Alumni Networks, Board Characteristics, and Artificial Intelligence Adoption: Evidence from Chinese Listed Companies

Abstract

Alumni relationships are essential social capital that are significant in companies’ resource acquisition and information sharing. Using 2018 data from Chinese listed companies, this study examines the impact of the chairperson–alumni network on corporate artificial intelligence (AI) adoption. The results show that chairperson–alumni relations are positively associated with AI adoption. Moreover, the impact of chairperson–alumni networks on AI adoption may span industrial, administrative, and geographical boundaries. This study shows that chairperson–alumni networks can indirectly influence AI adoption by influencing board size. Finally, this study demonstrates the heterogeneity of the impact of the chairperson–alumni network on AI adoption.

1. Introduction

As a landmark technology, artificial intelligence (AI) is increasingly important in accelerating the transformation of traditional industries [1]. Unlike the previous three technological revolutions, AI has demonstrated a comparative advantage at the enterprise level in a wider range of business areas, including mechanized work in manufacturing [2] and service sectors such as retail, healthcare, security, and finance [3]. Using AI in the workplace can improve overall business efficiency and reduce business costs [4,5]. While an increasing number of organizations recognize AI’s potential, few companies have adopted it. AI has been adopted by only 16% of companies in Germany [6] and 23% of companies globally [7]. Therefore, investigating the factors influencing AI adoption and promoting the actual adoption and use of AI technology to fully exploit its advantages is of great theoretical and practical significance, respectively.

Extant research on the antecedents of AI adoption at the firm level has focused on organizational factors, such as digital skills [6], return on investment [8], speciesism [9], productivity [10], institutional pressures [11], and pressure to change [12]. However, few studies have explored the effects of social factors on AI adoption. Although Hangl, Behrens, and Krause [5] attempted to make a breakthrough in this area, the social factors affecting AI adoption that they explored were limited to human–robot collaboration. Social networks are relational structures through which enterprises access information and resources [13]. Today’s enterprises are within dynamic social networks, and most require using social networks for knowledge exchange and resource acquisition [14]. Given the importance of social networks in the adoption and innovation of new technologies [15,16], exploring how social networks affect AI adoption is necessary.

Companies are frequently linked through alumni relationships, resulting in large alumni networks [17]. Alumni have a natural affinity toward their alma mater and a strong sense of identity [18], which is important in this process as a type of social glue [19]. The alma mater bond acts as a silent lubricant that brings business executives closer together [20]. The sum of the various ties between graduates of a school and between graduates and their alma mater constitutes the alumni network, and its existence and occurrence assume various forms [21]. At the corporate level, alumni networks refer to the social networks among executives based on the same alma mater [22]. This study focuses on the alumni network of a company’s board chairperson, considering the extent to which the chairperson is active in the alumni network and their leadership role in the company [23]. Following definitions in previous studies [22,24], this study uses “alumni network” to refer specifically to the alumni network of corporate chairpersons, that is, the social network relationship between corporate chairpersons based on their alma mater.

Based on the above theoretical and practical background, this study aims to explore the impact of the chairperson–alumni network on a firm’s AI adoption using data from Chinese listed companies. This study’s main contributions are as follows. First, the existing literature on AI adoption focuses on technical and organizational factors [25,26,27], with limited research on socio-environmental factors. This study enriches the research on AI adoption by demonstrating the impact and specific mechanisms of informal institutions of alumni networks on AI adoption in companies. Second, this study empirically investigates the impact of alumni culture on corporate behavior at the micro-level, enriches social network research [28,29], and provides a new research path for subsequent social capital research. Finally, existing social network studies using spatial econometrics face the limitation of equal weighting [30,31]. This study pioneers the grouping small-difference coding (GSDC) method, which overcomes the technical bottleneck of unequal weight construction of microdata and improves the scientific measurement of alumni network effects.

The remainder of this paper is structured as follows: Section 2 introduces the concepts and hypotheses; Section 3 describes the data and research design; Section 4 presents the empirical results and analysis; Section 5 provides the discussion; and Section 6 concludes the paper.

2. Literature Review, Hypothesis Development, and Theoretical Framework

Although a wealth of theoretical research exists on the adoption of new technologies, academic research on AI adoption is still emerging [25,32]. Corporate executive alumni networks are important as widespread informal systems that help alleviate financial constraints and facilitate information sharing [33,34]. Therefore, company executives’ alumni networks promote corporate innovation [21,35] while increasing companies’ willingness to adopt AI. Exploring the impact of company executives’ alumni networks on AI adoption and their mechanisms of action can provide scientific methods and concepts for companies to adopt AI.

2.1. Literature Review

The term “artificial intelligence” was first coined by John McCarthy in 1956 and is described as the science and engineering of intelligent technologies [36]. In recent years, the number of organizations implementing AI has grown rapidly [37]. Existing studies have explored and explained the drivers of AI adoption, primarily at the individual and organizational levels.

AI adoption at the individual level is an important research direction in the field of AI technology adoption, and the theoretical results of this study are relatively rich. The antecedents affecting AI adoption at the individual level include perceived usefulness [38], trust in a company [39], cultural values [40], trust in AI techniques [41], technological optimism [42], hedonic motivation [43], and consumption values [44]. Unlike individual AI adoption, enterprise AI adoption depends on a combination of factors, making it a challenging area of study. Therefore, this study focuses on AI adoption at the organizational level. According to Kumar, Raut, Mangla, Ferraris and Choubey [26] and Wang and Su [45], AI technology adoption in enterprises refers to integrating AI with corporate strategies, systems, processes, people, and workflows, as well as collaborative work with corporate managers.

Several studies have explored and identified multiple factors at the organizational level that influence AI adoption [32]. A study of 655 manufacturing companies located in 16 leading industrialized countries shows that organizational factors such as digital skills, company size, and R&D intensity significantly influence AI adoption [6]. Other organizational factors that influence AI adoption include AI governance [46], organizational decision-making [47], organizational agility [48], organizational innovation [49], existing technical limitations [5], managerial capability and organizational readiness [50], current marketing activities [51], organizational culture [52], technology infrastructure [53], and organizational perception [54]. A rich and diverse body of research has focused on the antecedents of AI adoption at the organizational level. However, insufficient attention has been paid to the socio-environmental factors that influence AI adoption, and even fewer studies have explored the impact of social networks on AI adoption.

Empirical evidence from Engelberg et al. [55] shows that alumni relationships create approximately four times more economic value than colleagues and other social relationships. New technology adoption from a social network perspective has been studied for a long time [56,57]. Several recent studies confirm that social networks significantly affect the adoption of new technologies [16,58,59]. Unfortunately, this area of research is limited to the field of AI adoption. This study aims to address this gap by investigating the impact of chairpersons’ alumni networks in Chinese listed companies on AI adoption. In this study, “alumni network” specifically refers to the corporate chairpersons’ alumni network, that is, the social network relationship between corporate chairpersons based on their alma mater [22,24].

While social networks, such as alumni ties, are often viewed as valuable social capital that facilitates resource exchange and trust [29,60], a growing body of literature highlights their potential disadvantages. Overembeddedness in closed networks may lead to insularity, reduce exposure to diverse information, and stifle innovation [61]. Moreover, exclusionary dynamics within elite networks—such as those formed through shared educational backgrounds—can reinforce homogeneity, limit thought diversity, and perpetuate power imbalances [17]. In the context of AI adoption, while alumni networks may provide access to technical knowledge and financial resources, they may also discourage the exploration of non-network-based innovations or create barriers to entry for firms outside these circles. This study acknowledges these dual aspects, but primarily focuses on the facilitative role of alumni networks, leaving the examination of potential negative effects for future research.

2.2. Hypothesis Development

2.2.1. Link Between Alumni Networks and Artificial Intelligence

Social capital theory suggests that social networks are essentially connections between various channels through which resources and information flows are traded [29,60]. An alumni network is a network of relationships; that is, a communication channel for information and a community of interest [62]. Alumni have various resources to assist and share with each other [34]. Alumni relationships play an important role in business processes by making company executives value the establishment and maintenance of these relationships [17]. All these factors help develop a favorable external environment and provide certain resources and information for companies to adopt AI technologies.

