How Does Artificial Intelligence Capability Affect Product Innovation in Manufacturing Enterprises? Evidence from China

Abstract

In today’s fast-changing business environment, artificial intelligence (AI) capability plays a critical role in fostering product innovation (PI). Resource-based theory (RBT) posits that resources and capabilities characterized as valuable, rare, inimitable, and non-substitutable can generate a sustained competitive advantage, providing an appropriate theoretical framework for this study. Using RBT this study examines how business intelligence transforming capability (BITC) mediates the relationship between AI capability and PI and how formal and informal knowledge governance mechanisms (FKGMs and IKGMs, respectively) moderate the effect of AI capability on BITC. Using partial least squares structural equation modeling on 516 Chinese manufacturing enterprises, we empirically test a mediated moderation model. The findings reveal that BITC significantly mediates the relationship between AI capability and PI. Both FKGMs and IKGMs strengthen the effect of AI capability on BITC (with IKGMs showing a stronger influence). This study theoretically contributes by identifying BITC’s mediating role, defining AI capability and BITC boundary conditions, revealing FKGMs’ and IKGMs’ asymmetries, and extending RBT. In terms of practical contributions, the findings emphasize the necessity of developing BITC and strategically applying both FKGMs and IKGMs to maximize AI capability-driven PI benefits.

1. Introduction

Product innovation (PI), characterized by the continuous evolution of intelligent information technologies, is crucial for the survival of manufacturers in dynamic environments [1,2]. As the core of intelligent information technology, artificial intelligence (AI) has the potential to fundamentally transform the design and architecture of new products, creating innovative ways to generate and capture value [3,4,5,6]. AI applications surpass human performance in various activities, while simultaneously highlighting the essential technical competencies required for integrating AI into products within the workforce [7]. Studies have explored how enterprises manage AI to drive iterative PI, such as establishing data governance guidelines, combining offline human expertise with online AI analysis, and implementing standards for intelligent device management [8]. However, other studies demonstrate that the introduction of AI technology by enterprises is unlikely to provide any competitive advantage, as it can be easily acquired and replicated in the market [9,10]. Many researchers believe that one of the primary reasons AI has not yet achieved the expected results is the lag in implementation and reorganization [10].

AI capability enhances an organization’s capacity in regard to the extensive deployment and utilization of AI [10]. This further enables businesses to expand operations and acquire strategic assets, thereby eliminating innovation barriers in key organizational processes [11]. The influence of AI capability on business model innovation, organizational creativity, and open innovation has been extensively studied [12,13]. Although prior studies have investigated the direct relationship between AI capability and PI [14], the specific mechanisms by which AI capability facilitates PI remain poorly understood, representing a critical gap in the literature. In the context of digital transformation, this theoretical gap manifests in practice as a pervasive “data-rich but innovation-poor” paradox, where organizations accumulate vast data yet struggle to effectively transform these data assets into PI [10]. To address these study challenges, this study aims to systematically investigate the mediating mechanisms through which AI capability influences PI, along with its critical boundary conditions.

As discussed above, studies on the influence and mechanisms of AI capability on PI within enterprises remain at an early stage. Current studies exhibit the following limitations. First, although prior studies have established AI’s transformative impact on PI [3,4,5,6,7,8], these studies predominantly treat AI as a monolithic technology, and the causal pathways through which AI capability enables PI remain primarily unexplored [12,13,14,15,16]. Some studies indicate that expanding and optimizing business operating capabilities is critical for enhancing PI [17]. Furthermore, prior studies have established that both AI capability and optimizing business operating capability constitute critical organizational resources for enterprises [18], with knowledge-sharing efficiency playing a vital role in linking these organizational resources [19,20]. Nonetheless, although the existing literature has identified the role of knowledge sharing in facilitating organizational resources [21], the mechanisms that facilitate knowledge sharing of organizational resources have also been overlooked by scholars.

In the process of PI, transforming innovative outcomes into marketable products and continuously optimizing them represent core challenges for enterprises [1]. To address these challenges, scholars argue that expanding and optimizing business operating capabilities are essential for enhancing PI [22]. The development of AI has expanded the boundaries of business operations, positioning AI capability as a vital resource for enterprises to extend their operational frontiers [13,23]. Building on the role of AI capability in operational expansion, capabilities such as BITC enable enterprises to operationalize data-driven insights, thereby simultaneously optimizing business operations and strategically aligning them for PI [24]. Previous studies have extensively examined the impact of enhanced business operating capabilities on the optimization of PI [22,25]. However, only a limited number of studies have focused on the mechanism by which AI capability influences PI via BITC. Hence, our first study question is derived: how does BITC mediate the relationship between AI capability and PI?

The influence of AI capability on BITC is potentially moderated by certain governance mechanisms [17,26]. AI capability and BITC, as critical organizational resources and orientations, serve distinct functions and applications [23,27]. Identifying and applying the appropriate governance mechanisms can help modern enterprises share knowledge across business process technologies [28,29]. In particular, for PI-oriented enterprises that hold a variety of resources, formal and informal knowledge governance mechanisms (FKGMs and IKGMs, respectively) are required to enhance knowledge-sharing capabilities [30]. However, although the value of KGMs in facilitating knowledge sharing is widely recognized [31,32], their moderated mediating role in the relationship between AI capability and BITC remains underexplored. Investigating this role could provide deeper insights into the processes driving PI during AI implementation. Thus, the second study question is derived: how do FKGMs and IKGMs moderate the relationship between AI capability and BITC?

In summary, this study investigates the mediating role of BITC and the influence of FKGMs and IKGMs as boundary conditions on the effect of AI capability on BITC. This study offers three main contributions to the literature. First, it identifies BITC as the mediating factor through which AI capability enhances PI, emphasizing its optimizing business operating role in transforming valuable assets into strategic actionable insights. Second, it develops a framework to explore how FKGMs and IKGMs act as boundary conditions, offering new insights into their influence on the relationship between AI capability and BITC. Third, it extends resource-based theory (RBT) by highlighting how AI capability serves as a strategic asset to drive PI.

The remainder of the paper is organized as follows: Section 2 outlines the theoretical framework and formulates the hypotheses. Section 3 details the methodology, while Section 4 presents the results. Section 5 discusses the theoretical and managerial implications, limitations, and suggestions for future study.

2. Theoretical Framework and Hypothesis Development

2.1. Theoretical Framework

2.1.1. AI Capability, BITC, KGMs, and PI

PI, encompassing the enhancement of existing products alongside the development of new ones, is vital for competitiveness [18,20]. It reflects a shift or evolution in an enterprise’s innovation pathway, requiring evolutionary or iterative PI resources [1]. The product’s nature and the heterogeneous resources required to achieve its differentiated advantages are crucial for PI [33]. Therefore, the ability to acquire unique, valuable, scarce, and difficult-to-imitate heterogeneous resources is essential for enterprises to develop PI [1,34].

