information asymmetry.

These findings indicate that AI-enabled process capability contributes to enterprise value creation through both productivity enhancement and information transparency mechanisms.

4.8. Heterogeneity Analysis

To explore the boundary conditions of the value creation mechanism, we conduct a heterogeneity analysis across ownership structures and corporate life-cycle stages.

Ownership structure analysis:

The positive effect of PI remains statistically significant for both state-owned enterprises (SOEs) and non-state-owned enterprises (Non-SOEs). The similarity in coefficient magnitudes indicates that AI-enabled process capability contributes to enterprise value creation across different institutional governance structures.

Life-cycle analysis:

The effect of PI is strongest among growth-stage firms, followed by mature firms, and remains positive but comparatively smaller for declining firms. This pattern suggests that AI-enabled process capability plays a particularly important role during periods of rapid organizational expansion, when efficient coordination of resources and digital integration are critical for sustaining competitive advantage.

These findings indicate that process capability contributes to enterprise value across diverse institutional environments, with stronger marginal effects observed in dynamic operational contexts.

4.9. Discussion

The empirical results provide consistent evidence that AI-enabled process capability contributes positively to abnormal earnings and enterprise value creation. The Process Performance Index serves as an effective quantitative bridge linking artificial intelligence adoption to firm-level financial performance within the Feltham–Ohlson valuation framework. These findings directly respond to the limitations identified in Section 2, where existing indicator-based, data-driven, and hybrid evaluation approaches often lack an interpretable theoretical linkage between AI-enabled operational capability and enterprise value creation, thereby strengthening the connection between the literature review and empirical discussion.

From a theoretical perspective, the findings extend the Feltham–Ohlson model by incorporating digital operational capability as a measurable driver of abnormal earnings. This integration contributes to the literature on intangible assets by demonstrating that AI-enabled operational capability functions as economically meaningful value-relevant information. Unlike traditional digitalization indicators that primarily capture technology adoption intensity, the proposed Process Performance Index reflects multidimensional operational effectiveness derived from intelligent investment, governance environment, and process output performance, thereby improving interpretability and theoretical consistency between intelligent manufacturing capability and enterprise valuation.

This linkage is particularly important within the context of Industry 5.0 and intelligent manufacturing, where technological development is expected to support not only efficiency improvement but also broader objectives such as human-centric production, resilient operational systems, and sustainable resource allocation. The empirical results indicate that AI-enabled process capability contributes to enhanced coordination between human expertise and intelligent decision systems, supporting human–machine collaborative innovation. In addition, the importance of governance quality, supply chain stability, and operational flexibility demonstrates that resilient process structures are positively associated with enterprise value creation. Improved asset utilization efficiency further indicates that AI-enabled intelligent manufacturing promotes more sustainable resource allocation patterns, contributing to long-term productivity improvement.

From a managerial perspective, the results suggest that investments in intelligent process capability generate both operational efficiency improvements and capital market benefits. Firms that strengthen digital infrastructure, knowledge-based human capital, and governance mechanisms achieve improved productivity and enhanced investor recognition. These findings suggest that intelligent manufacturing should be viewed as a strategic capability development process rather than a short-term automation upgrade, encouraging firms to align AI investment with long-term organizational transformation objectives.

The mechanism analysis further demonstrates that AI-enabled process capability enhances enterprise value through complementary channels involving internal productivity improvement and external information signaling. Internally, intelligent decision support improves resource allocation efficiency and operational coordination, while externally, improved process performance provides credible signals of technological competitiveness and strategic adaptability, strengthening market confidence.

The heterogeneity analysis indicates that process capability is particularly valuable during periods of organizational growth, suggesting that strategic timing of AI investment may influence long-term firm value creation. This result highlights that firms in growth stages may obtain stronger benefits from intelligent manufacturing due to higher flexibility in adopting digital operational capability, further supporting the importance of aligning AI strategy with firm lifecycle characteristics.

4.10. Limitations and Future Research

Despite providing empirical evidence supporting the value relevance of AI-enabled process capability, several limitations should be acknowledged and provide opportunities for future research grounded in existing literature on digital transformation, intelligent manufacturing, and intangible asset valuation [20,21,30,36,37].

First, although the Process Performance Index captures multiple dimensions of intelligent operational capability, the current indicator system primarily relies on firm-level financial and governance proxies commonly used in empirical accounting and digital economy studies. Prior research in Industry 4.0 and cyber-physical systems emphasizes the importance of real-time sensing integration, edge intelligence deployment, and human–AI collaboration mechanisms as critical drivers of intelligent operational performance [38,39]. Future studies may incorporate micro-level indicators derived from IoT-enabled production environments, digital twin systems, and real-time operational analytics to further improve measurement precision and theoretical alignment with intelligent manufacturing literature.

Second, the empirical analysis focuses on listed firms in the Chinese capital market, which has been widely used in studies of digital transformation and AI capability due to high data availability and standardized disclosure structures. However, institutional environments may influence the effectiveness of AI adoption strategies, as suggested in studies on technological capability development and institutional economics [40,41]. Future research may extend the proposed framework to cross-country datasets in order to examine how regulatory systems, technological infrastructure, and market maturity moderate the relationship between AI-enabled process capability and enterprise value creation.

Third, although the instrumental variable approach improves causal inference, additional research designs may further strengthen the identification of the economic impact of AI adoption. Prior empirical studies on technological innovation frequently employ natural experiments, policy shocks, or quasi-experimental methods to address endogeneity concerns [42,43]. Future work may exploit exogenous variations in digital economy policies, industrial AI pilot programs, or regional technology infrastructure development initiatives to further validate the causal mechanism linking AI-enabled operational capability to abnormal earnings.

Fourth, the current framework operationalizes AI capability as a firm-level composite indicator, which provides interpretability but may not fully capture dynamic learning effects emphasized in the dynamic capability literature [44,45]. Future research may integrate longitudinal process learning indicators, adaptive capability metrics, and knowledge diffusion measures to better reflect the evolutionary nature of intelligent operational capability development.

Finally, future studies may explore integration of the proposed framework with emerging research on digital platform ecosystems, human–AI collaborative decision-making, and Industry 5.0 value co-creation mechanisms [46,47]. Extending the hybrid Feltham–Ohlson framework to incorporate dynamic technology diffusion and learning effects may provide deeper theoretical insight into how AI-enabled operational capability contributes to sustained enterprise value creation in intelligent industrial systems.

5. Conclusions

Artificial intelligence is increasingly embedded in industrial processes, influencing investment decisions, operational environments, and production outcomes. This study develops an interpretable evaluation framework that links AI-enabled operational capability to enterprise value creation by integrating the Feltham–Ohlson valuation model with a multi-level fuzzy comprehensive evaluation approach.

The proposed framework evaluates AI-driven transformation across intelligent investment, operational governance environment, and process output dimensions. The constructed Process Performance Index (PI) provides a normalized quantitative indicator of intelligent process capability, enabling systematic measurement of AI-enabled operational performance.

Empirical results show that PI is positively associated with abnormal earnings and firm profitability, confirming that AI-enabled process optimization contributes to sustained enterprise value creation. Mechanism analysis further indicates that AI capability improves productivity efficiency and enhances the external information environment, thereby strengthening both operational performance and market valuation.

The framework contributes a structured bridge between operational performance evaluation and financial valuation theory, offering practical decision support for assessing AI investment effectiveness in the context of Industry 5.0 transformation.