From the perspective of the chairperson–alumni network, a network of alumni connections can fund the adoption of AI in a business. AI innovation and adoption are ongoing cumulative investment processes requiring significant financial support throughout the project duration [63]. The executive alumni network frequently acts as a bridge to improve the understanding between borrowers and lenders and overcome financing barriers [64]. For example, when a mutual fund manager has an alumni relationship with a member of the company’s board of directors, the fund manager’s longstanding position is significantly higher in companies with alumni relationships [65]. Additionally, company executives who share information through alumni networks facilitate AI adoption by their companies. Adopting AI systems is difficult for companies; every stage is accompanied by challenges, and the risk of failure remains high [12]. Therefore, alumni networks are important channels for sharing information among company executives [33]. Sharing AI adoption information among executives of different companies can reduce the “cost of trial and error,” identify the right direction in time, and increase the probability of successful AI adoption [25].

Based on the aforementioned analysis, this study proposes the following hypothesis:

H1. 

The chairperson’s alumni network positively affects AI adoption in business.

2.2.2. Boundary-Spanning Effects of Alumni Networks

Social networks generally exhibit boundary-spanning effects [66]. The alumni network is extensive and direct, reaching different sectors and industries [62]. Through alumni networks, alumni from different sectors and industries can build direct bridges for communication and collaboration across sectors [35]. AI technology is a foundational technology in many industries, with clear cross-industrial characteristics [67]. Therefore, the impact of the chairperson–alumni network on AI adoption should also span industries. Alumni relationships have strong spatial and industrial penetration, and alumni from different regions and industries can achieve optimal allocation of resources, such as capital and technology, through exchange and cooperation [68]. As the chairperson–alumni network is translocal [69], its impact on AI adoption in businesses should also be considered translocal. This implies that the impact spans administrative boundaries. Social networks are generally considered to be systems that effectively span geographical boundaries [70]. Information and knowledge sharing occur within alumni networks [21,22], and information and knowledge flows often span geographic boundaries [71]. Therefore, the impact of a chairperson’s alumni network on AI adoption should also cross geographic boundaries; that is, they are not bound by geographic proximity. Therefore, this study proposes the following hypotheses:

H2a. 

The impact of a chairperson’s alumni network on AI adoption is not limited to industrial boundaries.

H2b. 

The impact of a chairperson’s alumni network on AI adoption is not limited to administrative boundaries.

H2c. 

The impact of a chairperson’s alumni network on AI adoption is not limited to geographical proximity.

2.2.3. Link Between Alumni Networks and Board Size

The factors influencing board size and composition can be divided into four categories: firm complexity and private interests; outside directors’ regulatory costs; ownership incentives; and executive characteristics [72,73]. A study on family-owned firms in Thailand shows that executives with rich alumni ties exhibit improved business performance [74]. Therefore, these companies are more complex and have larger boards. In terms of personal interests, the chairperson’s alumni network is also the “coterie” [33]. Inviting alumni to join the board of directors, thereby expanding its size, is easy for the chairperson. Additionally, inviting suitable candidates from among their alumni to serve as outside directors is easy for the chairperson; this increases the size of the board and reduces the cost of supervision by outside directors because of the natural trust among the alumni [34]. The chairperson may select alumni to join the company as executives [75], give them equity, and bring them to the board, which also increases their size. Therefore, this study proposes the following hypothesis:

H3a. 

The chairperson’s alumni network positively affects board size.

2.2.4. Link Between Board Size and Artificial Intelligence

Employee attitudes influence AI adoption; when employees are resistant, companies’ willingness to adopt AI decreases [27]. As the size of the board increases, particularly the number of inside directors, the directors are more likely to coordinate the willingness of the company’s employees to embrace AI technology, which helps them adopt it. Inside directors play a key role in safeguarding R&D investments and corporate technological innovation by monitoring the CEO and mitigating the information asymmetry among independent directors [76]. Therefore, the greater the number of inside directors on the board, the better the company can guarantee the innovation and application of AI technology and the more likely it is to adopt AI [32]. Additionally, the greater the number of internal directors on the board, the more likely that risk-averse and conservative policies will be adopted in highly uncertain environments [77]. Consequently, interest in using AI technologies to achieve efficiency in uncertain environments increases, along with the likelihood of firms adopting AI [12,78]. Therefore, this study proposes the following hypothesis:

H3b. 

The board size positively influences AI adoption.

According to previous studies [79,80], if H3a and H3b hold simultaneously, board size is a mediator between the chairperson’s alumni network and AI adoption. In addition to the direct impact, the chairperson’s alumni network indirectly influences AI adoption by influencing board size. Based on the above analysis, the chairperson’s alumni network may increase board size (H3a), which, in turn, facilitates AI adoption (H3b). Therefore, board size may serve as a mediator between a chairperson’s alumni network and corporate AI adoption. According to mediation theory [79,80], if the alumni network positively affects board size (H3a), board size positively affects AI adoption (H3b), and the direct effect of the alumni network on AI adoption remains significant after including the mediator, a mediation effect can be inferred. Thus, this study proposes the following hypothesis:

H3. 

Board size mediates the link between alumni networks and AI adoption.

2.3. Theoretical Framework

In summary, AI adoption by enterprises is affected by various factors, and related studies have revealed the impact of these factors on AI adoption [6,32,46], providing a basis for understanding the antecedents of AI adoption in enterprises. However, these studies have not effectively answered the question of the action mechanism of enterprise AI adoption in the context of social networks, nor have they unearthed the firmly established linkage between the internal and external factors affecting enterprise AI adoption. Additionally, existing AI adoption studies generally ignore the boundary-spanning effects of social networks on the impact of AI [16,58,59], which weakens the explanatory power of existing theories on corporate AI adoption. Therefore, this study considers the chairperson alumni network as the starting point, explores the complex mechanisms of the social network and board structure affecting enterprise AI adoption from an organic system perspective, and proposes a theoretical model, as shown in Figure 1. The fish diagram shown in Figure 1 exemplifies the organic system theory, which explains AI adoption in an enterprise. Alumni networks located in the caudal fin can span industrial, administrative, and geographic boundaries and directly impact AI adoption in the fish head. Simultaneously, alumni networks in the caudal fin can indirectly influence AI adoption in the fish head by influencing board size in the dorsal fin.

Figure 1. Schematic of the theoretical framework. Alt text: Alumni networks located in the caudal fin can span industrial, administrative, and geographic boundaries and directly impact AI adoption in the fish head. Simultaneously, alumni networks in the caudal fin can indirectly influence AI adoption in the fish head by influencing board size in the dorsal fin.

Alumni relationships are widespread social networks and are important components of social capital. Each school has its own unique cultural imprint that can strengthen the sense of identity and trust as well as promote information exchange and experience sharing among alumni. Information can be transmitted through alumni networks, with business executives generally assigning high credibility and value to this information. Therefore, such information is more likely to guide decision-making. Two companies whose presidents have alumni relationships tend to share information on AI adoption decisions and experience with AI applications, resulting in a high degree of similarity between the two companies’ AI adoption decisions. Therefore, alumni networks can facilitate AI adoption by companies. When company executives face limited information and noisy environments, the collective concerted efforts of the board of directors are likely to effectively utilize information from the alumni network to make informed AI adoption decisions and increase the efficiency of AI importation. Therefore, the board size can act as a mediator and bridge between alumni networks and AI adoption.

Notably, the impact of alumni networks on AI adoption has a “boundary-spanning” effect, capable of overcoming industrial, administrative, and geographic boundaries. As shown in Figure 1, alumni networks positively influence AI adoption. However, when industry boundaries are introduced, that is, when the scope of the study is limited to one industry, the effect of alumni networks on AI adoption becomes insignificant (n.s). This indicates that the impact of alumni networks on AI adoption overcomes industry boundaries. The same is true for administrative and geographic boundaries. This reflects the boundary-breaking properties of the trust, resources, and information embedded in alumni networks. However, this emerging field has not attracted significant attention from researchers.

