Boost or burden? The nonlinear impact of artificial intelligence on innovation resilience within a dynamic capabilities framework: evidence from Chinese listed companies

1 Introduction

Faced with complex and dynamic environments characterized by sudden public health emergencies and accelerated technological iteration, corporate innovation activities face severe challenges. High-quality corporate development relies not only on static innovation paradigms but also on the resilience of innovation—the ability to withstand shocks, adapt, and achieve recovery and leapfrogging. Simultaneously, Artificial intelligence (AI) is reshaping the corporate ecosystem with unprecedented depth and breadth, creating new opportunities and promoting a series of innovations, including new technologies and products [1]. Therefore, how enterprises leverage AI to enhance their innovation resilience has become a popular topic with significant theoretical breakthrough potential and practical demand.

Existing research has widely acknowledged the positive role of AI in driving corporate innovation. Through pathways such as accelerating knowledge creation [2, 3], enhancing management efficiency and digital governance capabilities [4–6], AI injects new momentum into corporate innovation activities. However, the relationship between the two remains complex in practice. For instance, some enterprises experience rapid short-term innovation gains after AI adoption but face innovation stagnation during external shocks due to technological dependency and organizational rigidity. Conversely, other firms leverage AI to develop dynamic response mechanisms, successfully achieving innovation transformation amid crises. These phenomena suggest that the relationship between AI and innovation resilience may not be a simple linear correlation. Academic research on the factors influencing innovation resilience has focused primarily on digital transformation [7, 8], digital finance development [9–11], and innovation diversification [12], offering valuable perspectives for understanding how enterprises enhance innovation resilience under AI. The integration of artificial intelligence in the domain of autonomy has the capacity to substantially alleviate financial strain [13]. Such cost-reduction and capital-preserving effects provide a crucial foundation for firms to sustain R&D investment, pursue continuous technological innovation and consolidate corporate innovation resilience amid volatile market environments.

As external business environments grow increasingly uncertain and unpredictable, enterprises face escalating risks and challenges. Enhancing innovation resilience to navigate these external shifts has become critical for building sustainable competitive advantages and achieving high-quality development. Existing research predominantly focuses on “Smart+,” intelligent manufacturing, generative AI, and the optimization and integration of internal resources [14–16], often neglecting the embedded process of environmental dynamism. Considering the above mentioned findings, this research explores the nonlinear influence process of AI on corporate innovation resilience. We accurately identify the mediating mechanisms of the role of corporate learning and absorption capacity, coordination and integration capacity, and technological innovation capacity. We also examine the moderating mechanism, which reveals that environmental dynamism provides theoretical references and policy recommendations for advancing AI applications and enhancing corporate innovation resilience.

Our study makes three major contributions. Firstly, while prior studies have largely centered on the linear, unidirectional direct impact of AI on corporate innovation resilience [17, 18], this study empirically tests the inverted U-shaped relationship between AI and innovation resilience, thereby providing data support and decision-making references for enhancing innovation resilience through AI in practice. Secondly, it systematically reveals the intrinsic mechanisms and boundary conditions linking AI and innovation resilience. This study builds upon dynamic capability theory to construct a “technology-capability-behavior” framework that systematically validates the intrinsic mechanisms of firms' learning and absorption capability, coordination and integration capability, technological innovation capability. Concurrently, it incorporates environmental dynamism to expand the boundary conditions. Thirdly, it provides empirical support for leveraging AI to precisely reshape corporate innovation resilience and achieve high-quality development. On the one hand, it conducts AI’s heterogeneous impact on innovation resilience across three dimensions—firm ownership structure, industry differences, and scale variations—offering theoretical foundations for firms to enhance resilience tailored to their specific contexts and providing crucial references for policy formulation. On the other hand, the enabling effects of AI exhibit an optimal range, offering empirical evidence for corporate managers to avoid “technology application pitfalls” and achieve high-quality corporate development.

This paper is structured as follows for the subsequent sections: Section 2 sets forth the framework as well as the hypotheses put forward in this study; Section 3 outlines the research method system; Section 4 reveals the evidence-based findings and analysis; Section 5 shows further analytical conclusions; Section 6 shows the final implications of the findings and implications.