AI capability is defined as the ability of an enterprise to integrate AI-based technologies, skills, knowledge, and complementary resources to build a competitive advantage [35]. It encompasses three core components: infrastructure, business spanning, and a proactive stance [35,36,37,38]. According to RBT, distinct capabilities such as technical know-how and managerial acumen are critical components for enterprises to acquire and leverage valuable resources [34,39]. Enterprises can access, acquire, and assist a wide variety of valuable resources such as advanced technologies, data, human expertise, and organizational infrastructure by leveraging their AI capabilities [10,11,40]. Additionally, AI capability can help enterprises secure proprietary data and manage resources [35]. Haefner et al. [9] emphasize that AI capability facilitates data-driven decision-making, optimizes resource allocation, and enhances an enterprise’s ability to detect market trends, all of which contribute to PI.

BITC refers to the ability of business intelligence systems to capture information, analyze real-world relationships, and exploit these relationships to achieve organizational goals, thereby enabling the timely adjustment, integration, and reallocation of resources [18]. It optimizes business operations; enhances the quality of strategic analysis, decision-making, and plan execution; and enables enterprises to identify their strategic orientation in rapidly changing and dynamic markets [36,37,38]. For innovative enterprises, this capability is critical for converting strategic resources into strategic orientation and formulating new action plans [30]. In other words, BITC provides a solid foundation for PI by enabling the precise optimization of business operations [17,19]. AI capability enhances enterprises’ ability to expand business operations by leveraging BITC to align valuable resources with organizational goals, thereby achieving a competitive advantage driven by innovation [10,18,41].

KGMs influence the creation, sharing, integration, utilization, and retention of knowledge [42]. Knowledge is an essential strategic resource to gain intangible resources and competencies [43]. Resources can be released and transformed during the knowledge-sharing process [44]. KGMs are categorized into FKGMs and IKGMs [42]. Both FKGMs and IKGMs are essential for knowledge sharing, enabling enterprises to acquire valuable resources and avoid resource scarcity [31,45,46]. As Zhou and Li (2012) emphasized, enterprises that possess various resources and aim for innovation require both FKGMs and IKGMs to enhance knowledge sharing, thereby unlocking resources [26]. Therefore, by implementing FKGMs and IKGMs, knowledge sharing between AI capability and BITC liberates more resources [47,48,49].

2.1.2. Theoretical Model

RBT emphasizes that an enterprise’s resources and capabilities serve as its foundation for developing new products [34,50]. This theory argues that coordinating resources and capabilities to transform strategic assets into strategic orientations is a prerequisite for enterprises to achieve sustained innovation [34,50,51], especially in dynamic markets [52]. In the era of intelligent systems, exploring how to leverage valuable resources as strategic assets and transform them into strategic orientations [53] as well as investigating the microfoundations underlying this process are crucial [38,50,54].

According to RBT, AI capability enables enterprises to acquire valuable resources and unlock strategic assets, forming the foundation for BITC [18,53]. BITC, in turn, helps convert data into actionable insights and determines the strategic orientation for PI. The efficiency of knowledge sharing is crucial to the transformation of strategic assets into strategic orientations, as its effectiveness directly influences the efficiency of resource flow [55]. The integration of FKGMs and IKGMs facilitates the efficiency of knowledge sharing between AI capability and BITC [56]. Therefore, the moderating role of FKGMs and IKGMs in the relationship between AI capability and BITC is particularly critical. In contrast, once BITC is established, its impact on PI is more direct and execution-driven, relying less on FKGMs and IKGMs [57]. This explains why, in the theoretical model, KGMs only moderate the first-stage path from AI capability to BITC. Consequently, we focus on first-stage moderation in our hypotheses (H2–H3) and empirically test this boundary condition. Based on this, we propose the research model shown in Figure 1.

2.2. Mediating Effect of BITC

AI capability utilizes fact-based support systems to expand business operations, while BITC adjusts enterprises’ resources, processes, and strategies with the extended content from AI capability. This optimizes business operations and strengthens the decision-making process of the enterprise [36,37,38], thereby promoting PI across the entire enterprise. First, AI capability integrates the dynamic design and deployment of AI in the organizational environment, ensuring that the enterprise dynamically adjusts based on real-time resource inputs [35]. This dynamic input enhances the enterprise’s business intelligence, continuously aiding in the transformation and creation or absorption of new resources, optimizing business operations, and providing the latest and most accurate resources for PI [18]. Second, AI capability utilizes simulation technologies, such as digital twins and Monte Carlo simulations, to test the effects of different process designs [58]. The analysis of these simulations enhances BITC in selecting the optimal process solutions, thereby assisting enterprises in adjusting and optimizing their processes [59]. This optimization effect of AI capability on BITC enables enterprises to continuously improve their processes to support PI. Third, AI capability provides real-time information and predictive analytics, enabling enterprises to respond swiftly in dynamic environments and allowing BITC to adjust strategies promptly to address uncertainties encountered during PI implementation [18,60]. Based on these points, we present the following hypothesis:

H1. 

BITC mediates the relationship between AI capability and PI.

2.3. Moderating Effect of FKGMs and IKGMs

2.3.1. Formal Knowledge Government Mechanisms

FKGMs refer to structured guidelines such as performance evaluations, incentives, and reward systems that coordinate, direct, and control knowledge-sharing processes to create value [31]. These mechanisms can be viewed as the degree of formalization or standardization within an organization. Enterprises empowered by FKGMs exhibit standardized decision-making processes that allow managers to structure, govern, and manage the flow of knowledge, ensuring that decisions are based on accurate, reliable, and timely information [25,61]. When FKGMs are highly standardized, they facilitate collaboration, enhancing the efficiency and effectiveness of knowledge sharing [28].

FKGMs align specialists’ goals and foster collaboration by establishing clear standardized procedures for exchanging knowledge and decision-making [62]. This alignment reduces organizational friction and promotes cohesion [63]. Consequently, knowledge flows more smoothly from AI to business intelligence, enhancing the integration of AI capability into BITC. In addition, structured and institutionalized FKGMs help minimize information asymmetry by providing a standardized shared understanding that improves the effectiveness of large knowledge transfers [64]. By overcoming information asymmetry, FKGMs mitigate key challenges, such as path dependence, causal ambiguity, and resource asymmetry [31,62]. This facilitation of knowledge sharing strengthens the presumed benefits of AI capability for BITC. For example, Huawei’s standardized knowledge-sharing processes have effectively integrated AI capabilities into its business intelligence systems. This demonstrates how FKGMs can strengthen the relationship between AI capability and BITC by reducing information asymmetry and promoting systematic knowledge sharing. Consequently, we propose the following hypothesis:

H2. 

FKGMs positively moderate the effect of AI capability on BITC.

2.3.2. Informal Knowledge Government Mechanisms

IKGMs refer to the unstructured, non-formalized practices, norms, and social interactions within an organization that influence the creation, sharing, and use of knowledge [31]. These mechanisms are typically based on relationships, trust, shared values, and social interactions rather than formal policies and structured systems [36,65]. IKGMs include activities such as mentoring, networking, informal meetings, water cooler conversations, and other social gatherings that facilitate knowledge sharing and create an environment supportive of collaboration and information exchange [31]. They are particularly essential for extracting and sharing tacit knowledge [31]. Thus, they complement formal governance structures by fostering a knowledge-sharing culture in which knowledge flows freely.