3. Data, Variables, and Methods

This study selected Chinese A-share companies listed on the Shanghai and Shenzhen stock exchanges in 2018 as the research sample. China’s A-share market comprises the Shanghai and Shenzhen Stock Exchanges. In China, the disclosure of the academic qualifications of listed company executives is mandatory; however, the disclosure of information about their graduation institutions is not. Consequently, some listed companies do not disclose information on their executives’ graduation schools. The year 2018 was selected because a relatively large number of listed companies disclosed information about the graduation institutions of their executives. Therefore, the sample size was relatively large. As disclosing information about their executives’ graduation institutions is not mandatory for listed companies, some choose not to disclose this information; therefore, identifying the alumni networks of the chairpersons of these listed companies is not possible. In addition to obtaining the graduation schools of the chairpersons of listed companies, this study developed a web crawler program using Python 3.13.6 to crawl data on the names of graduation schools for the chairpersons of listed companies. However, this effect was not ideal because China has a high renaming rate. Most Chinese universities do not disclose their lists of alumni. Therefore, publicly disclosed information was used in this study. Data on listed companies were obtained from the China Stock Market Accounting Research Database (CSMAR), which includes financial data on listed companies, company executive data, and information on the nature of listed companies. After excluding specially treated (ST) companies and those with missing values, 579 companies were included in the study sample. Of the 579 chairpersons in the sample, 148 (25.56%) did not have an alumni relationship with anyone else and the remaining 431 (74.44%) had at least one alumni.

3.1. Artificial Intelligence

In previous studies, the metrics used for AI technologies at the organizational level are typically indicators such as industrial robots [81,82], AI metrics based on text mining [83], and the value of equipment per capita [84,85]. However, each of these measurements has its own advantages and disadvantages. Although Wang and Dong [82] estimate the industrial robot penetration of listed companies through a layer-by-layer decomposition method, many human subjective factors exist in this estimation process. Therefore, ensuring that the estimated industrial robot penetration is the real situation is difficult. Zhong, Xu and Zhang [83] measure AI usage in companies by calculating the frequency of AI-related terms used in texts such as annual reports of listed companies through text mining. However, the limitations of this measurement are also evident. First, what is said is not necessarily what is being done; great talkers are little doers. Second, a time lag exists between companies conceptualizing AI and implementing it. Finally, AI-related terms often have varying connotations across different industries and businesses. All these factors contribute to the biased nature of AI measurements obtained through text mining.

However, the metric of the value of equipment per capita to measure the use of AI in an enterprise also has some limitations, such as the existence of equipment in the enterprise that is unrelated to AI, as well as the rental of AI equipment. However, from a multi-industry perspective, the value of equipment per capita is a better indicator of an enterprise’s AI adoption. AI in an enterprise does not exist on its own, but is attached to related machines and devices [86]. AI makes machines and equipment smarter, thereby increasing their value [87]. The AI usage in enterprises has accelerated the streamlining of their workforce [88]. Thus, a positive relationship exists between enterprise AI adoption and the value of equipment per capita. This suggests that the value of equipment per capita can be used as a proxy variable for enterprise AI adoption. Following He, Li, Cheng and Li [84] and Lei and Xie [85], we use the ratio of the value of machines and equipment to the number of employees as a measure of AI technology adoption level, denoted as AI. The indicator is standardized for use.

Additionally, following previous studies [82,89], industrial robot penetration is used to measure companies’ AI adoption in the robustness tests. Furthermore, following Zhong, Xu and Zhang [83], text-mining-based AI metrics are used to measure companies’ AI adoption in robustness tests. The empirical results of these measurements are consistent with the results of the measurements using the value of equipment per capita. This further illustrates the validity of using value of equipment per capita as a measure of AI adoption. It is noteworthy that the models using both industrial robot penetration and text-mining-based AI metrics have a low level of model fit (Pseudo R²), whereas the models using the value of equipment per capita have a high level of model fit.

Therefore, this study uses the value of equipment per capita as the primary dependent variable in the baseline model, while industrial robot penetration and text-mining-based AI metrics are used in the robustness tests. However, the limitations of this proxy must be acknowledged. Although the value of equipment per capita is a reasonable indicator of capital intensity that may reflect AI-related investments, it does not directly measure AI adoption, and may be influenced by other types of machinery and equipment unrelated to AI. Furthermore, variations in equipment leasing practices and industry-specific capital structures may also affect the measure. These limitations must be considered when interpreting the findings.

While the value of equipment per capita provides a useful cross-industry indicator of capital intensity that may correlate with AI adoption, the relationship between equipment value and AI investment may vary across industries. For example, in capital-intensive manufacturing sectors, high equipment values may reflect traditional automation rather than AI-specific technologies, whereas in service- or knowledge-intensive industries, AI adoption may rely more on software and intangible assets that are not fully captured by this measure. Although our robustness tests and the use of alternative proxies (e.g., text-based AI indicators) help mitigate such concerns, future research could further examine how industry-specific factors moderate the link between capital intensity and AI adoption.

3.2. Alumni Networks

In terms of the operational definition of alumni networks, network centrality has been the main indicator used in previous studies [22,34]. This method is suitable for studying the alumni network of a company’s executive team, because the team generally has a large number of personnel, including the chairperson, vice chairperson, general manager, and deputy general manager [35]. This method can be traced back to previous studies [29,90]. However, the alumni network centrality obtained using this method is an exogenous variable with a potential endogeneity problem of reverse causality [91]. This problem can be effectively overcome if the alumni network is an intrinsic part of the empirical model [30]. Fortunately, the method of the spatial autoregressive model offers this option [31].

According to this method [92,93], the chairperson’s alumni network can be represented by a weighted matrix: a commonly used matrix is a weight matrix comprising zeros and ones, that is, the spatial contiguity weight matrix [79]. If two chairpersons share an alma mater, the weight of these two chairpersons is 1; otherwise, it is 0. If a total of $n$ chairpersons exists, these weights form an $n\times n$ weight matrix $W$. It can be observed that the alumni network of the chairperson is reflected in this weight matrix. Accordingly, the average AI adoption level (${AI}_{-i}$) of companies other than the target company (company $i$) in the chairperson’s alumni network is calculated. The formula used is as follows:

$${AI}_{-i} = W\times AI$$

(1)

However, in this method, the weights are the same for chairpersons who have graduated from the same school; this is known as the equal-weight method. In reality, a difference always exists in the closeness of relationships among alumni who graduated from the same school, and assigning equal weights is not possible. In a spatial weighting matrix, assigning the same weights to all members is misleading [94]. Therefore, the weights should differ among the chairpersons who graduated from the same school.

To solve this problem, spatial inverse distance weights are used. In the sample data, the graduation code of the chairperson is a seven-digit integer. Those with the same graduation code are alumni, whereas those with different graduation codes are not. The weight is then obtained by calculating the gap between the graduation school codes of the chairperson and the spatial inverse distance using the following formula:

$$w_{ij} = {\displaystyle \frac{1}{\left|U_j-U_i\right|}}$$

(2)

where $w_{ij}$ represents the spatial inverse distance weight between chairpersons $i$ and chairperson $j$, and $U_i$ and $U_j$ represent the graduation school codes of chairpersons $i$ and chairperson $j$, respectively. If chairpersons $i$ and $j$ belong to the same graduation school, they are alumni, and the value of $w_{ij}$ can be augmented by reducing the difference between the two codes in terms of graduation school coding. For example, recoding is performed on a sample of chairpersons in a particular graduation school so that each chairperson is coded from the smallest to largest, plus a very small decimal Δ in that order. If $n$ chairperson samples exist from the graduating school, then the maximum gap in the newly generated school codes for that graduating school is (n−1) $\times$ Δ. If the value of Δ is given as ${10}^{-8}$, then the minimum $w_{ij}$ value for that school is ${\displaystyle \frac{{10}^8}{\left|n-1\right|}}$. The weight of the spatial inverse distance between alumni is observed to be significant. If chairpersons $i$ and $j$ belong to different graduate schools, the $w_{ij}$ value must not be greater than 1. Since graduation school codes are seven-digit integers, the difference between the codes of different graduation schools must be an integer greater than or equal to one. Subsequently, its inverse must be less than or equal to one. Accordingly, the spatial inverse distance weights between alumni were large, whereas the spatial alumni weights for non-alumni were limited. This method is characterized as the GSDC. In principle, the smaller the value of the small difference parameter Δ, the better; however, ${10}^{-8}$ is typically sufficient. Therefore, the baseline nuanced gap parameter Δ value set in this study is ${10}^{-8}$, and the spatial inverse distance weight matrix of the sample chairperson is constructed using the GSDC method, which is the chairperson’s alumni network. As each firm has only one chairperson, this spatial inverse distance weight matrix is also an alumni network among the sample firms.