2 Theoretical model and research assumptions

2.1 Theoretical model

First emerging in ecology, the concept of “resilience” refers to the disturbance that an ecosystem can withstand before its function is altered. It pioneered the introduction of resilience into organizational research, defining it as an organization’s capacity to adapt rapidly in the face of adversity [19]. Subsequently, building upon organizational resilience studies, the concept was further applied to innovation systems. Innovation resilience is defined as an enterprise’s ability to sustain or achieve innovation, overcome adversity, and innovate against the odds when facing external crises [20, 21]. It categorized organizational resilience into “dynamic” and “static” perspectives on the basis of attributes and outcomes [22]. The dynamic perspective views resilience as a dynamic capability or developmental process, whereas the static perspective regards it as an ultimate state or performance outcome. This paper adopts the dynamic perspective, integrating the capability and process perspectives of innovation resilience and accordingly defines corporate innovation resilience as the capacity and dynamic process through which an enterprise absorbs disturbances, maintains stability, rapidly adapts, and restores innovation capabilities when subjected to uncertain shocks.

Dynamic capabilities theory emphasizes how enterprises cope with the fast-changing environment through the integration, building and reconfiguration of intra- and extra-organizational resources [23]. The dynamic capabilities constitute the critical process through which firms respond to environmental shocks [24, 25]. Hence, on the basis of dynamic capability theory, this study integrates the capability perspective and process perspective of innovation resilience to construct a theoretical framework that links AI applications with corporate innovation resilience. AI applications enhance enterprises' dynamic capabilities across three dimensions—“sensing opportunities, seizing opportunities, and reconfiguring resources”—thereby influencing their innovation resilience. Learning and absorption capacity serves as the “early warning radar” and “knowledge foundation” of the corporate innovation system. Coordination and integration capacity relates to innovation’s “execution capability” and “agility”. Technological innovation capability is the core of the innovation system’s evolution.

However, this influence does not follow a simple linear promotion relationship, but rather, it may exhibit complex nonlinear characteristics. This process accounts for environmental shocks to innovation resilience and underscores the need for enterprises to flexibly respond to external changes when adaptive strategies are formulated. The theoretical framework for AI and corporate innovation resilience is illustrated in Figure 1.

2.2 Research assumptions

2.2.1 Artificial intelligence and innovation resilience

At the early stage when AI is adopted, enterprises face lower trial-and-error costs, faster response times, and more robust operational buffers when confronted with disruptions. This reduces the complexity and uncertainty of information processing, thereby enhancing the ability to rapidly recover from operational disruptions and adapt flexibly to volatile risk environments [26–28]. Simultaneously, through technological identification and predictive analytics, AI enables faster and more precise strategy adjustments, allowing enterprises to capture market trends and technological frontiers, thereby enhancing innovation resilience [29, 30], and it provides new sources of momentum for navigating unstable environments and crisis scenarios, thereby achieving an enabling effect on innovation resilience [31–33].

As Al adoption progresses, discriminatory boundaries based on information asymmetry arise within companies between “people who can effectively use AI (those with cognitive capital)”and “people who cannot.' Furthermore, as algorithm standardization progresses, there is a risk that minority ideas and non-routine tasks will be excluded as “errors' or “inefficiencies' (structural exclusion by algorithms). When AI adoption surpasses a critical threshold, it imposes heightened demands on corporate organizational frameworks, resource allocation, along with management systems [34–36]. To avoid falling into a digital trap, enterprises must enhance their capabilities in tandem with the rapid advancement of digital technologies. For example, organizations may become overly reliant on AI-driven decision support systems while neglecting corporate culture, employee creativity, and other factors, thereby limiting the breadth and depth of innovation. Furthermore, the increasing complexity of AI technologies may heighten management challenges, while data security and privacy protection issues could impose additional risks and burdens on enterprises. As AI investments grow, their inhibitory effect on corporate innovation resilience gradually becomes more pronounced, outpacing the enabling effect and leading to a decline in innovation resilience.