IKGMs influence the effectiveness of knowledge sharing between AI and BITC, facilitating a stronger alignment and integration of AI capability and BITC [66]. The social interactions encouraged by IKGMs help build trust and interpersonal relationships, which are essential for effective knowledge exchange [67]. These mechanisms allow specialists to share tacit knowledge, insights, and practical experiences that are difficult to codify or communicate using formal systems [31]. As employees from the AI and business intelligence domains collaborate informally, they bridge gaps in understanding and foster mutual respect, which enhances the integration of AI capability into business intelligence processes [61]. This improved collaboration leads to better alignment between AI and BITC [43]. For instance, Alibaba’s emphasis on cross-departmental collaboration, mentorship programs, and a culture of trust has facilitated the flow of tacit knowledge between AI teams and business units. This has enabled the development of more accurate recommendation algorithms, showcasing how IKGMs can enhance the integration of AI capabilities into business intelligence systems by fostering trust and collaboration. Based on this, we present the third hypothesis:

H3. 

IKGMs positively moderate the effect of AI capability on BITC.

3. Methodology

3.1. Methodology Selection and Rationale

Survey data can be analyzed using various methods, such as regression, covariance-based structural equation modeling (CB-SEM), and partial least squares structural equation modeling (PLS-SEM). This study employed PLS-SEM for two primary reasons. First, because this study aims to develop a predictive theory of how AI capability influences PI through BITC, PLS-SEM is more appropriate than regression analysis. Unlike regression, which focuses solely on parameter significance, PLS-SEM maximizes the explained variance of dependent variables (i.e., PI), aligning with our goal of identifying key drivers of innovation outcomes [68]. Second, PLS-SEM can simultaneously evaluate complex model structures with both mediation and moderation effects in a single analytical framework. By contrast, traditional regression requires separate models for each path, increasing the risk of Type I errors and failing to account for interdependencies among constructs [69]. By integrating these analyses, PLS-SEM provides a more holistic and statistically robust test of our hypotheses.

3.2. Data Collection and Sample

To test our three hypotheses, we collected survey data from manufacturing enterprises in China. According to the latest data released by the Central People’s Government of the People’s Republic of China, as of August 2024, the country had 6.03 million manufacturing enterprises. Priority was given to enterprises actively engaged in advanced technologies such as AI to ensure the sample’s high relevance to the study objectives. Surveys were distributed both online and in person, with follow-up conducted to improve the response rate. Preliminary data validation checks confirmed the accuracy of the sample in terms of sector and organizational representation. This approach ensured a robust and representative sample, providing a solid foundation for analyzing AI capability, business intelligence, and PI in China.

To ensure the randomness and representativeness of the sample, we employed a computer-generated random number system to select samples within each stratum, ensuring true randomness in the selection process. First, we stratified the manufacturers into multiple layers based on enterprise size (large, medium, and small) and industry sector (i.e., automotive, electronics, machinery, and pharmaceuticals). Subsequently, we randomly selected samples in proportion to each stratum to ensure the representativeness of each size and industry combination (e.g., large automotive enterprises and medium-sized electronics enterprises). Finally, considering the study objectives and resource constraints, we determined the total sample size (700 enterprises) and ensured the diversity and balance of the sample during the sampling process. This approach reduced sample bias and enhanced the generalizability and reliability of the study findings.

To avoid cultural bias and ensure validity, we followed a standard procedure. First, the questionnaire was developed in English and then translated into Chinese by bilingual professionals with extensive research experience in the relevant field. Subsequently, the Chinese version was back-translated into English, and the research team compared it with the original English version to identify and resolve any potential misunderstandings that might have occurred during the translation process. Prior to conducting the formal survey, a pretest was performed using in-depth interviews with six senior executives and five experienced researchers to validate the content, clarity, and linguistic appropriateness of the questionnaire. Each interview lasted approximately 45 to 60 min. Certain problematic items were revised based on the feedback provided by the respondents. Subsequently, an e-mail invitation was sent to key informants from randomly selected manufacturing enterprises, followed by a phone call to reiterate the study objectives and ensure the confidentiality and anonymity of potential participants.

To mitigate common method bias (CMB), enhance causal inference capabilities, and improve the accuracy and reliability of the data, we adopted the multi-wave, multi-source data collection approach, as suggested by Podsakoff et al. [70]. This study was divided into three phases, with each phase spaced two weeks apart. To encourage participants to complete the questionnaires, rewards or incentives were offered upon completion of each of the three phases. To reduce informant bias, we sampled respondents who play similar roles within their respective companies and explicitly stated that there are no right or wrong answers, ensuring the anonymity of respondents. The first phase involved collecting data on basic employee information, AI capability, FKGMs, and IKGMs. The second phase focused on collecting data on BITC, and the third phase involved collecting data on PI and marker variables. The survey items were organized into general thematic sections rather than grouped by construct to prevent respondents from identifying specific constructs or guessing hypothesized relationships. This design helped reduce CMB, enhanced causal inference capabilities, and improved the accuracy and reliability of the data.

Following Chan and Wong’s (2019) [71] recommendations, we performed Power and N Computations for mediation using the MedPower program (accessed in March 2025). We assumed a standardized path coefficient of 0.15 for both the first stage (AI Capability to BITC) and the second stage (BITC to PI) of the mediation effect. The mediation effect size was calculated as 0.0225. With α = 0.05 and power = 0.8 [72], the required sample size was estimated to be 451. We obtained data from 516 enterprises in 2025, achieving a valid response rate of 73.7%, which not only meets but exceeds the calculated requirement.

Table 1. Sample demographics.

Variable	Attribution	Frequency	Percent
Position	CEO	96	18.6
	SVP	156	30.2
	CIO	140	27.1
	CSO	124	24.0
Working years	Less than one year	95	18.4
	1–2 years	94	18.2
	2–3 years	110	21.3
	3–4 years	129	25.0
	Over four years	88	17.1
Industry	Pharmaceutical manufacturing sector	130	25.2
	Automobile manufacturing industry	120	23.3
	Electrical machinery and equipment manufacturing industry	79	15.3
	Computer, communication, and electronic equipment manufacturing industry	127	24.6
	Others	60	11.6
Established years	Less than three years	67	13.0
	3–5 years	104	20.2
	6–10 years	130	25.2
	11–15 years	138	26.7
	Over 15 years	77	14.9
Enterprise size	Less than 300 employees	103	20.0
	300–500 employees	87	16.9
	500–1000 employees	137	26.6
	1000–2000 employees	122	23.6
	More than 2000 employees	67	13.0

Note: CEO = Chief Executive Officer. SVP = Senior Vice President. CIO = Chief Information Officer. CSO = Chief Strategy Officer.

presents the respondents’ demographics. To address non-response bias, we compared responding enterprises with non-responding enterprises and found no significant differences in terms of enterprise size and age. Table 1 presents the descriptive statistics of the sample demographics.

3.3. Measures

The dependent variable, PI, was assessed using the measurement from Castro et al. (2013) [73], comprising three items that assess an organization’s ability to render current product lines obsolete. The independent variable, AI capability, was assessed based on the scale developed by Mikalef et al. (2023) [35]. The scale encompasses 12 items evaluating AI capability across three aspects: infrastructure, business spanning, and a proactive stance. The mediating variable, BITC, was gauged using the scale from Chen and Lin (2021) [23], which includes six items that measure BITC across two dimensions, including upgrading capability and regeneration capability. The moderating variables, FKGMs and IKGMs, were assessed using the six-item scale developed by Huang et al. (2013) [31]. All these measures use a five-point Likert scale with responses ranging from “strongly disagree” to “strongly agree.” The survey scales and corresponding items are presented in

Table 2. Survey scales and items.