The GSDC method employed in this study assigns differentiated inverse-distance weights to chairpersons who graduated from the same institution, thereby implicitly capturing variations in tie strength based on graduation coding proximity. Although this approach moves beyond binary (0/1) measures by introducing weight variability, it does not directly measure the actual frequency of interaction, depth of relationship, or other qualitative aspects of tie strength that may influence resource and information flows. Future research could enhance the measurement of alumni networks by incorporating survey-based data on interaction frequency, collaboration history, or perceived closeness between alumni.

3.3. Covariates

In terms of the control variables, the necessary controls are exercised at the chairperson’s personal and corporate levels. Company executives’ overseas experiences frequently influence their corporate decisions [95]. Therefore, the presence or absence of overseas study experience of the chairperson (Overseas) is used as a control variable. Additionally, this study controlled for the chairperson’s gender (Gender) and age (Age), which are commonly used control variables at the individual level [96,97]. Aligning with previous studies [98,99], this study controls for the commonly used attributes of firm size (Size), firm age (Fage), and business performance (ROA). The financial gearing ratio (Gear) is an important indicator of business risk [100]. Therefore, it is used as a control variable. External audits are also an important factor influencing company behavior [101]. In the auditing industry, the term “Big Four” refers to four of the world’s leading accounting firms: PricewaterhouseCoopers (PwC), Deloitte & Touche (DTT), KPMG, and Ernst & Young (EY). This study also uses the Big Four audits by listed companies (Audit) as a control variable. Finally, following Ying et al. [102], the shareholdings of the top ten shareholders (Top), an indicator of the shareholding structure, is also used as a control variable.

3.4. Mediating Variable

Board size is defined as the number of members on a company’s board of directors [103,104]. Board size can also mediate corporate behavior [105]. This study focuses on the number of inside directors, as opposed to Chang et al. [106], who study the number of outside directors on boards. Referring to previous studies [72,107], the number of inside directors is the number of directors on the board of directors who are employees of a company. Therefore, board size (Board), which is operationally defined as the number of directors among the employees on the board, is used as a mediating variable in this study.

A summary of these concepts and variables is provided in Table 1. The details of the correlation coefficients are provided in the Appendix A.

Table 1. Summary information on concepts and variables.

Symbol	Name	Operational Definition	Mean	Std. Dev.	Min	Max
AI	Artificial intelligence	Standardized ratio of value of machines and equipment to the number of employees	0	1	−0.102	23.923
$W$	Alumni networks	Weight matrix of alumni network of chairpersons constructed by the GSDC method	-	-	-	-
Overseas	Overseas experience	1 for overseas study experience, 0 for vice versa	0.183	0.387	0	1
Gender	Gender of chairperson	1 for men, 0 for women	0.934	0.248	0	1
Age	Age of chairperson	Specific age of the chairperson	52.491	7.124	24	78
Size	Company size	Natural logarithm of the company’s total assets	22.385	1.673	19.371	30.776
Fage	Company age	Year difference between the year studied and the listed year	8.513	7.04	0	27
ROA	Return on total assets	The company’s return on total assets in the year under study	0.02	0.144	−1.629	0.228
Gear	Financial gearing ratio	The company’s financial liabilities divided by its assets	0.389	0.245	0	0.922
Audit	External audit	1 for those audited by the “Big Four,” 0 for others	0.093	0.291	0	1
Top	Shareholding structure	Shareholding of top ten shareholders (%)	61.641	15.086	21.81	100
Board	Inside director representation	Total number of directors from employees on the board	2.739	1.946	0	8

3.5. Methods

Referring to existing research [30,31], the impact of a chairperson’s alumni network on firms’ AI adoption can be examined using the following spatial lag model:

$$\begin{array}{c}AI = \mathsf{\alpha }+\mathsf{\rho }W\times AI{+\beta }_{1}Overseas+{\beta }_{2}Gender+{\beta }_{3}Age+{\beta }_{4}Size+{\beta }_{5}Fage\\ +{\beta }_{6}ROA{+\beta }_{7}Gear{+\beta }_8\times Audit{+\beta }_9\times Top+\mathsf{\epsilon }\end{array}$$

(3)

where $W$ is the spatial weight matrix constructed using the GSDC method to represent the alumni network relationships between the chairpersons of the sample companies. The estimated coefficient $\mathsf{\rho }$ represents the core parameter in this study [92,93]. If $\mathsf{\rho }$ is significantly positive, it shows that the chairperson’s alumni network will have a catalytic effect on AI adoption in the business. If $\mathsf{\rho }$ is significantly negative, it shows that the chairperson’s alumni network can have a disincentive effect on AI adoption in the business. If $\mathsf{\rho }$ is not significant, it shows that the chairperson’s alumni network does not impact AI adoption in the business.

To examine whether the impact of the chairperson’s alumni network on firms’ AI adoption is confined to the same industry, further tests must be conducted using the following model:

$$\begin{array}{c}AI = \mathsf{\alpha }+{\rho }_{1}W\times M_1\times AI{+\beta }_{1}Overseas+{\beta }_{2}Gender+{\beta }_{3}Age+{\beta }_{4}Size+{\beta }_{5}Fage\\ +{\beta }_{6}ROA{+\beta }_{7}Gear{+\beta }_8\times Audit{+\beta }_9\times Top+\mathsf{\epsilon }\end{array}$$

(4)

where $M_1$ is the spatial weight matrix comprising zeros and ones constructed to reflect whether the sample companies belong to the same industry. If two companies belong to the same industry, their weighting is 1; otherwise, it is 0. If ${\rho }_1$ in Formula (4) is nonsignificant; however, $\mathsf{\rho }$ in Formula (3) is significant. This suggests that the impact of the chairperson’s alumni network on firms’ AI adoption is not limited to the same industry alone, but is cross-industry and transcends industry boundaries.

Following the same reasoning, we construct a provincial spatial weight matrix $M_2$ and spatial contiguity weight matrix $M_3$. For $M_2$, the weight is 1 if the two companies are located in the same province, and 0 otherwise. For $M_3$, the weight is 1 if the two companies are located next to each other, and 0 otherwise. The corresponding models are given by Formulas (5) and (6):

$$\begin{array}{c}AI = \mathsf{\alpha }+{\rho }_{2}W\times M_2\times AI{+\beta }_{1}Overseas+{\beta }_{2}Gender+{\beta }_{3}Age+{\beta }_{4}Size+{\beta }_{5}Fage\\ +{\beta }_{6}ROA{+\beta }_{7}Gear{+\beta }_8\times Audit{+\beta }_9\times Top+\mathsf{\epsilon }\end{array}$$

(5)

$$\begin{array}{c}AI = \mathsf{\alpha }+{\rho }_{3}W\times M_3\times AI{+\beta }_{1}Overseas+{\beta }_{2}Gender+{\beta }_{3}Age+{\beta }_{4}Size+{\beta }_{5}Fage\\ +{\beta }_{6}ROA{+\beta }_{7}Gear{+\beta }_8\times Audit{+\beta }_9\times Top+\mathsf{\epsilon }\end{array}$$

(6)

Depending on the parameters $\mathsf{\rho }$ and ${\rho }_2$, determining whether the impact of the chairperson’s alumni network on AI adoption by companies is limited to the same province or across provinces is possible. Similarly, depending on the parameters $\mathsf{\rho }$ and ${\rho }_3$, determining whether the impact of the chairperson’s alumni network on AI adoption by companies is limited to geographical proximity or transcends geographical proximity is possible.