H1 The nexus between AI and corporate innovation resilience takes an inverted U-shape, increasing at first and then decreasing.

2.2.2 The mediating role of learning and absorption capacity

Learning and absorptive ability is defined as an organization’s capability of identifying, acquiring, utilizing, and commercializing external knowledge. It is categorized into latent absorption capacity and actual absorption capacity [37]. Latent absorption capacity emphasizes knowledge acquisition and assimilation. When facing external environmental uncertainties, AI leverages natural language processing and knowledge graphs to rapidly capture cross-domain, cross-industry external knowledge, thus providing capabilities for swift data perception, capture, and dynamic allocation. This means that through AI’s network connectivity and data collection functions, enterprises can better access their internal data resources. Human‒machine collaboration technologies also facilitate rapid knowledge diffusion and widespread absorption through real-time data analysis and feedback, which helps enterprises swiftly identify and seize diverse business opportunities, thus significantly enhancing information processing efficiency and accelerating the internalization of corporate knowledge [38–40].

Actual absorption capacity, however, emphasizes knowledge transformation and utilization. AI possesses capabilities for memory, storage, and knowledge processing. Accordingly, it not only integrates scattered, unordered distributed fragments into structured knowledge systems [41, 42], but it also facilitates the conversion of implicit knowledge into explicit knowledge by adding tag information to tacit knowledge. This effectively broadens the scope of knowledge dissemination and enhances the knowledge absorption capacity of R&D personnel [43, 44].

When AI surpasses a certain threshold, excessive AI leads to “quantitative overload” and “qualitative dilution” in knowledge acquisition, as its standardized decoding approach limits a deep understanding of knowledge and thereby inhibits learning and absorption capabilities. Learning and absorption capability is clearly pivotal to innovation development, as this capability influences corporate innovation resilience.

H2 Learning and absorption capacity acts as a mediator in the link between AI implementation and the resilience of enterprise innovation.

2.2.3 The mediating role of coordination and integration capability

Coordination and integration capability refers to a company’s ability to transform resources into value, encompassing both internal resource coordination and external resource integration. On the one hand, internal resource coordination focuses on converting technical knowledge within the company, fostering innovation and development, and facilitating interdepartmental collaboration. In response to external environmental changes such as market demand, AI technology coordinates resources across R&D, production, marketing, and finance departments. This enables the automatic synchronization of order, logistics, and financial data along with anomaly alerts, thereby enhancing the enterprise’s perception of cutting-edge technological trends and market opportunities while achieving the efficient integration of innovation activities [45–47]. During the initial AI implementation phase, internal resource coordination enables companies to react more flexibly to market dynamics and challenges, thereby enhancing their coordination and integration capabilities.

On the other hand, external resource integration emphasizes collaboration with upstream and downstream suppliers, universities, government departments, and other entities. AI-built external resource matching platforms overcome physical space and cooperation boundary constraints, rapidly aligning resource demands between enterprises and external partners. This significantly expands the scope and depth of enterprise resource integration.

Once AI applications reach a certain level and enter the ecosystem application phase, they exceed the capacity of external coordination and integration. The enabling effect then shifts to an inhibiting effect, with internal coordination and integration capabilities reaching “adaptive saturation.” As a consequence, AI’s standardized processes conflict with internal innovation demands, leading to a “peak-and-decline” state.

H3 Coordination and integration capability acts as a mediator in the link between AI implementation and the resilience of enterprise innovation.

2.2.4 The mediating role of technological innovation capability

Technological innovation capability refers to dynamically responding to market changes and iterating innovation models. As such, it encompasses the capacity to effectively absorb, master, and refine existing technologies, as well as the skills and knowledge required to create new technologies based on this foundation [48, 49].

AI possesses highly efficient information retrieval and massive data processing capabilities, facilitating real-time data evaluation and enabling automated decision-making through intelligent algorithms. This capability not only helps enterprises gather abundant innovation information and identify and generate new ideas during the innovation process, but it also drives the transformation from traditional human-centric innovation models to human‒machine collaborative innovation models, thereby more effectively supporting innovation decision-making. AI technology can fully leverage existing information technologies to enhance operational efficiency and promote the development of exploratory and exploitative innovation [50–52].

