How does digital-intelligent policy synergy foster ambidextrous innovation of manufacturing enterprises? The mediating role of resource endowments

The improvement of innovation ability is crucial in facilitating the competitiveness of manufacturing enterprises, responding to the dynamic external environment, and achieving long-term growth (Nie et al., 2022; Li et al., 2024c). At the current time, as the restructuring of global value chains and the realignment of industrial division of labor accelerate, manufacturing firms face an increasingly complex and volatile market and institutional environment. Only through continuous innovation can they flexibly adapt to such dynamic changes (Chung et al., 2020; Zhang and Cui, 2017), determining corporate market competitiveness (Aghion et al., 2012). As the world's largest manufacturing economy, particularly against the backdrop of the new round of global technological and industrial revolutions, the Chinese government places great emphasis on the innovative development of the manufacturing sector and has created a favorable environment for innovation through the introduction of a series of national policies, such as the 14th Five-Year Intelligent Manufacturing Development Plan and the Plan for Promoting Innovation-Driven Development of Service-Oriented Manufacturing (2025–2028).

Ambidextrous innovation refers to enterprise behaviors that affect or shape how these enterprises explore and utilize learning. By combining advantages and different resources or enterprises, ambidextrous innovation that consists of two different modes (i.e. exploitative and exploratory) provides more growth opportunities and maintains development stability (Hu et al., 2023; Gupta et al., 2023; Li et al., 2024c). Exploitative innovation builds on existing knowledge and involves replication, promotion, and refinement processes to ensure organizational stability (Kammerlander et al., 2020). In other words, this mode emphasizes corporate survival and short-term performance to ensure organizational stability (Wen et al., 2021). In contrast, exploratory innovation means that an enterprise needs to face possible changes and actively invest in the R&D of new materials and new processes. This mode entails high uncertainty and costs, focusing on long-term competitiveness (Jiang et al., 2025). Although this kind of innovation may occupy resources in the near term, it is key to coping with industrial upgrading and building long-term core competitiveness. Additionally, exploitative innovation can make up for the risks that exploratory innovation may bring through its short-term benefits (Jiang et al., 2025). Therefore, for manufacturing enterprises, how to effectively allocate resources and meet a dynamic equilibrium between ensuring daily operational efficiency (i.e. exploitative) and investing in future technical capabilities (i.e. explorative) is not only a strategic proposition, but also relates to their position and survival resilience in the global value chain (Jin et al., 2022).

Amidst rapid digital transformation, advanced technologies represented by big data and artificial intelligence (AI) are driving the continuous enhancement of the innovative ability of economic entities, including manufacturing firms. Studies have shown that the improvement of digitalization provides a technical basis for enterprises to quickly integrate information and management experience, enabling them to adjust and utilize resources more flexibly in rapidly changing markets (Papadopoulos et al., 2022). Against this background, the Chinese government has also begun to boost the expansion of the digital-based economy and digitalization through the implementation of relevant policies. The Broadband China Policy (BCP) seeks to expedite the rollout of broadband – an integrated, secure, and ubiquitous next-generation national information infrastructure. The selection of BCP demonstration cities began in 2013. After three rounds of approval from 2014 to 2016, 119 cities were identified (Li et al., 2025b). Leveraging its strong digital technology base, including the industrial internet, cloud computing, and AI, BCP promotes the restructuring and upgrading of traditional industries, driving the shift toward networked and collaborative production models (Feng et al., 2023). In addition, the Intelligent Manufacturing Policy (IMP) was released in 2015 to improve the intelligent capabilities of enterprises in research and development, production, management, and service, which, in turn, would cultivate multi-level talents, support industrial upgrading, and enhance the competitiveness and sustainability of the manufacturing sector. Therefore, the IMP is identified as an important measure to deepen the digitalization level in the sector by advancing the intelligence, precision, reliability, and productivity of the production procedure (Tao et al., 2018).

It is worth noting that whether digital and intelligent investment can be transformed into considerable performance improvement, especially the improvement of enterprise innovation capacity, is still controversial. Some scholars believe that the “Solo Paradox” in the digital economy era still exists (Brynjolfsson and Collis, 2019), and the deployment of digital tools could foster internal competition for scarce resources (Ardito et al., 2019). The information burden induced by digital technological applications might exceed the security capacity of enterprises, thereby increasing the risk of information leakage due to data errors, security vulnerabilities, or privacy issues (Tang and Veelenturf, 2019). Accordingly, the possibility of the failure of innovative initiatives is then increased. In addition, some scholars have found that intelligent manufacturing might not bring positive effects during the initial phase. For instance, one study revealed that the IMP allowed the listed companies targeted by the policy to receive more innovation subsidies, but there was little statistical evidence that this policy brought higher productivity and R&D (Li and Branstetter, 2024). Others observed that the enforcement of the policy caused increased R&D investment, yet its short-term innovation outcomes remained insignificant (Wen and Zhao, 2021). At the same time, when multiple policies are conducted simultaneously, complex interactions may arise, resulting in either synergistic effects (i.e. mutual reinforcement) or antagonistic effects (i.e. mutual weakening). Therefore, whether the coordination of China's broadband strategy and intelligent manufacturing policy can promote corporate ambidextrous innovation has not been adequately addressed. Accordingly, the first question of this article is: Can digital-intelligent policy synergy promote the ambidextrous innovation of manufacturing enterprises?