To test whether board size is a mediating variable, following previous studies [31,80], two additional models are added to Formula (3) as follows:

$$\begin{array}{c}Board = \mathsf{\alpha }+\mathsf{\rho }W\times AI{+\beta }_{1}Overseas+{\beta }_{2}Gender+{\beta }_{3}Age+{\beta }_{4}Size+{\beta }_{5}Fage\\ +{\beta }_{6}ROA{+\beta }_{7}Gear{+\beta }_8\times Audit{+\beta }_9\times Top+\mathsf{\epsilon }\end{array}$$

(7)

$$\begin{array}{c}AI = \mathsf{\alpha }+\mathsf{\rho }W\times AI {+\mathsf{\gamma }Board+\beta }_{1}Overseas+{\beta }_{2}Gender+{\beta }_{3}Age+{\beta }_{4}Size+{\beta }_{5}Fage\\ +{\beta }_{6}ROA{+\beta }_{7}Gear{+\beta }_8\times Audit{+\beta }_9\times Top+\mathsf{\epsilon }\end{array}$$

(8)

Board is the mediating variable if the parameters $\mathsf{\rho }$ in Formulas (3), (7) and (8) are all significant and the parameter $\mathsf{\gamma }$ in Formula (8) is also significant [108].

4. Results

4.1. Baseline Results

A stepwise regression method is used to conduct the empirical tests. The first step examines the impact of the chairperson’s alumni network on firms’ AI adoption without controls. Column (1) of Table 2 presents the results. The second step examines the impact of the chairperson’s alumni network on firms’ AI adoption, controlling for provincial effects. Column (2) of Table 2 presents the results. The third step examines the impact of the chairperson’s alumni network on firms’ AI adoption, controlling for both provincial and industry effects. Column (3) of Table 2 presents the results. The fourth step examines the impact of the chairperson’s alumni network on the adoption of AI in their firms while simultaneously controlling for provincial effects, industry effects, firm ownership, and listing location. Column (4) of Table 2 lists the results. In the fifth step, the three control variables at the chairperson’s personal level are added to Column (4). Column (5) of Table 2 presents the results. Finally, the six control variables at the firm level are added to Column (5). Column (6) of Table 2 presents the results.

Table 2. Empirical results of baseline models.

	Explained Variable: AI
	(1)	(2)	(3)	(4)	(5)	(6)
$\rho$	0.657 **	0.732 **	0.555 **	0.609 **	0.612 ***	0.545 **
	(2.24)	(2.41)	(2.32)	(2.59)	(2.64)	(2.40)
Overseas					0.284 ***	0.229 **
					(2.91)	(2.35)
Gender					0.003	0.005
					(0.02)	(0.03)
Age					−0.002	−0.005
					(−0.44)	(−0.99)
Size						0.013
						(0.39)
Fage						0.009
						(1.29)
ROA						0.144
						(0.56)
Gear						0.344 **
						(2.15)
Audit						0.431 ***
						(2.70)
Top						0.002
						(0.55)
Constant	−0.053	−0.02	−0.098	0.059	0.163	−0.468
	(−0.18)	(−0.2)	(−0.32)	(0.19)	(0.36)	(−0.56)
Control of ownership	No	No	No	Yes	Yes	Yes
Control of listed location	No	No	No	Yes	Yes	Yes
Control of industry	No	No	Yes	Yes	Yes	Yes
Control of province	No	Yes	Yes	Yes	Yes	Yes
N	579	579	579	579	579	579
Pseudo R2	0.141	0.16	0.319	0.32	0.33	0.354
Wald chi2	145.21 ***	122.12 ***	280.94 ***	286.87 ***	301.67 ***	329.88 ***
Wald test of spatial terms	5.01 **	5.81 **	5.36 **	6.68 **	6.95 ***	5.75 **

z-statistics are in parentheses; ** p < 0.05, *** p < 0.01.

According to the empirical results in Table 2, the parameter $\rho$ is significantly positive in all models. This suggests that the chairperson’s alumni network significantly contributes to AI adoption businesses. In Column (6) of Table 2, the estimated coefficient of the effect of the chairperson’s alumni network on AI adoption is 0.545, which is significant at the 5% confidence level. Therefore, H1 is supported. The chairperson’s alumni network, as important social capital for a company, can bring rich innovation and information resources [22]. This can provide additional knowledge, technology, and intellectual support for the adoption of AI technology [26], thus strengthening the application and adoption of AI technology in the company [33,35].

Columns (5) and (6) show that the regression coefficient of Overseas is significantly positive, indicating that chairpersons with overseas backgrounds are more likely to adopt AI in their firms. However, the chairperson’s gender and age do not significantly impact the company’s AI adoption. Additionally, according to Column (6), the regression coefficients of both Gear and Audit are significantly positive, indicating that both the financial gearing ratio and “Big Four” audits have a positive and significant impact on AI adoption by firms. However, the effects of firm age, size, return on assets (ROAs), and equity structure on AI adoption are not significant.

4.2. Robustness and Sensitivity Analysis

To test the robustness of the findings, a multi-round test is conducted by adjusting for the spatial inverse distance weighting strategy. The results are summarized in Table 3. Column (1) shows the results obtained from the spatial inverse distance matrix of the baseline model, which is the result in Column (6) of Table 2. The corresponding value of the parameter Δ is ${10}^{-8}$. From Columns (2)–(6), the Δ value is then adjusted from ${10}^{-7}$ to ${10}^{-3}$. The value of the parameter Δ is observed to be increasing. According to Table 3, the corresponding nonzero minimum (minimum > 0) of the spatial inverse distance weight also shows a simultaneous increasing trend but is generally close to 0. The maximum value of the corresponding spatial weights, which is between 0.15 and 0.161, does not significantly change, and the mean value of the corresponding spatial weights remains the same, which is always 0.001. In terms of the impact of the chairperson’s alumni network on AI adoption in their firms, the effect is significant and positive regardless of the adjusted spatial weights, as described above, and the regression coefficient values do not significantly change and remain between 0.545 and 0.637. This indicates that the impact of the chairperson’s alumni network on a firm’s AI adoption is robust.

Table 3. Sensitivity analysis.

	Explained Variable: AI
	(1)	(2)	(3)	(4)	(5)	(6)
$\rho$	0.545 ** (2.4)	0.637 *** (2.7)	0.63 *** (2.68)	0.63 *** (2.68)	0.63 *** (2.68)	0.632 *** (2.68)
variables controlled	Yes	Yes	Yes	Yes	Yes	Yes
Control of ownership	Yes	Yes	Yes	Yes	Yes	Yes
Control of listed location	Yes	Yes	Yes	Yes	Yes	Yes
Control of industry	Yes	Yes	Yes	Yes	Yes	Yes
Control of province	Yes	Yes	Yes	Yes	Yes	Yes
N	579	579	579	579	579	579
R2	0.354	0.348	0.348	0.348	0.348	0.348
Wald chi2	327.05 ***	327.05 ***	327.05 ***	327.05 ***	327.05 ***	327.01 ***
Wald test of spatial terms	5.75 **	7.32 ***	7.19 ***	7.19 ***	7.19 ***	7.19 ***
Spatial matrix
$\mathsf{\Delta }$	${10}^{-8}$	${10}^{-7}$	${10}^{-6}$	${10}^{-5}$	${10}^{-4}$	${10}^{-3}$
Elements
Minimum	0	0	0	0	0	0
Minimum > 0	3.24 × 10−15	3.24 × 10−14	3.24 × 10−13	3.24 × 10−12	3.24 × 10−11	3.24 × 10−10
Mean	0.001	0.001	0.001	0.001	0.001	0.001
Max	0.161	0.153	0.151	0.151	0.151	0.15

z-statistics are in parentheses; ** p < 0.05, *** p < 0.01.