Furthermore, a close connection exists between AI technology and innovation technology development. This relationship enhances AI’s potential with respect to perception, execution, and decision-making within R&D domains, thereby providing critical innovation momentum for enterprises. It is precisely through this continuous enhancement of technological iteration and perceptual capabilities that enterprises, by cultivating technological innovation capacity, can adapt flexibly and sustain innovation in volatile market environments [53, 54], ultimately influencing their innovation resilience.

When AI reaches a certain threshold, it may inhibit corporate technological innovation capabilities, potentially leading to excessive resource dispersion and reduced decision-making efficiency, thereby weakening innovation resilience. In summary, AI shapes and enhances corporate innovation resilience through the empowerment of technological innovation capabilities.

H4 Technological innovation capability mediates the relationship between AI and innovation resilience.

3 Research design

3.1 Sample selection and data sources

This study selects China’s listed companies in the Shanghai and Shenzhen A-share markets from 2010 to 2024. Data concerning innovation resilience is furnished by the CNIPA. The AI data, along with other relevant data, is furnished by the CSMAR database. To guarantee data quality, the initial data set underwent the following processing: (1) Exclusion of companies in the financial sector; (2) Exclusion of companies under ST and *ST; (3) Exclusion of samples with severe data omissions. Ultimately, the study obtained 23,213 observations. To mitigate the influence of outliers, all continuous variables were subjected to 1% trimming to eliminate extreme values.

3.2 Definitions of variables

3.2.1 Dependent variable

Innovation Resilience (EIR): Drawing upon the study [55], the innovation resilience index is established on the foundation of three dimensions: innovation input, innovation output and innovation efficiency. The entropy weight method is employed to calculate the entropy weight of each indicator, and the innovation resilience variable is then reweighted accordingly (as shown in Table 1). The formula is as follows: EIR = ω _*1* RI+ω _*2* RE+ω _*3* PA+ω _*4* IE。

TABLE 1. Indicators of corporate innovation resilience.

3.2.2 Independent variable

Artificial Intelligence (AI): Specifically, AI is characterized through four metrics: software investment, hardware investment, total investment amount, and total assets. The calculation process draws upon the research methodology [56], in which the annual total investment in AI and digitalization is obtained by directly summing “software investment” and “hardware investment.” To eliminate the impact of company size differences and facilitate cross-company comparisons, the “total investment” obtained in the third step is divided by the company’s “total assets” to derive a relative ratio, which is termed “artificial intelligence”.

3.2.3 Mediating variables

This study references the research [57] to measure learning and absorption capacity, coordination and integration capacity, and technological innovation capacity.

• Learning and Absorption Capacity (Ab): This is measured by R&D expenditure intensity and is calculated as follows: Ab = Total R&D Expenditure/Operating Revenue, where Ab reflects the potential to acquire, digest, and apply external knowledge to enhance internal learning capabilities.
• Coordination and Integration Capability (Co.): This is measured by the coefficient of variation (CV) of three expenditures—R&D, capital, and advertising—and is calculated as follows: CV = Standard Deviation of Three Expenditures/Mean. A lower CV indicates more balanced resource allocation and stronger coordination and integration capability; conversely, a higher CV suggests weaker coordination. To align the direction of the coefficient with coordination and integration capability, a negative value is adopted for the coefficient. A higher adjusted coefficient indicates enhanced coordination and integration capability, and Co. reflects the enterprise’s flexibility in dynamically allocating and coordinating resources.
• Technological Innovation Capability (In): This is calculated by summing the Z-score standardized values of two variables—R&D Intensity (Total R&D Expenditure/Operating Revenue) and the Technical Personnel Ratio (number of Technical Staff/Total Employees). Thus, the formula is as follows: In = Z (R&D Intensity) + Z (Technical Personnel Ratio), thereby quantifying the enterprise’s overall technological innovation level.