Innovation activities are typically resource-consuming and investment-intensive; therefore, a company's ability to allocate and provide resources at different stages will influence its innovation performance (Barney, 1991). Emerging technologies are driving widespread digital transformation (Jiang et al., 2025) and have the potential to unlock significant untapped potential within organizations, thereby enhancing their resource endowment (He and Shen, 2019). For instance, enterprises that adopt digital intelligence technology can effectively boost production efficiency by replacing human labor with automated machines. Furthermore, the digitalization process endows data resources with unique asset value, enabling enterprises with massive operational data reserves to explore growth potential and optimize allocation efficiency and utilization efficiency of various types of resources (Brynjolfsson and McElheran, 2016). However, how digitalization and intelligence can further affect the corporate innovation by reshaping their resource endowments has not yet been systematically explored. Therefore, our second research problem is: Will policy synergy of digital intelligence improve manufacturing firms' ambidextrous innovation through resource endowments management?

Currently, research on ambidextrous innovation in firms has primarily focused on the impact of individual elements such as digitalization (Yang and Xiao, 2024), financial technology (Zhang et al., 2024a) and network communication (Xu and Jiang, 2024). However, comprehensive analyses of how policy-driven digitalization impacts ambidextrous innovation remain scarce, particularly regarding the synergistic effects of different policy types. In the context of digitalization, understanding and promoting innovation from a resource-based view (RBV) has become an emerging and pressing issue (Shao et al., 2024). Accordingly, this research explores mechanisms through which digital-intelligent policy synergy (DIPS, i.e. both the BCP and the IMP) influences ambidextrous innovation. Ambidextrous innovation not only involves technological progress, but also the strategic design and precise implementation of policies. These two elements are mutually reinforcing and jointly drive the inclusive and long-term advancement of enterprises.

This research enriches the existing discussion in the following ways: First, we expand the scope of application of RBV in a policy context. By examining how the synergistic effects of digital-intelligent policy enhance firms' resource management capabilities, this paper deepens our understanding of how policy-driven digitalization drives innovation and offers new perspectives on how firms can leverage digitalization to build and reshape their resource endowment management capabilities to enhance innovation. Second, we systematically evaluate the synergy effect of digital-intelligent policies on ambidextrous innovation for the first time, and specifically study its direct, indirect, and heterogeneous effects, as well as how those effects vary across different sequences of policy implementation. Third, we use DML technology and four-step mediation approach for empirical testing. Compared with the traditional DID method, DML has the advantage of reducing the redundancy of control variables and model mis-setting. The four-stage mediation method reveals the transmission mechanism between variables more systematically, enhancing the logical integrity and explanatory power of the empirical analysis. Moreover, it mitigates potential endogeneity bias among explanatory variables, mediating variables, and explained variables, thereby improving the reliability of the estimation results (Niu et al., 2023). Therefore, this method strengthens the robustness and credibility of the empirical outcomes. In addition, our conclusions provide empirical support for the optimization of policy tools and show how the digital-intelligent policy synergy can promote enterprise innovation and upgrading. These insights provide valuable guidance for formulating targeted industrial policies and adjusting enterprise innovation strategies.

According to RBV, resources are unevenly distributed among enterprises, with the costs of transferring them high and the process slow. Digital intelligence constitutes a critical technological resource that enhances firms' information processing and decision-making capabilities. This special and immovable resource helps enterprises gain a competitive advantage (Barney, 1991). First, BCP can promote the construction of digital infrastructure, especially network connectivity and information flow, and drive a firm's digital transformation. Digital technologies enable enterprises to transcend the restrictions of temporal and spatial boundaries to gain the backing of external resources (Appio et al., 2021; Zhang et al., 2023), thus reducing the cost of resource acquisition. Digitalization improves the effectiveness and productivity of R&D and innovation by optimizing resource utilization, strengthening Internet integration, and enabling the seamless integration of physical simulation and data analysis in virtual environments, thereby reducing human and material costs and accelerating innovation progress (Bian and Fan, 2024; Li et al., 2024b; Cheng and Zhao, 2025). Second, digitalization provides infrastructure and technical support for intelligent manufacturing. Intelligent manufacturing can not only reshape traditional production models and give rise to new production methods, it can also promote enterprises to shift from experience-driven to accurate research and judgment and dynamic optimization through data-driven intelligent decision-making capabilities (Kusiak, 2017). Through the IMP, firms can enhance resource distribution, foster efficiency gains, control costs, and accelerate R&D innovation (Davis et al., 2015; Parhi et al., 2023). Consequently, the integration of expanded digital infrastructure with intelligent manufacturing initiatives can enhance enterprises' abilities to transform data into productivity, drive the adoption of cutting-edge technologies, and ultimately achieve breakthrough innovations (Li et al., 2025a). Third, digital intelligence will enable more accurate demand forecasting in the market, improve adaptive potential of new product models, and optimize operational processes (i.e. exploitative innovation). Digital intelligence also fosters stable and enduring cooperation among stakeholders, enabling enterprises to cross the existing technological development path, that is, exploratory innovation (Jiang et al., 2025). In accordance, we posit that:

Resource accumulation reflects the ability of enterprises to obtain external technical and financial support (Wang et al., 2024); it is both a fundamental component of the RBV and a prerequisite for innovation (Zahra, 2021). Ambidextrous innovation requires not only skilled professional knowledge, but also substantial capital investment (Cao et al., 2024), while digital and intelligence techniques have significantly improved enterprises' ability to accumulate resources. For example, innovation activities inherently involve notable information asymmetry, which may weaken stakeholders' perception and understanding of enterprise innovation behavior and its potential value. The application and development of digital intelligence technology has boosted the timeliness and transparency of corporate information disclosure, enabling creditors to continuously track the progress of enterprise technology innovation. This reduces information asymmetry and debt financing costs, and enables enterprises to obtain capital accumulation (Xu et al., 2023). Evidence suggests that increased digitalization can reduce stakeholders' concerns about the innovation ability of target companies and enhance trust among partners (Chen et al., 2020). Moreover, the application of digital intelligence approaches facilitates firms to continuously accumulate technical knowledge and innovative experience, thus promoting the continuous accumulation of their technical capabilities (Shao et al., 2024). Sufficient financial support and mature technical capabilities help enterprises continuously improve production technology and product performance within the existing technical path and product system, reduce the marginal innovation cost, and thus promote utilization innovation characterized by the improvement of the process (Ju, 2023). In addition, stable financial security and deepening technology accumulation have enhanced the ability of enterprises to withstand high-risk and uncertain innovative projects, provided support for cross-field technology exploration and subversive technology research and development, and then promoted exploratory innovation oriented by new process research and development (Huang et al., 2024). For these reasons, we posit that.

The ambidextrous innovation of enterprises depends not only on the resources they can obtain, but also on the utilization and efficiency of resources (Purchase et al., 2014). Therefore, we pay attention to enterprises' adaptability to the dynamic market environment and their effective use of resources (Dahabiyeh and Constantinides, 2022). In the process of leveraging resources, digitalization ability can improve the efficiency of divergent resource utilization. First, business process digitalization, smart manufacturing, and digital office solutions can dynamically and accurately allocate resources based on business needs, enabling the synergistic and flexible application and utilization of diversified elements and factors. What is more, Digital intelligence-driven market analysis, digitally integrated supply chain connectivity, and digital customer interaction models can address external operational variations in a real-time manner (Wu et al., 2022). In the meanwhile, the use of digital-based or digital-driven techniques represented by blockchain can accurately plan logistics transportation and warehousing time, reduce inventory backlog, and further boost the level of optimized utilization of resources (Wang et al., 2025). In addition, the optimization of resource utilization processes facilitates enterprises to improve operational processes, update manufacturing technologies and adopt new materials, such as process optimization and product line extension, thus promoting utilization innovation (Gao et al., 2025). Augmenting the efficiency of resource utilization can also save resources for enterprise innovation activities, such as capital, human resources, and management cognitive resources. In view of the limited supply of resources for enterprises at a specific time, and considering exploratory innovation requires extensive resource investment, the stable cash flow and profits created by efficiency improvement provide financial security for highly uncertain exploratory innovation. This enables enterprises to invest some of the resources released in the exploration of new technologies and new markets beyond existing business. Therefore, we propose the following arguments.

Existing research indicates that enhancing digital intelligence helps to break down information silos within organizations, thereby significantly promoting the optimal allocation of resources (Greenstein, 2020; Jin et al., 2024). Although innovation is crucial to a firm's long-term development, it doesn't yield immediate returns; commercialization of innovation is a lengthy and uncertain process, and improvements in firm performance often occur with a time lag (Vasileiou et al., 2022). Leadership of a company that controls the allocation of corporate resources may act in pursuit of its own short-term economic interests, thereby sacrificing the potential long-term profits brought by innovation (Shao et al., 2024). If there are disparities in information access among different economic actors within a firm, or if their expectations regarding the timeline of the same innovation project differ, this can lead to agency problems—including misallocation of resources—and ultimately hinder technological innovation (Galende, 2006). Through a digital technology platform, shareholders can access relevant information in a timely manner, thereby reducing the scope for speculative and self-serving behavior by management and lowering agency costs. More effective shareholder oversight not only curbs redundant investment and resource waste in inefficient projects but also encourages enterprises to reassign limited resources to exploitative innovative activities with more certain returns and manageable risks, thereby improving the efficiency of existing technology and product enhancements. Digital intelligence empowers decision-making processes through digital tools and intelligent strategies, improving decision-making accuracy and response efficiency and thereby promoting the efficient allocation of capital and labor. On this basis, resources are freed from over-invested and ineffective sectors. These factors collectively increase the long-term resource supply for exploratory innovation, enabling firms to pursue such innovation while stabilizing their existing business (Teece, 2018; Yu and Chen, 2025). Given this, we propose Hypothesis 4.

In summary, this paper argues that the synergy between digital-intelligent is a critical factor influencing enterprises' ability to enhance ambidextrous innovation, while the endowment of enterprise resources serves as the mechanism linking the two. The conceptual framework of whether and how policy synergy is achieved is illustrated in Figure 1.