A common method to achieve robustness is to change the explanatory variable, control variable, and so on. Here, industrial robot penetration is used as a proxy variable for AI. Acemoglu and Restrepo [89] construct a measure of “robot penetration” at the regional level in the U.S. However, this measure is not at the firm level. Here, Wang and Dong’s method [82] is used to estimate industrial robot penetration at the industry level in China, using industrial robot data released by the International Federation of Robotics, and then further estimate the firm-level industrial robot penetration of listed companies in China. Industrial robot penetration, a new dependent variable, is used with industrial robot penetration as a measure of AI adoption by listed companies. Regressions are conducted using a stepwise procedure based on the steps listed in Table 2, and the results are presented in Table 4. According to the results listed in Table 4, the regression coefficients of $\rho$ are significantly positive in all six models, even when the AI adoption is switched to a new proxy variable, under different control situations. This finding is consistent with the results presented in Table 2. This suggests that this study’s findings are robust.

Table 4. Robustness analysis for industrial robot penetration.

	Explained Variable: AI
	(1)	(2)	(3)	(4)	(5)	(6)
$\rho$	0.124 **	0.147 **	0.143 **	0.15 **	0.143 **	0.151 **
	(1.97)	(2.23)	(2.17)	(2.29)	(2.17)	(2.29)
variables controlled	No	No	No	No	Yes	Yes
Control of ownership	No	No	No	Yes	Yes	Yes
Control of listed location	No	No	No	Yes	Yes	Yes
Control of industry	No	No	Yes	Yes	Yes	Yes
Control of province	No	Yes	Yes	Yes	Yes	Yes
N	579	579	579	579	579	579
Pseudo R2	0.033	0.039	0.063	0.079	0.085	0.086
Wald chi2	122.54 ***	126.84 ***	143.46 ***	154.61 ***	157.77 ***	159.66 ***
Wald test of spatial terms	3.89 **	4.96 **	4.69 **	5.23 **	4.72 **	5.24 **

z-statistics are in parentheses; ** p < 0.05, *** p < 0.01.

Additionally, the approach of Zhong, Xu and Zhang [83] is followed to generate an AI lexicon using machine learning methods and textually analyze the annual reports of listed companies to construct enterprise-level AI metrics. Using this machine learning and text-mining route, four firm-level metrics of AI transformation are constructed: AI vocabulary as a percentage of the vocabulary in the managerial discussion module (percentage); AI vocabulary as a percentage of the vocabulary in the firm’s strategy development module (percentage); number of AI words in frequency; and AI investment as a percentage of intangible asset investment. Using these indicators as proxy variables for AI, regression analyses are performed for each, and the corresponding results are summarized in Columns (1)–(4) of Table 5. According to Table 5, the regression coefficient of $\rho$ is significantly positive in all four models. These results are consistent with those presented in Table 2. This suggests again that the study’s findings are robust.

Table 5. Robustness analysis.

	Explained Variable: AI
	(1)	(2)	(3)	(4)
$\rho$	1.61 ***	1.432 ***	1.182 ***	1.278 ***
	(6.16)	(4.65)	(4.06)	(4.63)
variables controlled	Yes	Yes	Yes	Yes
Control of ownership	Yes	Yes	Yes	Yes
Control of listed location	Yes	Yes	Yes	Yes
Control of industry	Yes	Yes	Yes	Yes
Control of province	Yes	Yes	Yes	Yes
N	579	579	579	579
Pseudo R2	0.013	0.019	0.009	0.055
Wald chi2	99.3 ***	80.78 ***	96.51 ***	97.3 ***
Wald test of spatial terms	37.9 ***	21.58 ***	16.48 ***	21.4 **

z-statistics are in parentheses; ** p < 0.05, *** p < 0.01.

4.3. Boundary-Spanning Effects

To test whether the influence of the chairperson’s alumni network on AI adoption by the company can overcome the industry, administrative district, and geographical boundaries, corresponding spatial proximity weight matrices, such as $M_1$, $M_2$, and $M_3$ are constructed. Regressions are then performed based on Formulas (3)–(5), and the corresponding results are listed in Columns (2)–(4) of Table 6. Column (1) of Table 5 presents the results of the baseline model obtained based on Formula (2), which is the same as that in Column (6) of Table 2.

Table 6. Empirical results for boundary-spanning.

	Explained Variable: AI
	(1)	(2)	(3)	(4)
$\rho$	0.545 **	1.848	−0.262	0.008
	(2.40)	(0.73)	(−0.15)	(0.08)
Overseas	0.229 **	0.240 **	0.231 **	0.232 **
	(2.35)	(2.47)	(2.40)	(2.41)
Gender	0.005	−0.001	0.002	0.001
	(0.03)	(−0.01)	(0.01)	(0.01)
Age	−0.005	−0.006	−0.006	−0.006
	(−0.99)	(−1.17)	(−1.21)	(−1.21)
Size	0.013	0.009	0.008	0.008
	(0.39)	(0.27)	(0.24)	(0.25)
Fage	0.009	0.011	0.012 *	0.012 *
	(1.29)	(1.62)	(1.69)	(1.71)
ROA	0.144	0.165	0.156	0.156
	(0.56)	(0.65)	(0.62)	(0.62)
Gear	0.344 **	0.328 **	0.332 **	0.329 **
	(2.15)	(2.08)	(2.09)	(2.08)
Audit	0.431 ***	0.449 ***	0.449 ***	0.445 ***
	(2.70)	(2.84)	(2.79)	(2.82)
Top	0.002	0.001	0.002	0.001
	(0.55)	(0.49)	(0.54)	(2.41)
Constant	−0.468	−0.352	−0.325	−0.333
	(−0.56)	(−0.43)	(−0.39)	(−0.40)
Control of ownership	Yes	Yes	Yes	Yes
Control of listed location	Yes	Yes	Yes	Yes
Control of industry	Yes	Yes	Yes	Yes
Control of province	Yes	Yes	Yes	Yes
N	579	579	579	579
Pseudo R2	0.354	0.169	0.361	0.364
Wald chi2	329.88 ***	331.82 ***	331.41 ***	331.37 ***
Wald test of spatial terms	5.75 **	0.54	0.02	0.001

z-statistics in parentheses; * p < 0.1, ** p < 0.05, *** p < 0.01.

In Table 6, a comparison of Columns (1) and (2) reveals that the estimate of $\rho$ is significantly positive in Column (1) and nonsignificant in Column (2). This suggests that the impact of a chairperson’s alumni network on AI adoption is not limited to the same industry; that is, this impact transcends industry boundaries. Thus, the chairperson’s alumni network spans industry boundaries. Therefore, H2a is supported. Comparing Columns (1) and (3) reveals that the estimate of $\rho$ is significantly positive in Column (1), but appears nonsignificant in Column (3). This suggests that the impact of the chairperson’s alumni network on AI adoption is not limited to the same administrative district; that is, this impact extends beyond administrative district boundaries. In other words, the chairperson–alumni network spans administrative district boundaries. Therefore, H2b is supported. Finally, a comparison of Columns (1) and (4) reveals that the estimate of $\rho$ is significantly positive in Column (1) and nonsignificant in Column (4). This suggests that the impact of the chairperson’s alumni network on AI adoption is not limited to geographical proximity; that is, this impact extends beyond geographical boundaries. That is, the chairperson–alumni network spans geographical boundaries. Therefore, H2c is supported. These findings corroborate those of previous studies on the boundary-spanning effects of social networks [66,71]. However, for the alumni network, this is the first piece of evidence identified in this study.

4.4. Analysis of Mediating Mechanism

Columns (1)–(3) of Table 7 report the results of testing the mediating effect of board size, under which we expect the chairperson’s alumni network to influence AI adoption through board size in turn. Column (1) examines the relationship between the chairperson’s alumni network and AI adoption; Column (2) examines the relationship between the chairperson’s alumni network and board size; and Column (3) examines the relationship between board size and AI adoption. The results in Column (2) show that the coefficient of $\rho$ is 0.766, with a coefficient greater than 0 and significant at the 5% level, indicating that the chairperson’s alumni network has a significant positive relationship with board size and that the chairperson’s alumni network can increase board size. Therefore, H3a is supported. The coefficient of board size (Board) is 0.055 and is significant at the 5% level. Therefore, H3b is supported. The results in Column (3) show that the coefficient of $\rho$ is 0.518, with a coefficient greater than 0 and significant at the 5% level, indicating that the chairperson’s alumni network has a significant relationship with AI adoption. These results indicate that board size has a mediating effect on a chairperson’s alumni network and AI adoption [31,80]. Therefore, H4 is supported.