Variable	Scale	Items
AI capability [35]	Infrastructure	AI Capability1: Relative to our industry rivals, our organization has data management services and architectures for AI.
		AI Capability2: All network communication services and cloud services are connected to the central office for analytics.
		AI Capability3: Our organization utilizes AI application portfolio and services (i.e., Microsoft Cognitive Services, Google Cloud Vision).
		AI Capability4: Our organization has AI facilities’ operations/services (i.e., servers, large-scale processors, performance monitors) to ensure that data is secured from to end to end with state-of-the-art technology.
	Business spanning	AI Capability5: Developing a clear vision regarding how AI contributes to business value.
		AI Capability6: Integrating business strategic planning and AI planning.
		AI Capability7: Enabling functional area and general management’s ability to understand value of AI investments.
		AI Capability8: Establishing an effective and flexible AI planning process and developing a robust AI plan.
	Proactive stance	AI Capability9: Our organization are capable of and continue to experiment with new AI tools and techniques as necessary.
		AI Capability10: Our organization have a climate that is supportive of trying out new ways of using AI.
		AI Capability11: Our organization constantly seek new ways to enhance the effectiveness of AI use.
		AI Capability12: Our organization constantly keep current with new AI innovations.
BITC [23]	Upgrading capability	BITC1: BI assists in setting a favourable position and exploring new opportunities in a turbulent environment.
		BITC2: BI assists in continuously creating or absorbing new knowledge, developing new products, or innovating business processes.
		BITC3: BI assists in discerning and integrating new knowledge through exogenous sources of external network and social capital.
	Regeneration capability	BITC4: BI can prompt the reallocation of available resources in line with strategic goals and fully utilize the knowledge for organizational changes.
		BITC5: BI improves the capability of optimizing the allocation and utilization of resources in light of new business practices.
		BITC6: BI assists in enhancing organizational learning to capture, create, and utilize new capabilities.
FKGMs [31]		FKGM1: Knowledge sharing is an index of performance evaluation and rewards.
		FKGM2: Experts are invited as instructors in internal training.
		FKGM3: There are company newsletter or journal to encourage knowledge sharing.
IKGMs [31]		IKGM1: There are water-cooler, coffee lounge for colleagues to make friendship.
		IKGM2: There are leisure activities for colleagues to make friendship.
		IKGM3: There are athletic team or birthday party for colleagues to make friendship.
PI [74]		PI1: In the last three years, the number of product innovations developed by our organization is higher than my competitors’.
		PI2: The percentage of sales with respect to new products, on the total of sales, is higher than the one of my competitors.
		PI3: In the last three years, the number of new products with respect to my product portfolio is higher than the one of my competitors.

Note: Numbers in square brackets (e.g., Ref. [35]) refer to the sources of measurement scales, as listed in the References Section.

.

3.4. Control Variables

Based on an extensive review of the empirical literature on the determinants of innovation outcomes [69], we included five control variables that influence the relationship between AI capability and PI among manufacturers. Position was a dummy variable measured based on the respondents’ roles. As Damanpour [74] highlights, there are significant differences in innovation behavior and performance among employees in different positions within an organization. Therefore, we included position as a dummy variable in our study to control for the impact of positional differences on the study outcomes. Working years, defined as the tenure of the respondents in AI and innovation management, was included as a control variable because it can influence an individual’s adaptability to change and innovation outcomes. Longer tenures may indicate greater experience but could also be associated with resistance to change [75,76]. Working years was assessed using five categories: less than 1 year, 1 to 2 years, 2 to 3 years, 3 to 4 years, and more than 4 years [17]. Industry was included as a dummy variable because the industry context significantly influences an enterprise’s responsiveness to deployment strategies and technological capabilities. Different sectors have varying levels of technological intensity, market competition, and regulatory environments, which can affect innovation outcomes [76]. The selection of established years and enterprise size as control factors was based on their direct relevance to AI capability and innovation performance. A prior study suggests that enterprise age and size influence the availability of internal resources, absorptive capacity, and decision-making structures, which are critical to AI-driven innovation [9]. The variable established years was categorized as follows: less than 3 years, 3–5 years, 6–10 years, 11–15 years, and over 15 years. Enterprise size was measured based on the number of employees and categorized into five groups: less than 300, 300 to 500, 501 to 1000, 1001 to 2000, and over 2000 employees.

3.5. Evaluation of Common Method Bias

This study employed three methods to assess common method bias (CMB) in the survey data: Harman’s single-factor test [70], a combined approach of the marker variable technique and Confirmatory Factor Analysis (CFA) [77], and full collinearity testing [78].

First, it employed Harman’s single-factor test to assess CMB. All items included in this study’s scales were subjected to factor analysis using the principal component analysis method. The results showed that the variance explained by the first unrotated factor was 23.364%, which is below the threshold of 50% (see

Table 3. Harman’s single-factor test.

Component	Initial Eigenvalues
	Total	% of Variance	Cumulative%
1	7.944	23.364	23.364
2	4.514	13.276	36.640
3	2.794	8.217	44.858
4	2.084	6.128	50.986
5	1.641	4.825	55.811
6	1.419	4.173	59.983
7	1.303	3.831	63.815
8	1.195	3.514	67.329

), indicating no significant CMB.

Second, a combined approach using the marker variable technique and CFA was employed to systematically assess CMB. In the questionnaire design, this study introduced a marker variable theoretically unrelated to other variables, namely, “attitude toward the color blue” (ATT). The Cronbach’s alpha for this marker variable was 0.903, indicating good reliability. The correlation coefficients between the marker variable and other variables are presented in

Table 4. Pearson correlation analysis.

Variable	M	SD	1	2	3	4	5	6	7	8	9	10	11
1 INF	3.858	0.899	1
2 BS	3.676	0.972	0.477 **	1
3 PS	3.735	0.923	0.409 **	0.522 **	1
4 AI Capability	3.756	0.750	-	-	-	1
5 UC	3.764	0.924	0.305 **	0.351 **	0.375 **	0.428 **	1
6 RC	3.850	0.876	0.285 **	0.348 **	0.384 **	0.422 **	0.531 **	1
7 BITC	3.807	0.788	0.337 **	0.400 **	0.433 **	0.486 **	-	-	1
8 FKGMs	3.841	0.878	0.068	0.215 **	0.149 **	0.181 **	0.196 **	0.215 **	0.234 **	1
9 IKGMs	3.909	0.872	0.136 **	0.271 **	0.200 **	0.254 **	0.202 **	0.310 **	0.291 **	0.464 **	1
10 PI	3.621	0.955	0.359 **	0.375 **	0.365 **	0.456 **	0.363 **	0.414 **	0.443 **	0.159 **	0.195 **	1
11 ATT	5.592	0.891	0.017	0.013	−0.009	0.009	0.013	0.018	0.017	−0.005	−0.016	−0.006	1

Note: ** = p < 0.01. INF = Infrastructure. BS = Business Spanning. PS = Proactive Stance. UC = Upgrading Capability. RC = Regeneration Capability. ATT = Attitude Toward The Color Blue. Correlations between subscales and total scores are omitted (-) due to inherent conceptual overlap in the measurement model.