The research covers manufacturing companies listed on China's A-share market during 2011–2023. Firms in the financial sector, those that were suspended from trading or subject to price limits, and those with missing data are omitted. All continuous variables are winsorized at the 1 and 99% levels. Financial variables are sourced from China Stock Market and Accounting Research and WIND, and patent data are generated from the Chinese Patent Data Project (CPDP) platform. The final sample consists of 24,128 observations from 2,984 companies.

Ambidextrous innovation (⁠$AINNOi,t$⁠). Ambidextrous innovation consists of exploitative (⁠$Exploiti,t$⁠) and exploratory (⁠$Explorai,t$⁠) innovation, and patent data are used to construct the dependent variables. We employ the IPC classification system and the first 4-digits of the IPC code and a common 5-year window period. Exploitative innovation is measured by the logarithm of the number of patents in existing classifications plus one (those used within the past five years), while exploratory innovation is measured by the logarithm of patents in new classifications plus one (Li and Lu, 2023).

This paper uses BCP, IMP, DIPS as proxy variables for digital policy, intelligent policy, and digital-intelligent policy synergy, respectively. We match the list of all BCP pilot cities with enterprise data, creating BCP dummy variable according to the implementation timeline of these pilots. It maps the ten key areas of the IMP with industries classified by the China Securities Regulatory Commission, identifying industries adopting smart manufacturing and the corresponding enterprises, and constructs IMP dummy variable based on the policy implementation dates (Shen et al., 2024). The core independent variable (⁠$DIPSi,t$⁠) indicates whether an enterprise is covered by the policy programs of BCP and IMP simultaneously. Specifically, if a firm i is registered in a pilot city that has implemented the BCP and participates in the smart manufacturing policy within its industry in year t, then $DIPSi,t=1$⁠; otherwise, $DIPSi,t=0$⁠.

Resource accumulation (FC): Considering that capital is one of the most common and critical resources needed for innovation, we use the degree of financing constraints to evaluate resource accumulation status (Shao et al., 2024).

Resource utilization (OE): Given that the level of resource utilization mirrors a company's ability to convert available resources into goals, this paper adopts operational efficiency as an indicator to reflect this concept (Li et al., 2021a).

Resource allocation (AC): The intrinsic long-term orientation and information asymmetry issues of dual innovation could introduce conflicts of interest between a company's managerial group and shareholders, leading to rising agency costs (Li et al., 2021b). The former exercises responsibility for resource allocation choices while the latter focuses on long-term returns, typically manifested. We refer to Rashid's (2016) study and use expense rate as the direct proxy variable for agency costs.

Firm size (Size): Larger enterprises often possess more resources to foster technical R&D; Corporate establishment years (Age): Older enterprises tend to have a stronger innovation mindset and are more inclined to initiate technological innovation; Asset-liability ratio (LEV): A higher LEV may indicate financial risk, which in turn, could impact an enterprise's ability to invest in innovation (Xiao, 2025); Return on assets (ROA): Enterprises with higher profit margins may have more funds available to support innovation; Revenue growth rate (Growth): Enterprises experiencing higher revenue growth may prioritize core business development over technological innovation (Guo et al., 2023); State-owned enterprise (SOE): SOEs tend to support R&D due to their alignment with national policies (Yang and Xiao, 2024); The proportion of independent directors (INDEP): this indicator can influence key enterprise decisions (Zhang et al., 2024a); Cash flow ratio (Cashflow): Enterprises with sufficient cash flow possess a stronger capacity to ensure steady financial support for their innovation initiatives; Shareholding ratio of institutional investors (INST): Higher institutional ownership may impact enterprise decision-making and distribution of resources (Jian et al., 2024); Proportion of fixed assets (Fixed): A higher fixed asset ratio may indicate a greater asset base but reduced flexibility in technology upgrades (Li et al., 2024a) (see Table A1 for details).

The current mainstream regression methods rely on a strict assumption system and set a linear connection between dependent and independent variables, which limits the implementation of the technique in the empirical process. In order to productively observe the influence of digital-intelligent synergy on ambidextrous innovation, this study adopts a DML model (Chernozhukov et al., 2018) for causal relationship identification. Compared with the above mainstream methods, the DML method loosens the linear dependence between variables and allows nonlinear and interactive influence relationships. At the same time, this method can consider high-dimensional control variables, which improves the accuracy of causal estimation. Chernozhukov et al. (2018) proved that the treatment effect coefficient estimated by this method is unbiased. As a high-dimensional non-parametric statistical method, various machine learning models are based on data-driven fitting of complex relationships existing in the samples, without the need to preset a model form to avoid model mis-specification issues. In addition, leveraging regularization algorithms, various sample-splitting cross-validation methods, and other techniques, machine learning models can automatically select the most effective subset of variables from a high-dimensional set of control variables for fitting data relationships. This ensures the consistency of causal effect estimates even with small sample sizes, thereby mitigating the “curse of dimensionality”. Researchers need only to select a possible set of control variables based on economic theory (Fang et al., 2024).