Table 7. Test for the mediating mechanism.

	Explained Variable
	AI	Board	AI
	(1)	(2)	(3)
$\rho$	0.545 **	0.766 **	0.518 **
	(2.40)	(2.06)	(2.28)
Board			0.055 **
			(2.53)
control variable	Yes	Yes	Yes
Control of ownership	Yes	Yes	Yes
Control of listed location	Yes	Yes	Yes
Control of industry	Yes	Yes	Yes
Control of province	Yes	Yes	Yes
N	579	579	579
Pseudo R2	0.354	0.372	0.363
Wald chi2	329.88 ***	342.35 ***	341.24 ***
Wald test of spatial terms	5.75 **	4.22 **	5.22 **

z-statistics are in parentheses; ** p < 0.05, *** p < 0.01.

4.5. Heterogeneity Analysis

To examine the heterogeneous influence of the chairperson’s alumni network on AI adoption, the influence of the chairperson’s position and employment experience is analyzed from four perspectives: whether the chairperson is also a general manager, has R&D experience, has employment experience in an investment bank, and has employment experience in a research institution. The results are presented in Table 8 and Table 9. Columns (1) and (2) of Table 8 show that the effect of the chairperson’s alumni network on AI adoption is significant and positive only when the chairperson is not the CEO. The effect of the chairperson’s alumni network on AI adoption is not significant when the chairperson is also the CEO. Columns (3) and (4) of Table 8 show that the effect of a chairperson’s alumni network on AI adoption is significant and positive only when the chairperson has no R&D experience. The effect of the chairperson’s alumni network on AI adoption is not significant when the chairperson has R&D experience.

Table 8. Heterogeneity test of position and overseas employment.

	Explained Variable: AI
	CEO		Research and Development
	(1)	(2)	(3)	(4)
	Yes	No	Yes	No
$\rho$	−0.041	0.440 *	0.178	0.556 **
	(−0.72)	(1.95)	(1.05)	(2.38)
control variable	Yes	Yes	Yes	Yes
Control of ownership	Yes	Yes	Yes	Yes
Control of listed location	Yes	Yes	Yes	Yes
Control of industry	Yes	Yes	Yes	Yes
Control of province	Yes	Yes	Yes	Yes
N	156	423	165	414
Pseudo R2	0.865	0.399	0.611	0.396
Wald chi2	1002.08 ***	289.37 ***	260.61 ***	283.75 ***
Wald test of spatial terms	0.52	3.8 *	1.11	2.38 **

z-statistics in parentheses; * p < 0.1, ** p < 0.05, *** p < 0.01.

Table 9. Heterogeneity test of institutional characteristics.

	Explained Variable: AI
	Investment Banking		Scientific Research Institutions
	(1)	(2)	(3)	(4)
	Yes	No	Yes	No
$\rho$	0.351 ***	−0.010	0.176 ***	0.506 **
	(2.69)	(−1.15)	(4.02)	(2.16)
control variable	Yes	Yes	Yes	Yes
Control of ownership	Yes	Yes	Yes	Yes
Control of listed location	Yes	Yes	Yes	Yes
Control of industry	Yes	Yes	Yes	Yes
Control of province	Yes	Yes	Yes	Yes
N	528	51	44	535
Pseudo R2	0.406	0.998	0.987	0.359
Wald chi2	372.78 ***	271,164.36 ***	4805.13 ***	309.57 ***
Wald test of spatial terms	7.25 ***	1.32	16.16 ***	4.66 **

z-statistics are in parentheses; ** p < 0.05, *** p < 0.01.

Columns (1) and (2) of Table 9 show that the effect of the chairperson’s alumni network on AI adoption is significant and positive only when the chairperson has employment experience in an investment bank. The effect of the chairperson’s alumni network on AI adoption is insignificant when the chairperson has no employment experience at an investment bank. Columns (3) and (4) of Table 8 show that the effect of the chairperson’s alumni network on AI adoption is significant and positive, regardless of whether the chairperson has employment experience in a scientific research institution. In terms of the estimated coefficient value of $\rho$, AI adoption’s impact is greater for the alumni networks of chairpersons without employment experience in research institutions than for those with it.

A degree of heterogeneity is observed in the impact of the chairperson’s alumni network on AI adoption, which can vary depending on the chairperson’s position and employment experience. Alumni networks of chairpersons with employment experience in investment banking but no R&D experience and who do not serve as CEOs are more likely to help promote AI adoption.

5. Discussion

5.1. Theoretical Implications

This study contains theoretical breakthroughs and innovations through several aspects.

First, it proposes an integrative analytical framework for enterprise AI adoption, which enriches the theoretical research on organizational AI adoption [16,58,59] and extends the theoretical connotations of social network research [28,29]. In the past, analyses of alumni networks by company executives focus on areas such as turnover [22], corporate social responsibility [33], investment strategy [21], and innovation [35], but the impact of alumni networks on AI adoption has generally been overlooked. To the best of our knowledge, this study is the first to explore the relationship between alumni networks and AI adoption. This confirms that a chairperson’s alumni network has a catalytic effect on companies’ AI adoption. Further research highlights the boundary-spanning effects of alumni networks on AI adoption, which can break industry, administrative, and geographical boundaries. This study adopts an organic system perspective, integrating the chairperson’s alumni network outside the enterprise, the board of directors’ structure inside the enterprise, and boundary-spanning effects into one organic system. It also adopts an integrative analytical framework for enterprise AI adoption, laying a theoretical foundation for empirical research on the synergistic influence of multiple factors on enterprise AI adoption. Focusing on AI adoption by listed companies, an in-depth study of the phenomenon of the chairperson’s alumni network and its impact is of theoretical significance for improving the theoretical research framework for organizational AI adoption [47,52,53], enriching relevant research on the social network of company executives [24,62,69].

Existing theories on organizational AI adoption tend to focus on only one aspect, either the process of technology adoption [47,54], social relationships [16,59], or general organizational structure issues [48,53]. This has led to a relatively low overall efficiency of AI adoption in organizations [6], and many organizations are yet to adopt AI technologies [7]. To solve the issue of the overall efficiency of AI adoption in enterprises, analyzing organizational AI from a holistic perspective is necessary; therefore, the organic system theory emerges. This study proposes an organic system theory that integrates alumni networks, board size, and AI adoption, emphasizing that corporate AI adoption is fundamentally an organizational behavior in an integrated organic system comprising people, materials, machines, and other resources. Therefore, both internal and external factors affect AI adoption. Particularly, a firm is a subsystem of a social system; therefore, AI adoption by a firm depends not only on internal conditions but also on conditions external to the firm, such as informal institutional factors, social networks, and social capital. Thus, this study’s results can reconstruct or improve the knowledge system of AI adoption in organizations [46,52], and provide a scientific and reasonable theoretical framework for enterprises to successfully adopt AI by balancing internal and external factors in a dynamic environment [32,49].

Second, this study not only proposes an organic system of alumni networks and AI adoption, but also further discovers the path of action between the two, which helps to extensively excavate and anatomize the decision mechanism of AI adoption in enterprises and provide new perspectives for fine-tuned research in the field of organizational AI adoption [25,27]. Existing research on the antecedents of AI adoption at the firm level includes factors such as R&D intensity [6], return on investment [8], human–robot collaboration [5], and revenue growth [63]. However, existing research has not effectively revealed the influence of social networks on corporate AI adoption [16,58,59]; thus, it is not possible to test and explain how external and internal factors synergize to influence corporate AI adoption and their specific micro-action mechanisms. In other words, the mechanism and path of social network influence on enterprise AI adoption remain a “black box.” This study not only finds an intrinsic association between chairperson alumni networks and firms’ AI adoption, but also highlights a specific pathway of action, finding that board size plays a key role in bridging the gap between chairperson alumni networks and firms’ AI adoption. As this study uses cross-sectional data, it can only reveal the correlation between alumni networks and AI adoption, not causation. However, an indirect mediation effect analysis illustrates a possible causal relationship between the independent and dependent variables [109]. This effectively addresses the gap in the existing knowledge and provides a foundation for theoretical research on AI adoption at the enterprise level [12,110].