3.2.4 Moderating variable

Environmental Uncertainty (EU): To more accurately reflect the uncertainty faced by companies, drawing upon the study [58], we measure environmental uncertainty using the industry-adjusted standard deviation of a company’s sales revenue over the past 5 years.

3.2.5 Control variables

A series of controls are included to address possible biases linked to firm organizational structure and financial indicators, and governance-related variables [59–61]. The control variables are as follows: Firm Size (Size), Firm Growth (Growth), Largest Shareholder (Top1), Debt-to-Asset Ratio (Lev), Return on Assets (Roa), Dual Role (Dual), Largest Shareholder (Top1), Board Size (Board), Firm Age (Firmage), Ownership Structure (Soe), as well as whether the company’s auditor is among the Big Four (Big4). Additionally, to reduce the interference arising from industry and year effects, this study incorporates fixed effects for year and industry.

3.3 Model design

To verify how AI influences corporate innovation resilience, the following benchmark econometric model (Equations 1, 2) are constructed.

E I R i , t = β 0 + β 1 A I i , t + ∑ β C o n t r o l i , t + μ j + θ t + ε i , t (1)

E I R i , t = β 0 + β 1 A I i , t + β 2 A I i , t 2 + ∑ β C o n t r o l i , t + μ j + θ t + ε i , t (2)

Where the variable EIR_*i,t* measures the innovation resilience of firm i in year t, while the variable AI_*i,t* reflects the application of AI by firm i in year t. Additionally, Control_i,t incorporates a series of control variables to ensure the accuracy of the research findings. Where μ_j represents the firm-fixed effect, θ_t represents the year-fixed effect, and ε_i,t represents the error term.

To further test the mechanism underlying the innovation resilience of AI-enabled enterprises, this paper uses the stepwise regression method [62] to construct the subsequent mediation effect model (Equations 3–5) which are constructed based on Models (1) and (2): where Mediate_i,t denotes the mediating variable, while all other variables are defined consistently with Models (1) and (2).

M e d i a t e i , t = β 0 + β 1 A I i , t + ∑ β C o n t r o l i , t + μ j + θ t + ε i , t (3)

M e d i a t e i , t = β 0 + β 1 A I i , t + β 2 A I i , t 2 + ∑ β C o n t r o l i , t + μ j + θ t + ε i , t (4)

E I R i , t = β 0 + β 1 A I i , t + β 2 A I i , t 2 + β 3 M e d i a t e + ∑ β C o n t r o l i , t + μ j + θ t + ε i , t (5)

4 Empirical findings and analysis

4.1 Descriptive statistics

Table 2 presents a summary of the descriptive statistics. With respect to AI, the mini-recorded value registers 0.000, while the max reaches 5.640; additionally, the mean equals 0.030 and the standard deviation amounts to 0.110, thus indicating significant variation in AI. The minimum value for corporate innovation resilience was 0.440, the maximum was 7.110, the mean was 3.750, and the standard deviation was 3.780. These results indicate that corporate innovation resilience levels are generally low and exhibit considerable variation.

TABLE 2. Descriptive statistics.

4.2 Baseline regression results

The regression findings are detailed in Table 3, and Column (1) includes solely the core explanatory variables for AI, excluding fixed effects of industry and year. The primary term of AI is positively and significantly related to corporate innovation resilience (β = 0.445, p < 0.01). Column (2) includes only the primary and secondary AI core explanatory variables in the absence of controls for fixed effects associated with industry and year. The secondary AI term is significantly negatively correlated with innovation resilience (β = −0.256, p < 0.01). These results indicate an inverted U-shaped relationship between AI and innovation resilience. Column (3) includes only the core AI explanatory variable while incorporating control variables and controlling for industry and year fixed effects, whereas Column (4) adds the AI quadratic term to the regression model from Column (3). Because the AI quadratic term remains significantly negative, the validity of H1 is supported. Furthermore, plotting the association between AI and corporate innovation resilience (as shown in Figure 2) further confirms the relationship between the two.