Ambidextrous innovation is estimated by referencing Chernozhukov et al. (2018):

where, $AINNOi,t$ represents ambidextrous innovation of firm i in year t, $DIPSi,t$ represents the digital-intelligent policy synergy; $Ui,t$ is the error term, with a conditional mean of 0. $Xi,t$ constituting a set of high-dimensional controls. $θ0$ denotes the treatment effect of concern. The specific form of $l(Xi,t)$⁠, $lˆ(Xi,t)$ is measured through machine-learning algorithms. Based on Equation (1), the coefficient estimator can be obtained:

To address the issue of small-sample bias in the coefficient estimators, the following auxiliary regression is proposed:

where, $m(Xi,t)$ is the regression function of the core explanatory variable regarding the high-dimensional controls, and its specific form $mˆ(Xi,t)$ can also be measured through machine-learning algorithms. $Vi,t$ represents the error term with a conditional mean of 0. We first conduct regression on Equation (3) to get the residuals $Vi,t=DIPSi,t−mˆ(Xi,t)$⁠, $Vˆi,t$ as the instrumental variable for $DIPSi,t$ to derive an unbiased estimated coefficient:

This study tests the mediating mechanism of DIPS and firm ambidextrous innovation by designing a four-stage approach to optimize the analysis process and make up for the limitations of the three-stage method. Specifically, the three models in the three-step approach entail the estimation of three sets of variables, potentially entailing three endogeneity issues. At least two instrumental variables are needed, and the three error terms are needed to be pairwise uncorrelated. As empirical studies mostly utilize observational data, taking these endogeneity issues into account would render the research highly complex. Accordingly, we utilize the four-stage framework listed below for mediation tests to strengthen completeness of the empirical research (Aguinis et al., 2017).

where, $MCi,t$ are the mediating variables, namely resource endowments. To test whether the potential mediating variables are effective, we should first evaluate whether $DIPSi,t$ affects $AINNOi,t$ (Eq. (5)), then test whether $DIPSi,t$ affects $MCi,t$ (Eq. (6)), then test whether $MCi,t$ affects $AINNOi,t$ (Eq. (7)), and lastly estimate the combined effect of $DIPSi,t$ and $MCi,t$ on $AINNOi,t$ (Eq. (8)).

The outcome presented in Table 1—specifically in columns (1)–(3), (4)–(6), and (7)-(9)—analyze the effects of policy synergy on ambidextrous, exploratory, and exploitative innovation across models. The estimated results indicate that policy implementation simultaneously increases ambidextrous innovation (θ₀ = 0.4578), exploitative innovation (θ₀ = 0.5124), and exploratory innovation (θ₀ = 0.1351), respectively. The results demonstrate that policies not only encourage enterprises to optimize the existing knowledge, but also advocate the exploration of new technology. The influence of policy synergy on exploitative innovation is particularly notable, which may be explained by the fact that exploitative innovation is more dependent on resource integration and efficiency improvement, both of which are directly optimized through the popularization of digital foundational facilities and the improvement of intelligent technology. The above confirms that digital-intelligent policy synergy has significantly boosted corporate performance in various innovation modes, and H1, H1a and H1b have been confirmed.

Table 2 illustrates the outcome of three mediating effects. We first examine the mediating effect of resource accumulation, and the outcome is presented in columns (1) to (4). First, the outcome reveals that the observed coefficient of DIPS on AINNO in column (1) is remarkably positive, with the coefficient as 0.4727, confirming the overall effect of DIPS on ambidextrous innovation. In column (2), we tested the coefficient of DIPS on FC, which was significantly negative (β₁ = −0.0734), indicating that DIPS alleviated financing constraints. The regression results in column (4) show that the coefficient of DIPS decreased compared to column (1), but is still positive (β₄ = 0.4702), and both FC in columns (3) and (4) are significantly negative (β₂ = −0.0315 and β₃ = −0.0260), demonstrating that financing constraints inhibit the development of ambidextrous innovation. This outcome indicates that firms affected by DIPS will come across fewer financial constraints. In other words, DIPS helps improve resource accumulation, which is crucial for AINNO. Therefore, the four-step test indicates a partial mediating effect. Overall, Hypothesis 2 is supported.

Columns (5) to (8) demonstrate the mediating effect of resource utilization between DIPS and ambidextrous innovation. First, we find that the estimated coefficient of DIPS on ambidextrous innovation in Column (5) is significantly positive (β₀ = 4,709), confirming the total effect of DIPS on ambidextrous innovation. Second, we find DIPS positively affects OE with the co-efficiency as 0.4329 (i.e. column (6)) and OE positively affects AINNO with the co-efficiency as 0.0014(i.e. column (7)); the joint effects of DIPS and OE on AINNO are positive and significant with the co-efficiency as 0.4441 and 0.0013 (i.e. column (8)), suggesting that policy synergy not only directly increases ambidextrous innovation but also can be further affected through increasing resource utilization. Additionally, DIPS is significantly positive correlated with AINNO (i.e. column (9), β₀ = 4,578), negative correlated with AC (i.e. column (10), β₁ = −0.0033) and AC negatively impacts AINNO (i.e. column (11), β₂ = −0.7058), with column (12) presenting the combined significant effects of DIPS and AC on AINNO (β₃ = −0.6984, β₄ = 0.4023), indicating the mediating effect of resource allocation. Hypothesis 3 and Hypothesis 4 are thus supported. In addition, the results of Sobel test and Bootstrap (i.e. 1,000 times) sampling tests show that FC, OE, and AC are playing partial mediating roles, demonstrating that DIPS promotes AINNO by improving the levels of resource accumulation, utilization, and allocation.