. The results indicate that all correlations were non-significant, further suggesting no severe CMB.

Furthermore, this study systematically tested for common method variance (CMV) using CFA [77]. As shown in

Table 5. Results of CMB testing using the marker variable technique.

Fit	χ2	df	χ2/df	RMSEA	SRMR	CFI	TLI
1 CFA	803.334	491	1.636	0.035	0.032	0.963	0.958
2 Baseline	804.465	512	1.571	0.033	0.032	0.965	0.962
3 Method-C	804.430	511	1.574	0.033	0.032	0.965	0.962
4 Method-U	785.910	485	1.620	0.035	0.030	0.964	0.959
5 Method-R	785.918	521	1.508	0.031	0.030	0.969	0.966
Chi-Square Model Comparison Tests
ΔModels	Δχ2	Δdf	p Value
1 Baseline vs. Method-C	0.035	1	0.852
2 Method-C vs. Method-U	18.520	26	0.856
3 Method-U vs. Method-R	0.008	36	1.000

Note: CMB = Common Method Bias. CFA = Confirmatory Factor Analysis. RMSEA = Root Mean Square Error of Approximation. SRMR = Standardized Root Mean Square Residual. CFI = Comparative Fit Index. TLI = Tucker–Lewis Index.

, the analysis results indicate that both the baseline CFA model (χ²/df = 1.636, CFI = 0.963, RMSEA = 0.035) and the baseline model (CFI = 0.965, RMSEA = 0.033) met the criteria for good model fit (CFI > 0.95, RMSEA < 0.06). The Method-C model introduces four method factors with loadings constrained to be equal. Method-U adopts unconstrained loadings for the method factors, where ‘U’ indicates this unconstrained characteristic. The Method-R model, where ’R’ denotes restricted parameters, examines potential bias arising from marker variable method variance that may influence factor correlations or structural parameters [77]. Model comparisons revealed the following: The chi-square difference between the Method-C model and the baseline model was not significant (Δχ² = 0.035, Δdf = 1, p = 0.852), indicating no significant method effect. The difference between the Method-U and Method-C models was also not significant (Δχ² = 18.520, Δdf = 26, p = 0.856), supporting the hypothesis of equivalent method effects (i.e., the CMV model holds). The difference between the Method-R and Method-U models did not reach significance (Δχ² = 0.008, Δdf = 36, p = 1.000), suggesting that the method effect did not introduce systematic bias into the relationships among latent variables.

Finally, following the approach proposed by Kock [78], this study employed full collinearity testing to identify CMB. According to this method, if the Variance Inflation Factor (VIF) for all latent variables is less than 3.3, this study does not exhibit severe CMB. As shown in

Table 6. CMB testing based on full collinearity assessment.

Variable	VIF
Infrastructure (INF)	1.059
Proactive Stance (PS)	1.015
Business Spanning (BS)	1.172
Regeneration Capability (RC)	1.192
Upgrading Capability (UC)	1.029
FKGMs	1.041
IKGMs	1.030
PI	1.043

Note: VIF = Variance Inflation Factor.

, the maximum VIF value for all first-order latent variables in this study was 1.192, indicating no significant CMB.

This study also conducted normality and multicollinearity tests to further validate the multivariate assumptions. First, regarding the normality test, based on the criteria proposed by Kline [79], if the absolute values of skewness are ≤3.0 and the absolute values of kurtosis are ≤10.0, the distribution is not severely non-normal. In this study, the absolute values of skewness and kurtosis for all variables were less than 3, indicating that the data can be considered normally distributed. Second, the multicollinearity test was conducted based on two steps. In the first step, the VIF and tolerance were used to assess multicollinearity. The results demonstrate that the maximum VIF value for all predictor variables was 1.654, which is below the recommended threshold, suggesting no severe multicollinearity among the independent variables. Subsequently, this study further examined multicollinearity using correlation analysis. The maximum Pearson correlation coefficient among the independent variables was 0.487, which is below the standard threshold of 0.7 [80]. No severe multicollinearity issue was observed (Table 4).

3.6. Assessment of the Measurement Model

This study employed a formative second-order construct with reflective first-order dimensions for AI capability, as suggested by Edwards [81]. This approach allowed us to capture the distinct subdimensions of AI capability (i.e., infrastructure, business spanning, and a proactive stance) while maintaining the reflective nature of the first-order constructs. We used SPSS 26.0 and SmartPLS 4 for statistical data analysis. Consistent with Hair et al.’s [81] guidelines, we assessed internal consistency (Cronbach’s alpha), composite reliability (CR), convergent validity (average variance extracted, AVE), and outer loadings of the indicators [80]. In the model development process, the specific aspects represented by the first-order constructs are crucial for the measurement validity of the overall construct. Previous studies have suggested that the validation of the first-order constructs serves as a prerequisite for the validity of the second-order constructs [81]. Therefore, in this study, both the first- and second-order constructs were validated separately, providing a more solid foundation for the final study conclusions.

First, the reliability of the measurement models was assessed, with both the first- and the second-order Cronbach’s alpha values significantly exceeding 0.7, indicating strong consistency [82]. Next, convergent validity was assessed by examining the outer loadings (all exceeding 0.7) and AVE values (all above 0.5). These results indicate strong convergent validity for both the first- and second-order factors, as presented in

Table 7. Measurement outer loadings, reliability, and validity.

Construct	Items	Outer Loading	CR	AVE	Cronbach’s Alpha
First Order
Infrastructure (INF)	INF1	0.751	0.891	0.671	0.836
	INF2	0.842
	INF3	0.847
	INF4	0.833
Business Spanning (BS)	BS1	0.859	0.911	0.719	0.869
	BS2	0.810
	BS3	0.836
	BS4	0.885
Proactive Stance (PS)	PS1	0.860	0.905	0.704	0.859
	PS2	0.829
	PS3	0.810
	PS4	0.855
Upgrading Capability (UC)	UC1	0.864	0.902	0.754	0.837
	UC2	0.882
	UC3	0.859
Regeneration Capability (RC)	RC1	0.853	0.879	0.708	0.793
	RC2	0.850
	RC3	0.820
FKGMs	FKGM1	0.830	0.869	0.689	0.775
	FKGM2	0.885
	FKGM3	0.771
IKGMs	IKGM1	0.786	0.881	0.713	0.797
	IKGM2	0.868
	IKGM3	0.875
PI	PI1	0.815	0.887	0.724	0.809
	PI2	0.857
	PI3	0.879
Second Order
AI Capability	INF	0.773
	BS	0.838	0.889	0.847	0.649
	PS	0.805
BITC	UC	0.873
	RC	0.878	0.843	0.868	0.766

Note: CR = Composite Reliability. AVE = Average Variance Extracted.

.

Second, the heterotrait-to-monotrait (HTMT) ratio and the Fornell–Larcker criterion were employed to assess the discriminant validity among the variables [83]. The analysis results are as follows (see

Table 8. HTMT discriminant validity test.