TABLE 3. Benchmark regression results.

*, , and * indicate significance at the 1%, 5%, and 10% levels, respectively. The values in parentheses represent t-values. The same applies below.

To verify the dependability of the inverted U-shaped correlation, a three-step methodology advanced by Haans et al. (2016) [63] was applied to further assess the structure of this relationship. The results, as shown in Table 4, demonstrate that all inverted U-shaped relationships examined in this study satisfy the following conditions: (1) The regression coefficient for the quadratic term of artificial intelligence (AI²) is significantly negative; (2) The slope of the curve is positive at the minimum value point (X_L) and negative at the maximum value point (X_H); (3) The inflection point value (X_G) of the curve lies within the interval [X_L, X_H]. Thus, the relationship between the independent and dependent variables is once again validated.

TABLE 4. Re-examination of the inverted U-shaped relationship between AI and corporate innovation resilience.

4.3 Robustness tests

4.3.1 Replace the dependent variable

The measurement of innovation resilience is assessed by choosing key metrics that are responsive to responses to external shocks. The quantity of patent applications submitted by an enterprise has a close correlation with the R&D environment of the year in which it operates. Therefore, the number of patent applications is used as an important proxy indicator. This study draws on the measurement method of [20] to determine the level of innovation resilience of an enterprise by assessing the change in the number of patent applications. Model (Equations 6–8) are constructed as follows:

Δ P i = P i , t − P i , t − 1 (6)

Δ E = P c , t − P c , t − 1 / P c , t − 1 P i , t − 1 (7)

E I R i , t = Δ P i − Δ E / Δ E (8)

where EIR _i,t denotes the relative innovation resilience of firm i in year t; ΔP_i represents the actual change in number of patent applications of firm i from year t-1 to year t; ΔE denotes the predicted change in number of patent applications of the firm from year t-1 to year t, forecasted based on the overall change in number of patent applications in the city where the firm is located; P_i,t and P _*i,t*-1 are the number of patent applications of firm i in years t and t-1, respectively; and P_c,t and P_c,t-1 are the number of patent applications of the city where the firm is located in years t and t-1, respectively. The larger the value of EIR_i,t indicates stronger overall innovation resilience among enterprises in the corresponding city. As a whole, and vice versa With respect to EIR_i,t, a higher t value indicates stronger innovation resilience in enterprises, and vice versa. Column (1) of Table 5 presents the regression estimates. The regression coefficient value for AI² is −0.054, showing a statistically significant negative impact, with its significance meeting the 1% significance level, which further verifies the validity of the empirical results.

TABLE 5. Robustness tests - replacement variables.

4.3.2 Replace the explanatory variable

To further verify the reliability of the research findings, drawing on the research by Yao Jiaquan et al. [64], we use text analysis to construct an AI indicator. The natural logarithm plus ten of the number of AI keywords in listed companies’ annual reports (AI) is used as the firm-level AI indicator. Column (2) of Table 5 presents the regression estimates. The regression coefficient value for AI² is −0.027, showing a statistically significant negative impact, with its significance meeting the 5% significance level, which further verifies the validity of the empirical results.

4.3.3 Adding control variables

Considering the factors influencing corporate innovation resilience, the macro variable of GDP per capita was incorporated. As shown in Column (3) of Table 5, the regression coefficient for AI² was −0.246, indicating that AI remained significantly negative at the 5% confidence level. After testing, the conclusions of this paper remain robust.

4.3.4 Excluding abnormal years

To mitigate the impact of external shocks on corporate innovation resilience under ordinary economic conditions, we excluded particular periods from the analysis, e.g., the 2015 stock market crisis and the COVID-19 pandemic during the 2020–2021 period. As reported in Column (4) of Table 5, the coefficient on AI² is estimated to be −0.511 and significantly negative, with its significance meeting the 1% significance level. These results still hold robustly after relevant tests.