Furthermore, given that ambidextrous innovation consists of exploratory and exploitative innovation, we extend the mediation analysis to these two dimensions separately. As shown in Table 3 and Table 4, the results reveal differentiated mediating patterns. For exploitative innovation (Table 3), DIPS exerts significant indirect effects through FC, OE, and AC, with all mediators showing statistically significant coefficients. This provides empirical support for H2a, H3a, and H4a, respectively, indicating that resource accumulation, utilization, and allocation all serve as effective transmission channels for DIPS to enhance firms' exploitative innovation. For exploratory innovation (Table 4), the mediating roles of FC, OE, and AC are also significant, albeit with relatively smaller coefficient magnitudes compared to the exploitative model. The Sobel Z statistics further confirm the mediation effects, thereby supporting H2b, H3b, and H4b. These findings suggest that while DIPS facilitates both types of innovation through the same three resource-based mechanisms, the strength of these pathways varies between exploration and exploitation.

To enhance the reliability of the estimation outcome, multiple robustness tests are conducted, such as altering the dependent variable, modifying the sample range, incorporating interaction terms, controlling for concurrent implementation, substituting the DML approaches, and addressing endogeneity.

Referring to Wang et al. (2023), the total volume of invention patent and utility model transformation applications is utilized to represent AINNO. The influence of digital-intelligent synergy on ambidextrous innovation is still remarkably positive with the co-efficiency as 0.3836 (see column (1), Table 5).

To enhance robustness, we conducted an additional set of regression analysis, restricting the time limit to a narrower three-year window surrounding the policy's implementation (i.e. ± three years). As presented in Table 5, the results remain consistent and robust compared to the primary findings with the co-efficiency as 0.5044 (see column (2)).

Firms in the same city are likely to show similarities in terms of political and economic environment. Therefore, we added the city-year interaction term to the baseline model to control the influence of changes in each city over time. The result is still significantly positive with the co-efficiency as 0.4761 (see column (3), Table 5).

Another important policy implemented in 2015 that related to digital-intelligent policy was the “National Big Data Comprehensive Pilot Area”. We constructed a big data policy dummy variable (Bigdata) and included it in the regression test. The findings show a positive correlation between DIPS and AINNO with the co-efficiency as 0.4925 (see column (4), Table 5).

We chose the support vector machine (SVM) approach, the Boosting technique, and the Lasso regression model as replacements for the random forest regression used for the baseline regression. After using different models, the results are still significantly positive with the co-efficiency as 0.9493, 0.4129 and 0.1799 (see column (5), Table 5).

This paper also considers the correlation between industries. In the robustness test, the standard errors are bidirectionally clustered at the industry-year level, and the results do not show any substantial change. (see column (6), Table 5).

To further mitigate concerns regarding reverse causality, we employ the one period lagged variable of DIPS as the instrumental variable. The analysis employs Partially Linear Instrumental Variables (PLIV) to tackle the endogeneity issue. In contrast to traditional linear instrumental variable methods, PLIV facilitates a nonlinear treatment of the effects of control variables on both independent and dependent variables, thus enhancing robustness in high-dimensional settings. The findings indicate that the DIPS continues to be significantly positive with the co-efficiency as 0.4335 (see column (7), Table 5). Finally, a propensity score matching difference-in-differences (PSM-DID) approach is employed to mitigate potential sample selection bias. The results reported in column (8) of Table 5 remain to be significantly positive with the co-efficiency as 0.1431, confirming the positive impact of DIPS on ambidextrous innovation.

The previous sections have thoroughly demonstrated the enhancing effect of the dual policy in promoting ambidextrous innovation. To further reveal whether they mutually reinforce each other in promoting manufacturing enterprises' innovation, we conducted the following analysis. First, investigating the net effect of the single policy; second, comparing the total effect of the single policy with that of the dual policy; third, contrasting the net influence of the single and dual policy; and finally, exploring how the sequence of policy implementation impacts the synergistic effect. Through the analysis, we seek to offer a thorough assessment of how DIPS can empower ambidextrous innovation in enterprises, offering empirical support for governments to design more effective policy combinations.

The analysis first excludes the sample data of IMP implementation, and the BCP single policy is regarded as the experimental group, and the sample without any policy implementation is regarded as the control group to examine the net impact of the BCP. BCP significantly promotes exploitative innovation, but its effect on exploratory innovation is insignificant (i.e. columns (1) and (2), Table 6). The net effect of IMP is both significantly positive (i.e. columns (3) and (4), Table 6), but its effect on exploitative innovation is more notable than that on exploratory innovation. The stronger effect on exploitative innovation suggests that digital infrastructure and intelligent manufacturing policies primarily enhance a firm's ability to optimize existing technologies and production processes. In contrast, exploratory innovation, which holds more risks and requires higher absorptive capacity, tends to respond slowly to such policy interventions.