	1	2	3	4	5	6	7	8	9	10
Lower Order
1 INF	0.819	0.561	0.484	0.364	0.348	0.085	0.168	0.438	-	0.440
2 BS	0.481	0.848	0.604	0.411	0.417	0.261	0.323	0.448	-	0.512
3 PS	0.412	0.525	0.839	0.441	0.461	0.185	0.241	0.437	-	0.558
4 UC	0.304	0.351	0.375	0.869	0.653	0.243	0.249	0.442	0.546	-
5 RC	0.284	0.346	0.381	0.533	0.841	0.271	0.384	0.515	0.551	-
6 FKGMs	0.065	0.211	0.147	0.198	0.219	0.830	0.589	0.202	0.239	0.318
7 IKGMs	0.137	0.269	0.200	0.203	0.305	0.461	0.844	0.243	0.329	0.390
8 PI	0.361	0.378	0.365	0.364	0.413	0.163	0.197	0.851	0.594	0.591
Higher Order
9 AI Capability	-	-	-	0.427	0.419	0.177	0.253	0.457	0.806	0.678
10 BITC	0.336	0.398	0.432	-	-	0.238	0.291	0.444	0.484	0.875

Note: INF = Infrastructure. BS = Business Spanning. PS = Proactive Stance. UC = Upgrading Capability. RC = Regeneration Capability. Correlations between subscales and total scores are omitted (-) due to inherent conceptual overlap in the measurement model.

). The HTMT ratio was used to test discriminant validity. The HTMT ratios between all constructs were below the threshold of 0.85, indicating that the variables in this study can be well distinguished. Moreover, the Fornell–Larcker criterion was applied to further examine discriminant validity. The results show that the square root of the AVE for each variable was greater than the correlation coefficients between the variables, suggesting good discriminant validity among the variables in this study.

3.7. Hypothesis Testing

3.7.1. Evaluation of the Structural Model

To account for the measurement structure of higher-order constructs, this study employed Smart PLS 4.1 and adopted the embedded two-stage approach for hypothesis testing [84]. The first-stage path model (Figure 2) was constructed using the repeated indicators approach.

The second-stage path model includes the total effect model, the mediation effect model, and the moderation effect model. After performing bootstrapping with 5000 resamples, the analysis results of Model 1 reveal that AI capability has a significant positive impact on PI (β = 0.457, p < 0.001). Therefore, the baseline relationship is established. AI capability has a significant positive impact on BITC (β = 0.484, p < 0.001), and BITC has a significant positive impact on PI (β = 0.291, p < 0.001).

First, to ensure robust estimations, this study conducted collinearity tests on the models, and the results show that the maximum VIF for all predictor variables in the models was 1.431, which is below the threshold of 5. This indicates that no significant multicollinearity exists among the endogenous variables. The hypotheses were tested using 5000 bootstrap resamples with bias-corrected confidence intervals.

Second, model fit was assessed using the coefficient of determination (R²), which quantifies the variance explained in endogenous constructs by exogenous variables. The R² value for the endogenous latent variable PI in Model 1 (Figure 3) is 0.209. The R² values for the endogenous latent variable BITC in Models 2 (Figure 4) and 3 are 0.272 and 0.337, respectively, while the R² values for PI in Models 2 (Figure 4) and 3 are 0.275 and 0.284, respectively.

Third, we assess effect sizes using Cohen’s ƒ², which quantifies the contribution of individual exogenous constructs to R². The effect size analysis reveals small-to-medium effects across significant paths (ƒ² = 0.011–0.267). The strongest effect was observed for Model 3 (Figure 5; AI capability–BITC; ƒ² = 0.267).

Fourth, the predictive relevance (Stone-Geisser Q²) was assessed to evaluate the model’s predictive power. The Q² value for the endogenous latent variable PI in Model 1 (Figure 3) is 0.149, while those for PI in Models 2 and 3 (Figure 4 and Figure 5) are 0.194 and 0.195, respectively. For the endogenous latent variable BITC, the Q² values in Models 2 and 3 (Figure 4 and Figure 5) are 0.199 and 0.245, respectively. A Q² value greater than 0 indicates that the model has predictive relevance for the endogenous construct, while a value of 0 or below suggests a lack of predictive relevance.

Fifth, the PLSpredict results are presented in

Table 9. PLSpredict results.

	Model 1			Model 2			Model3
	Q2 Predict	PLS-LM		Q2 Predict	PLS-LM		Q2 Predict	PLS-LM
		ΔRMSE	ΔMAE		ΔRMSE	ΔMAE		ΔRMSE	ΔMAE
PLSPredict LV summary
BITC	-			0.259			0.307
PI	0.202			0.202			0.197
PLSPredict MV summary
RC	-	-	-	0.210	0.002	0.000	0.239	−0.015	−0.025
UC	-	-	-	0.185	−0.007	−0.002	0.231	−0.032	−0.021
PI1	0.135	−0.005	−0.005	0.130	−0.012	−0.015	0.127	−0.010	−0.011
PI2	0.134	−0.005	−0.004	0.136	−0.013	−0.011	0.131	−0.010	−0.008
PI3	0.170	−0.005	−0.004	0.173	−0.002	0.001	0.171	0.000	0.002

Note: UC = Upgrading Capability RC = Regeneration Capability. Correlations between subscales and total scores are omitted (-) due to inherent conceptual overlap in the measurement model.

. (1) The Q² predictive values for BITC and PI in Models 1–3 are all greater than 0, indicating that the models have predictive validity. (2) The Q² predictive values for all measurement indicators are also greater than 0, demonstrating their predictive relevance. (3) Following the criteria of Hair et al. [64], the RMSE and MAE values for the majority of the measurement indicators in the PLS-SEM model are lower than those of the linear regression model benchmark, further validating the superior predictive performance of the model.

3.7.2. Evaluation of the Mediating Effect

As shown in

Table 10. Assessment of the mediation effect.

	Std. Estimate	S.E.	95% CI
			2.50%	97.50%
Total Effect	0.457	0.049	0.354	0.545
Direct Effect	0.316	0.058	0.197	0.421
AI Capability → BITC → PI	0.141	0.027	0.091	0.199

, for the mediating role of BITC between AI capability and PI, the analysis results indicate that the 95% confidence interval for the mediation effect is [0.091, 0.199]. Because the confidence interval does not include 0, it suggests that BITC has a significant mediating effect between AI capability and PI, with a standardized mediation effect size of 0.141. Thus, hypothesis H1 is supported.

3.7.3. Evaluation of the Moderation Effect

Figure 6 shows that FKGMs exhibit a significant positive moderating effect on the relationship between AI capability and BITC (p = 0.020, β = 0.103). Figure 7 shows that IKGMs exhibit a significant positive moderating effect on the relationship between AI capability and BITC (p = 0.010, β = 0.155). Compared with FKGMs, IKGMs exhibit a stronger and more significant moderating effect on the path from AI capability to BITC. Therefore, hypotheses H2 and H3 are supported.

3.7.4. Robustness Tests and Endogeneity

First, we examined robustness by constructing an extended model (Model 2) with interaction terms and comparing its results against the baseline model (Model 1). As shown in

Table 11. Results of the study model.