4.3.5 Excluding distortions from the high-tech sector

Considering that the high-tech industry is highly innovation-dependent and tends to prioritize investments in these activities to build and strengthen its innovation resilience, the high-tech sector is excluded from the sample to reduce potential bias in the findings. Column (5) of Table 5 shows the corresponding regression results, the estimated coefficient of AI² stands at −0.377, showing a significantly negative effect and reaching the 1% significance level. In summary, these findings confirm the robustness of the conclusions.

4.3.6 PSM-DID

Made in China 2025 advocates a policy of “technology-driven innovation with quality as the priority.” Under this policy, China has entered a new era of innovation-driven strategic development. This paper further employs the intelligent transformation of key industries as a quasi-natural experiment, utilizing the PSM-DID method to address endogeneity issues between AI adoption and corporate innovation resilience. Specifically, we divided the sample into an experimental group and a control group based on the presence or absence of the keyword “artificial intelligence” in the annual reports. We used the control variables from Model (1) as the matching criteria and employed a 1:1 matching-with-replacement nearest-neighbor matching method. First, balance tests were conducted. As shown in the kernel density distributions pre- and post-matching in Figures 3, 4, prior to the matching process, the intervention group and its control counterpart have negligible overlap; the overlap of the two groups is significantly enhanced after matching, reaffirming the accuracy of the matching process. Second, The regression results after propensity score matching (PSM) are reported in Column (1) of Table 6. The coefficient of AI² is −0.174, which is significantly negative at the 1% level. As revealed by this finding, after accounting for endogeneity, the conclusion regarding the impact of AI on corporate innovation resilience remains robust.

TABLE 6. Endogeneity test.

4.3.7 Using instrumental variables to address endogeneity

From a reverse causality perspective, stronger corporate innovation resilience is correlated with higher levels of new technology adoption. A higher correlation indicates a stronger corporate capacity to mobilize and allocate resources in support of AI implementation processes. This issue of endogeneity, however, may give rise to the misleading interpretations of the inverted U-shaped association. To address this issue, the instrumental variables (IV) method is employed. This study uses one-period and two-period lagged AI as instrumental variables (Li and Bai, 2024). The results for the one-period lagged variable are presented in Columns (2) and (3) of Table 6. In the first-stage estimation, the coefficient of the instrumental variable (IV) is significantly positive at the 1% significance level, with an F-statistic of 2,717.173, which exceeds the conventional threshold of 10, indicating no issues of identification failure or instruments of low relevance. In the second-stage estimation, the estimated value corresponding to the key explanatory variable (AI²) estimated using the IV method remains significantly negative. Similarly, Columns (4) and (5) of Table 6 present the results of the two-period lagged test. In the first stage, the instrumental variable (IV) is significantly positive at the 1% level, and the F-statistic is 1,392.909, exceeding the commonly used threshold of 10, indicating no issues of underidentification or weak instrumentality. In the second stage, the coefficient of the core explanatory variable AI² estimated using the instrumental variables method remains significantly negative. The instrumental variable regression findings reveal that, even after accounting for endogeneity issues, an inverted U-shaped nexus persists between AI and corporate innovation resilience. These results support the paper’s fundamental conclusion, supporting the robustness of H1.

4.4 Mechanism tests

First, we identified the mediating effect of learning and absorption capacity, with the conclusions reported in Table 7. Column (2) indicates that the AI primary term exhibits the impact on learning-absorption capacity at the 1% level. Furthermore, the coefficient on the AI quadratic term in Column (3) stands at −0.017, with statistical significance at the 1% level, confirming a nonlinear nexus between learning-absorption capacity and AI. Column (4) presents the regression of learning and absorption capacity on corporate innovation resilience. The regression coefficient for learning and absorption capacity is 4.603, which is statistically significant at the 1% level and confirms a linear association between learning and absorption capacity and innovation resilience. Column (5) includes the regression analysis of the effects of the AI first-order term, AI second-order term, and learning and absorption capacity on innovation resilience. The results indicate that the regression coefficient for the quadratic term of AI is significantly positive at the 1% statistical level, further confirming the nonlinear relationship. Compared with the estimated coefficient of AI² in Column (1), the coefficient of AI² increases from −0.250 to −0.173 after incorporating learning and absorption capacity. This change suggests that learning and absorption capacity exert a certain mediating effect, thereby validating H2.