Second, we exclude all “no policy” samples, setting the dual policy sample as the experimental panel and the single policy sample as the control category. As listed in columns (5) and (6) of Table 6, the dual policy has a more remarkable incentive influence on ambidextrous innovation compared to a single instrument, and the synergistic effect is stronger on exploitative innovation as well. The combination of the BCP and IMP provides comprehensive digital infrastructure and intelligent technology support. These supports significantly reduce the technological costs and implementation challenges for enterprises engaging in exploitative innovation, thus directly and effectively promoting it. Since exploratory innovation carries higher risks and longer return cycles, the policy synergy provides more support for exploitative innovation than for exploratory.

Third, we exclude the “no policy” samples then further exclude samples that only receive the IMP, and compare the ambidextrous innovation incentive influences of the dual policy with the BCP. Similarly, we compare the dual policy with the IMP single policy. As listed in columns (7)–(10) of Table 6, the dual policy consistently delivers a stronger ambidextrous innovation incentive effect compared to either BCP or IMP instrument itself.

Last, two implementation paths are defined based on the sequence of BCP and IMP adoption. Path 1 means the BCP is implemented first followed by the IMP, while Path 2 denotes the opposite. In examining Path 1, enterprises that first became BCP pilots are treated as the experimental group, while enterprises without policy implementation are the control group. Path 1 in columns (11) and (12) of Table 6 demonstrates a notable positive effect on ambidextrous innovation. As shown in columns (13) and (14) of Table 6, Path 2 also promotes ambidextrous innovation. It is also noticed that the net effect of Path 2 is smaller compared with Path 1, as the difference between the coefficients of the two groups was statistically significant according to the Fisher's combined test (i.e. p-values of 0.002 and 0.030, respectively). This result shows that the order of policy implementation is also important. When BCP is implemented first and then IMP is implemented, the promotion effect of ambidextrous innovation will be more significant. This is likely because BCP usually focuses on the establishment of digital infrastructure, which can enhance the Internet penetration rate and network quality and provide the necessary infrastructure support for the implementation of IMP. This, in turn, facilitates the effective promotion of technological advancements. The IMP often requires substantial technological innovation. The successful deployment and diffusion of these technologies requires a robust network environment, which the BCP helps to establish.

A firm's ownership structure of manufacturing enterprises may introduce different responses to DIPS. Analyzing heterogeneity in ownership helps identify these differences and reveals how ownership influences innovation under policy conditions. Therefore, we classify firms into two categories, SOEs and non-SOEs, to observe how firms with different ownership structures respond to policy changes. As seen in Figure 2, the outcome shows that policy synergy significantly enhances ambidextrous innovation in both types of firms, with SOEs showing a slight advantage with a significant outcome for the difference in coefficients between groups. This may be because ambidextrous innovation usually requires substantial resource reserves, capacity accumulation and risk tolerance. With their industry-leading position, abundant cash flow and strong risk resistance, SOEs are more suitable for long-term value creation and strategic investment (Liu and Peng, 2023). In contrast, while policy incentives help non-SOEs alleviate short-term financial pressures and encourage innovation, their weaker cash flow and risk tolerance may lead them to be relatively conservative in ambidextrous innovation (Zhang et al., 2024b).

In order to test the possible divergent roles of DIPS for manufacturing enterprises at different stages, we separate the samples into growth, mature and decline groups (Dickinson, 2011). As listed in Figure 2, the promoting effect of policy synergy is more remarkable for growing and mature firms with remarkable coefficients of intergroup differences. The stronger effects of broadband and intelligent manufacturing policies in growing and mature firms can be explained by differences in resource endowments and absorptive capacity. These firms typically possess better financial resources, more advanced technological foundations, and stronger managerial capabilities, enabling them to effectively utilize digital infrastructure and translate policy support into ambidextrous innovation (Fonseca et al., 2022). In contrast, declining firms often face financial constraints, organizational rigidities, and limited innovation capacity, which weaken their ability to benefit from such policies. Moreover, these firms tend to prioritize short-term survival over long-term innovation investment. Consequently, the policy effects are not evident in firms in the declining phase, highlighting the importance of firm lifecycle in shaping policy effectiveness (MacCarthy et al., 2016).

Our analysis divides the samples into large-scale enterprise groups and small-scale enterprise groups to analyze the difference in the role of enterprise size in influencing DIPS on ambidextrous innovation (Xu et al., 2024). Figure 2 shows that policy coordination has a greater impact on innovation in large enterprises, with a significant difference in coefficients between groups, indicating that large-scale enterprises have a stronger advantage in promoting ambidextrous innovations. This might be because large-scale firms usually have more abundant resources, stronger R&D investment capabilities, and an ability to bear higher uncertainties and risks, thus giving them a greater incentive to engage in innovation activities. In addition, large and medium-sized enterprises are more efficient in optimizing existing technologies and making incremental improvements as they usually have mature production systems and better capabilities to cope with risks (Zhang et al., 2024b). Compared with large-scale enterprises, small-scale enterprises are constrained by limited resources and weaker digital capabilities, making it more difficult for them to fully leverage digital-intelligent transformation to enhance innovation (Xu et al., 2024).