Path	Std. Estimate	S.E.	t	p	95%CI		f2	VIF
					2.50%	97.50%
Model 1: Mediation Model; R2 (BITC = 0.272, PI = 0.275); Q2 (BITC = 0.199, PI = 0.194)
AI Capability → BITC	0.432	0.050	8.552	0.000	0.328	0.527	0.238	1.074
AI Capability → PI	0.310	0.059	5.276	0.000	0.190	0.418	0.100	1.329
BITC → PI	0.280	0.053	5.286	0.000	0.179	0.388	0.079	1.374
FKGM → BITC	0.099	0.046	2.135	0.033	0.003	0.183	0.011	1.276
FKGM → PI	0.031	0.049	0.636	0.525	−0.068	0.126	0.001	1.289
IKGM → BITC	0.137	0.056	2.460	0.014	0.031	0.244	0.019	1.320
IKGM → PI	0.022	0.046	0.491	0.623	−0.065	0.112	0.001	1.346
Model 2: Interaction Model; R2 (BITC = 0.337, PI = 0.284); Q2 (BITC = 0.245, PI = 0.195)
AI Capability → BITC	0.436	0.042	10.305	0.000	0.351	0.517	0.267	1.077
AI Capability → PI	0.352	0.057	6.180	0.000	0.233	0.458	0.098	1.762
BITC → PI	0.230	0.056	4.081	0.000	0.120	0.342	0.037	1.974
FKGM → BITC	0.116	0.045	2.566	0.010	0.022	0.199	0.015	1.303
FKGM → PI	0.035	0.044	0.794	0.427	−0.054	0.121	0.001	1.344
IKGM → BITC	0.160	0.053	3.037	0.002	0.062	0.265	0.029	1.345
IKGM → PI	0.013	0.043	0.298	0.766	−0.068	0.099	0.000	1.426
IKGM × AI Capability → BITC	0.155	0.060	2.590	0.010	0.044	0.278	0.031	1.430
IKGM × AI Capability → PI	−0.034	0.063	0.539	0.590	−0.157	0.092	0.001	2.282
FKGM × AI Capability → BITC	0.103	0.044	2.341	0.019	0.018	0.191	0.015	1.427
FKGM × AI Capability → PI	0.102	0.068	1.498	0.134	−0.035	0.230	0.009	2.284
IKGM × BITC → PI	−0.009	0.052	0.182	0.856	−0.116	0.088	0.000	2.545
FKGM × BITC → PI	−0.072	0.062	1.175	0.240	−0.194	0.046	0.005	2.444

Note: VIF = Variance Inflation Factor.

, the direction and significance of the coefficients in the core pathway (AI capability → BITC → PI) remain consistent, supporting the robustness of our hypotheses. However, the non-significance of certain interaction terms (i.e., IKGM × AI capability → PI) suggests that moderating effects may be pathway-specific.

Second, this study followed the procedure proposed by Hult et al. [85] to address endogeneity issues in PLS-SEM. First, omitted variable bias is one of the most common sources of endogeneity. To mitigate this, a series of control variables grounded in prior theoretical foundations, including position (dummy variable), industry (dummy variable), work years, established years, and enterprise size, were incorporated into the PLS-SEM model. Second, the Gaussian Copula method, recommended by Hair et al. [79], was employed to test for endogeneity. The Gaussian Copula method directly models the correlation between the error terms of the antecedent constructs and the endogenous constructs. The results of the model incorporating all Gaussian Copula paths are presented in the table below. None of the Gaussian Copula paths are significant (p > 0.05), indicating that the model (with control variables) does not exhibit significant endogeneity issues (see

Table 12. Endogeneity tests.

Path	Std. Estimate	S.E.	t	p
AI Capability → BITC	0.747	0.232	3.213	0.001
AI Capability → PI	0.562	0.263	2.137	0.033
BITC → PI	0.436	0.145	3.003	0.003
FKGM → BITC	0.135	0.111	1.223	0.221
FKGM → PI	0.079	0.119	0.662	0.508
IKGM → BITC	0.084	0.112	0.747	0.455
IKGM → PI	0.075	0.113	0.667	0.505
GC (AI Capability) → BITC	−0.323	0.233	1.389	0.165
GC (AI Capability) → PI	0.375	0.253	1.485	0.138
GC (FKGM) → BITC	−0.034	0.086	0.399	0.690
GC (FKGM) → PI	−0.045	0.100	0.446	0.656
GC (IKGM) → BITC	0.046	0.078	0.591	0.554
GC (IKGM) → PI	−0.049	0.094	0.516	0.606
GC (BITC) → PI	−0.140	0.117	1.196	0.232

Note: Gaussian Copula.

).

4. Results

This study developed a moderated mediation theoretical model to examine the mediating role of BITC in the relationship between AI capability and PI while introducing FKGMs and IKGMs as moderators of the AI capability–BITC linkage. Through multi-stage path models, the mediation and moderation effects among variables were empirically validated. The results demonstrate that AI capability enhances PI via the mediating pathway of BITC. Furthermore, both FKGMs and IKGMs significantly amplify the positive impact of AI capability on BITC, with the moderating effect of IKGMs being more pronounced. The research hypotheses, namely H1 (BITC’s mediation effect), H2 (FKGMs’ moderating role), and H3 (IKGMs’ moderating role), were fully supported. Robustness tests and endogeneity controls confirmed the reliability of the core pathways. By identifying BITC as the critical strategic pivot, this study reveals that FKGMs and IKGMs are key strategic levers between AI capability and BITC, thereby reconstructing the digital transformation pathway for enterprises. Furthermore, it highlights the necessity for policymakers to implement BITC-enhancing policies, including differentiated subsidies and refined incentive mechanisms for FKGMs and IKGMs.

5. Discussion

First, BITC positively mediates the relationship between AI capability and PI. Previous studies have clarified that the relationship between AI capability and innovation requires effective transformation pathways [14,17]. Prior studies have examined how AI capability, as a strategic resource, influences enterprises’ PI through optimized business operating capabilities and strategic orientation [10,17,18,19,30,41]. These studies have hinted that the specific capabilities necessary to achieve appropriate strategic orientation and organizational goals require further scholarly clarification. Building on this, this study demonstrates that AI capability requires BITC to navigate the challenges of integrating an organization’s resources with its optimized business operating capabilities to accomplish PI. By investigating BITC’s mediating role, this study provides a more holistic perspective on the drivers of manufacturing enterprises’ PI, elucidating how AI capability delivers value to enterprises engaged in productive innovation. Our study identifies a critical mechanism through which AI capability boosts manufacturing enterprise PI, offering new insight to the literature.