Finally, this study proposes a new method for solving the problem of different weights for the same group. This is a significant innovation in methodology. Previous studies on alumni networks have used indicators, such as concentration [20,24,35], which are exogenously given variables. This results in the endogeneity of reverse causation [111]. This study addresses endogeneity and reverse causality by constructing a spatial weight matrix that reflects alumni relationships between chairpersons, making the chairpersons’ alumni network a built-in component of the system. In this regard, the traditional method constructs a spatial proximity matrix with a weight of 1 between chairpersons who are alumni and 0 between those who are not [30,31]. However, in reality, alumni sharing the same alma mater are often closely and distantly related; therefore, the weighting cannot be the same for different alumni. Assigning equal weights to all the alumni is misleading [94]. Establishing unequally weighted alumni relationships with relatively large sample sizes is a difficult task, and no method is available for this purpose. In this regard, this study pioneers the GSDC, which uses spatial inverse-distance weights to define alumni relationships and solves the problem of unequal weights [92]. This method overcomes the technical bottleneck of building unequally weighted spatial matrices for large microdata samples, thereby allowing for further scientific measurement of alumni effects.

Although this study primarily highlights the facilitative role of alumni networks in promoting AI adoption, such informal networks may entail certain risks and limitations. Strong ties within closed alumni circles can lead to elite closure, where access to resources and opportunities is restricted to a homogeneous group, potentially reinforcing existing power structures and excluding outsiders. Moreover, over-reliance on alumni networks may result in reduced innovation diversity, as firms may become insular and less exposed to novel ideas or alternative technological pathways outside their immediate social circles. In the context of AI adoption, this could mean that firms may prioritize technologies endorsed within their alumni networks, overlooking potentially superior or more suitable alternatives. These dynamics suggest that, while alumni networks can serve as valuable conduits for information and resources, they may also inadvertently constrain organizational learning and innovation in the long run.

5.2. Practical Implications

This study’s results have several significant implications. Compared with the general AI innovation network, the innovation and application network formed based on alumni relationships has a stronger trust foundation, which is conducive to cooperation among these enterprises. Therefore, vigorously promoting the establishment of a cooperative platform and strategic AI alliances between enterprises with alumni relations is necessary; this is conducive to helping enterprises obtain more heterogeneous resources in the development and application of AI, and thus effectively improves the success rate of AI applications. For an enterprise, each executive’s alumni network is a potential external resource and each alumni network may be advantageous in certain business areas. Therefore, enterprises can strategically adjust and optimize the academic, geographic, and faculty structures of their executive teams by considering their business positioning and strategic direction to gain access to different alumni networks, improve their in-group identification in heterogeneous executive alumni networks, and form an optimal alumni network structure conducive to enhancing innovation performance and AI applications. Thus, enterprises should maintain a rational perception of executive alumni network values. After accumulating a certain alumni network size, enterprises should focus on building a platform that allows different executives’ alumni networks to form synergies with each other and facilitate the sharing of alumni information resources within the enterprise, which can help the enterprise have an information advantage in enhancing the overall efficiency of the enterprise’s AI adoption and application.

5.3. Limitations and Future Research

Based on this study’s limitations, the following research outlook is proposed. First, this study focused on the alumni network of the chairperson alone. The impact and paths of action of the alumni network of the company’s executive team on AI adoption of their company can be explored in depth in the future. Second, cross-sectional data were analyzed. In the future, panel data could be used to conduct a dynamic analysis of the impact of alumni networks on AI adoption. Third, although we employ multiple proxy variables for AI adoption, including the value of equipment per capita, industrial robot penetration, and text-based AI indicators, each measure has inherent limitations. For instance, the value of equipment per capita is an indirect measure that may capture capital intensity beyond AI-specific investments. Future research could benefit from more direct, survey-based, or administrative data on AI adoption to enhance measurement validity and precision. Finally, this study is based exclusively on Chinese listed firms as of 2018, which may limit the generalizability of the findings to other contexts. The Chinese institutional environment, characterized by strong state influence, unique corporate governance structures, and a rapidly evolving AI policy landscape, may shape the role of alumni networks and board characteristics in ways that differ from those in Western or other emerging economies. Furthermore, the year 2018 represents a specific point in China’s AI adoption trajectory, and the observed dynamics may not hold in earlier or later periods or under different macroeconomic conditions. Therefore, caution should be exercised when extending these results to other countries, time periods, or institutional settings. Future research could enhance external validity by replicating this study in different national contexts, using longitudinal data spanning multiple years, or conducting cross-country comparative analyses to examine how institutional factors moderate the relationship between alumni networks and AI adoption.

Moreover, this study’s conclusions may still be subject to certain limitations in terms of endogeneity and causality. Although a range of firm- and individual-level covariates are controlled for, and spatial econometric models are employed to mitigate peer effects, several sources of endogeneity may remain unaddressed. For instance, omitted variable bias could arise from unobserved factors—such as regional policy incentives for AI adoption, firm-specific innovation culture, or latent managerial capabilities—which may simultaneously influence alumni network formation and AI adoption decisions. Additionally, reverse causality, whereby firms adopting AI might attract chairpersons with stronger alumni networks, cannot be completely ruled out. While plausible pathways are suggested by our mediation analysis, the establishment of causal relationships could be better achieved in future research using panel data, instrumental variables, or quasi-experimental designs (e.g., policy shocks or alumni reunions as exogenous events). Furthermore, more granular controls for regional policy environments, firm innovation climates, and managerial traits could be incorporated, and longitudinal data, along with natural experiments, could be utilized to more robustly identify causal mechanisms.

While this study employs multiple quantitative proxies (e.g., the value of equipment per capita, industrial robot penetration, and text-based indicators) to measure AI adoption, these measures remain indirect and may not fully capture the nuanced organizational-level processes of AI integration. To enhance the validity and richness of AI adoption measurements, future research could incorporate qualitative methods such as in-depth interviews with executives, IT managers, or innovation officers, as well as longitudinal case studies of firms undergoing AI transformation. Such approaches would provide contextual insights into how AI is adopted, implemented, and leveraged within organizations and help validate whether quantitative proxies align with on-the-ground practices.

6. Conclusions

Alumni relationships have significant economic value. Unfortunately, relatively little is known about how alumni networks affect AI adoption in organizations, which challenges how enterprises develop innovative solutions to improve the efficiency and success of AI adoption. This study addresses this gap by exploring the relationship between alumni networks and enterprise AI adoption from an organic system perspective. Using a sample of Chinese-listed companies from 2018, this study constructs a spatial matrix of the chairperson’s alumni network and empirically tests its effect on corporate AI adoption. The findings show that the chairperson–alumni network facilitates AI adoption by firms. The effect is boundary-spanning, which can overcome industrial, administrative, and geographical boundaries. Furthermore, this study demonstrates that the size of a company’s board of directors plays an important role in bridging the relationship between alumni networks and corporate AI adoption. Finally, our findings suggest a heterogeneous relationship between alumni networks and corporate AI adoption. In summary, this study provides new insights into the contextual influence of social networks on organizational AI adoption. This study highlights the importance of corporate executives’ alumni networks, and the findings are instructive for firms to improve the motivation and success of AI adoption, providing empirical evidence for the important role of trust, information, and resources embedded in alumni networks in firms’ decisions to adopt AI. This study also emphasizes the crucial role of informal institutions, such as trust, in corporate decision-making and helps unravel the mystery of China’s miraculous progress in the development and application of AI technology in recent years. Future studies are needed to evaluate this study’s findings using data covering longer time spans and more regions worldwide.