We further divided firms into high-tech and non-high-tech groups to examine the heterogeneity of policy effects. As shown in Figure 2, the policy synergy effect is more pronounced in non-high-tech enterprises (HTEs) when it comes to innovation with significant inter-group coefficients. High-tech enterprises already possess strong innovation capabilities and substantial technological accumulation, with relatively high initial levels of innovation; therefore, the space for improvement through policy support may be relatively limited for these firms (Shi and Chang, 2024). In contrast, non-HTEs have a weaker existing innovation foundation, and policy support can effectively alleviate their resource and technological constraints, thereby incentivizing them to engage in more innovation activities (Jin et al., 2022). Moreover, the relevant policies help them absorb, improve, and reuse existing technologies, enhancing production efficiency and technological application capabilities, thereby enabling ambidextrous innovation and transformation and upgrading.

Relying on data from China's manufacturing listed companies, we employ a DML model to investigate the influence and mechanisms of the digital-intelligent policy synergy on firm-level ambidextrous innovation.

Our findings demonstrate that the implementation of the two policies noticeably enhances manufacturing enterprises' ambidextrous innovation. Additionally, compared to exploratory innovation, the policy has a greater promoting effect on exploitative innovation. The conclusion stands after a series of robustness checks. Moreover, resource accumulation, utilization, and allocation all serve as critical mechanisms between policy implementation and corporate innovation level, including both exploitative and exploratory innovation. Heterogeneity analysis shows that policy synergies are more pronounced among SOEs, growth and mature enterprises, large enterprises, and non-HTEs. Finally, the net effect of the BCP promotes exploitative innovation but has no significant effect on exploratory innovation, while the net effect of the IMP significantly promotes both exploratory and exploitative innovation. The effect of policy synergy is superior to any single policy effect, and the path where the BCP is implemented first followed by the IMP, has a greater influence on facilitating manufacturing enterprises' ambidextrous innovation than the reverse sequence.

First, we have expanded the discussion on the role of DIPS in promoting ambidextrous innovation. This study confirms that DIPS is a key driver for enhancing ambidextrous innovation in manufacturing enterprises. Moreover, we also compared the heterogeneity impacts of policy synergy on corporate exploitative and exploratory innovation, revealing the constraints in the process of transforming from efficiency-driven to innovation-breakthrough. Under a policy-driven context, the responses of the two types of innovation to external institutional stimuli are asymmetric. Exploitative innovation tends to be more sensitive to policy synergy, indicating that ambidextrous innovation does not advance in a simply synchronized manner, but rather exhibits differentiated response mechanisms.

Second, existing research has not explored how to enhance ambidextrous innovation through accumulation, utilization, and allocation from a resource-based view. This study, on the other hand, reveals the mechanisms of how DIPS promotes ambidextrous innovation from divergent dimensions of RBV. This study extends the resource-based view theory to the context of the impact of digital and intelligent policy synergy on corporate ambidextrous innovation, enriching the application of this theory in emerging research fields.

Third, we also reveal the heterogeneous effects of single-policy and dual-policy approaches, as well as the order of policy implementation, on dual-track innovation. This study extends the traditional framework that treats policies as independent shocks by proposing that policies must generate synergistic effects through complementarity and mutual reinforcement. Furthermore, the order of policy implementation has a differentiated impact on innovation outcomes, suggesting that policy effects are order-sensitive and exhibit temporal heterogeneity.

The government should strengthen the coordinated advancement of digital and intelligent policies and pay attention to the impact of policy combinations and implementation sequencing on innovation. Specifically, efforts should be made to improve digital infrastructure, reduce the cost of digital transformation for manufacturing enterprises, and encourage firms to prioritize exploitative innovation to improve production efficiency and resource allocation capabilities. On this basis, relevant policies on intelligent manufacturing should be further implemented to guide enterprises to increase investment in technology research and development and system integration applications, promote exploratory innovation, and achieve breakthroughs in key manufacturing technologies. In addition, policy formulation and implementation should adopt differentiated measures for different types of manufacturing enterprises.

Enterprises should proactively seize the opportunities presented by digital and intelligent policies, integrating digital transformation with intelligent manufacturing transformation. They should solidify their foundation through exploitative innovation and achieve breakthroughs through exploratory innovation. Specifically, manufacturing firms can improve innovation performance by strengthening resource accumulation, enhancing resource utilization efficiency, and optimizing resource allocation through digital and intelligent technologies, thereby providing sufficient innovation inputs, improving the efficiency of knowledge transformation, and directing resources toward high-value innovative activities.

This study still has certain limitations. First, the sample analyzed in this paper consists of publicly listed manufacturing companies; future research could expand the scope to include privately held companies, thereby enhancing the representativeness and coverage of the sample. Second, this paper only explores its mechanism from the perspective of resource endowment; subsequent research could further explore other potential mechanisms. Finally, given that this study is based on China's specific institutional and policy background, its conclusions are more applicable to economies with similar industrial structures and digital transformation paths. Future research could conduct comparative tests under different institutional environments to identify the boundary conditions and external validity of the research conclusions more comprehensively.

Xuena Gao: Conceptualization, Data curation, Formal analysis, Software, Writing-original draft; Wei Gu: Conceptualization, Supervision, Validation, Writing-review and editing, Funding acquisition; Xiaoling Wang: Conceptualization, Validation, Methodology, Funding acquisition, Writing-review & editing; Andrea Appolloni: Conceptualization, Supervision, Validation and Writing-review & editing.