Second, the moderating roles of FKGMs and IKGMs reveal a critical boundary condition for AI capability to translate into BITC. While prior studies acknowledge the general importance of KGMs in enhancing knowledge-sharing capabilities [30,31,32], knowledge sharing is critical for aligning distinct strategic resources (e.g., AI capabilities) and strategic orientations (e.g., BITC) [23,27]. Prior studies have pointed to the possibility of pivotal yet underexplored relationships in this area. Based on the studies above, our study pioneers the elucidation of the specific pathways through which FKGMs and IKGMs facilitate knowledge sharing among organizational resources and strategic orientations. Additionally, prior studies have identified that IKGMs complement FKGMs by fostering the sharing of tacit knowledge, thereby highlighting the indispensable role of IKGMs in bridging structured governance and emergent knowledge flows [28,30]. Building on this premise, our comparative analysis advances the discourse by demonstrating that IKGMs exert a significantly more substantial moderating effect than FKGMs on the relationship between AI capability and BITC. This finding not only corroborates the centrality of tacit knowledge in strategic resource alignment [23] but also provides a nuanced framework for orchestrating knowledge governance between AI capability and BITC.

6. Conclusions

6.1. Theoretical Implications

First, this study advances current research by differentiating how AI capability drives PI in distinct ways compared to prior work. We demonstrate that AI capability enables PI through a unique pathway by fostering BITC as a critical mediator. Previous research has primarily focused on how AI capabilities assist enterprises in safeguarding proprietary data and managing resources [35] as well as how AI capabilities facilitate data-driven decision-making, optimize resource allocation, and enhance enterprises’ ability to detect market trends to innovate [9]. Unlike studies that treat AI capability as a standalone strategic resource [10,13,55], we reveal that its transformative potential depends on BITC’s ability to integrate AI capability with strategic orientation, addressing a key gap in understanding AI capability’s value-realization mechanisms.

Second, this study makes a distinct contribution to the literature on AI capability-driven PI by elucidating the boundary conditions that govern the relationship between AI capability and BITC. Prior research has demonstrated that PI-oriented enterprises, when faced with vast knowledge resources, require both FKGMs and IKGMs to enhance their knowledge-sharing capabilities [30,38]. Furthermore, the value of KGMs in facilitating knowledge sharing has been widely acknowledged [31,32]. However, the moderating role of KGMs in the relationship between AI capability and BITC as strategic resources and orientations remains underexplored. Investigating this role could provide deeper insight for advancing PI during AI capability implementation. Our study advances the literature on AI-driven organizational transformation by demonstrating that FKGMs and IKGMs act as boundary conditions between AI capability and BITC, thereby filling a critical gap in this research domain.

Third, our study advances the knowledge governance literature by revealing the asymmetric moderating effects of FKGMs and IKGMs on the relationship between AI capability and BITC. Existing studies have identified key critical limitations of over-relying on FKGMs, notably demonstrating that material incentives frequently cannot maintain tacit knowledge flow [86], yet ways to transcend these FKGMs limitations remain underexamined. Our study demonstrates that IKGMs promote knowledge sharing more effectively than FKGMs, particularly between AI capability and BITC. Our findings reshape the understanding of knowledge-sharing implementation between AI capability and BITC. It also indirectly reveals that the PI process in this pathway is primarily driven by sharing tacit knowledge [87], in which IKGMs play a crucial and decisive role. This contribution addresses the confusion regarding the impact of FKGMs and IKGMs on knowledge sharing, thus elucidating the roles of FKGMs and IKGMs in knowledge sharing and innovation.

Fourth, our study extends RBT by revealing how AI capability serves as a strategic resource, BITC enables strategic resource allocation for PI, and FKGMs and IKGMs provide the microfoundations for mobilizing intangible resources across these capabilities. Studies on the impact of AI capability on innovation have confirmed its alignment with RBT’s core principles and acknowledged the need to integrate AI capability into strategic decision-making [37,87], yet ways to align with RBT remain underexamined. Our study advances RBT by theoretically framing AI capability as a critical strategic resource, demonstrating how BITC facilitates strategic resource allocation for PI, and revealing FKGMs and IKGMs as the microfoundational mechanisms that enable effective mobilization of intangible resources. These contributions collectively extend RBT’s theoretical boundaries to address the unique challenges and opportunities of AI capability-driven PI.

6.2. Managerial Implications

This study discusses the managerial implications from the perspectives of enterprises and policymakers. From an enterprise perspective, this study reconstructs the digital transformation pathway by identifying BITC as the critical strategic pivot. Specifically, enterprises should prioritize establishing robust BITC. An insufficient BITC endowment hinders the effective translation of investments in AI capabilities into tangible PI outcomes. This inherent bottleneck necessitates a prudent evaluation of BITC maturity alongside AI investment decisions to ensure the successful realization of PI. Additionally, for enterprises, the key practical implication lies in recognizing that both FKGMs and IKGMs significantly amplify the impact of AI capabilities on BITC. Our study results confirm that FKGMs and IKGMs are essential levers for converting AI capability into BITC. Crucially, IKGMs exert a far more substantial moderating effect than FKGMs, primarily due to their unique capacity to mobilize tacit knowledge, such as experiential insights, contextual intuition, and collaborative problem-solving, which is indispensable for interpreting and transforming complex data and capabilities and driving PI. Therefore, enterprises must not only maintain parallel development of both KGMs but also strategically prioritize IKGMs as the dominant governance force, actively cultivating trust-based networks, collaborative cultures, and informal learning systems to unlock the full value of AI investments for BITC development.

For policymakers, our study results underscore the need to move beyond isolated support for AI capability development and to develop policies that foster BITC growth. Specifically, funding programs, tax incentives, or public–private partnerships should be designed that explicitly reward enterprises for building BITC, such as designing a “BITC Translation Grant” to direct funding for establishing dedicated BITC to action unit enterprises. Enterprises providing auditable evidence of BITC-driven impacts on PI shall qualify for proportional tax credits. The government co-invests with industry consortia to build sector-specific BITC validation labs. According to the results of the moderating effects, PI incentive policies can be refined through the use of differentiated subsidies. Specifically, enterprises with underdeveloped FKGMs should be offered subsidies for the development of foundational knowledge management, while enterprises that have implemented integrated IKGMs should be supported in establishing open innovation platforms. Additionally, significant efforts should be directed towards establishing IKGM-intensive innovation districts, epitomized by environments fostering cultures similar to the “garage culture” of Silicon Valley.

6.3. Limitations and Future Study Directions

Our study has some limitations. Although this study adopted a multi-wave, multi-source data collection approach to mitigate CMB and strengthen causal inferences, several limitations remain. First, although the temporal separation of data collection across three phases reduced CMB and provided some insights into causality, this study did not completely capture the long-term dynamics of AI capability development and its impact on PI. Future research could employ longitudinal studies spanning several years to track the evolution of AI capabilities and their effects on innovation outcomes over extended periods. Second, although this study used a stratified random sampling approach to ensure the randomness and representativeness of the sample, several limitations should be noted. Although the stratification based on enterprise size and industry sectors enhances the diversity of the sample, it may not fully capture the heterogeneity within each stratum. For instance, variations in organizational culture, resource availability, or regional differences within the same size–industry category could influence the outcomes but remain unaccounted for in the current sampling design. Future research could incorporate additional stratification criteria to further refine the sampling process. Third, this study focused on the manufacturing sector, where technological innovation is a priority. Future studies could test whether our findings are generalizable to other industries, such as service sectors, where AI-driven innovation may follow different mechanisms. Additionally, manufacturers with similar AI capabilities may develop and leverage them differently [88]. Future studies could segment enterprises based on the maturity levels of their AI capability and investigate whether different development pathways lead to distinct innovation outcomes.