TABLE 7. Mechanism test - learning and absorption capacity.

Second, the results of identifying the mediating effect of coordination and integration capability are presented in Table 8. Columns (2) and (3) of Table 7 display the regression analysis of the effects of the first-order and second-order terms of AI on coordination and integration capability. The second-order AI term carries an estimated coefficient of −0.056, significant at the 1% level, further verifying the nonlinear nexus between coordination and integration capability and AI. Column (4) presents the regression analysis of coordination and integration capability on corporate innovation resilience. According to the results, the Ad coefficient is positive, with significance reaching the 1% significance level, thereby verifying a linear nexus between coordination and integration capability and innovation resilience. Column (5) presents the regression of AI’s linear and quadratic terms along with coordination and integration capability on corporate innovation resilience. Compared with the estimated coefficient of AI² in Column (1), the inclusion of coordination and integration capability shifts the coefficient of AI² from −0.250 to −0.232. This change indicates that coordination and integration capability exerts a certain mediating effect, thereby validating H3.

TABLE 8. Mechanism test - Coordination and Integration Capability.

Finally, an empirical analysis examines technological innovation capability in linking AI with corporate innovation resilience. Table 9 demonstrates that technological innovation capability (In) yields a regression coefficient of 0.147 in Column (2), with its significance meeting the 1% significance level. This finding indicates that AI significantly promotes corporate technological innovation capability and provides robust theoretical underpinnings and empirical evidence to support enterprises in leveraging AI in driving innovation activities. To further verify the nonlinear nexus between AI and technological innovation capability, we regress technological innovation capability using the quadratic term of AI. Column (3) presents the regression of technological innovation capability using both linear and quadratic AI terms, with regression coefficients of 0.197 and −0.024, respectively. Both show statistical significance corresponding to the 1% and 10% levels, confirming the nonlinear relationship between AI and technological innovation capability. In Table 4, the regression coefficient for In is 0.044, a result that is positively significant at the 1% level, indicating a linear relationship between technological innovation capability and corporate innovation resilience. Column (5) presents the regression of AI’s linear and quadratic terms alongside technological innovation capacity on corporate innovation resilience. Compared with the estimated coefficient of AI² in Column (1), the inclusion of technological innovation capacity causes the coefficient of AI² to shift from −0.250 to −0.249. This change indicates that technological innovation capacity exerts a certain mediating effect, thereby validating H4.

TABLE 9. Mechanism test - technological innovation capability.

4.5 Moderating effects of environmental dynamics

Environmental dynamism refers to the degree of change in the external environment and the unpredictability arising from such changes [65]. To further analyze the moderation effect, we set the lower and upper bounds as one standard deviation below and above the moderator variable’s environmental dynamism [41], respectively, to explore how the association between artificial intelligence and innovation resilience varies with differing degrees of environmental dynamism. The regression results are presented in Table 10. Under high environmental dynamism, Columns (1) and (2) present the regression results of AI’s effect on corporate innovation resilience, with Column (1) excluding variable control and Column (2) including variable control, respectively. Columns (3) and (4) show similar findings under low environmental dynamism, indicating that when environmental dynamism is high, the quadratic term coefficient for AI is significant, whereas it is insignificant under low environmental dynamism. These findings suggest that high levels of environmental dynamism strengthen the inverted U-shaped nexus between AI and innovation resilience.

TABLE 10. Moderating effects test.

5 Further analysis

5.1 Heterogeneity analysis

5.1.1 Heterogeneity of property rights

Based on enterprise ownership, the sample was divided into two groups: state-owned enterprises and non-state-owned enterprises. As shown in Columns (1) and (2) of Table 11, the coefficients of the AI interaction term are statistically significantly negative at the 1% level for both types of enterprises, which is consistent with the research Hypothesis H1. Further analysis testing for between-group differences regarding property rights heterogeneity (state-owned/nonstate-owned enterprises) yields an empirical p value of 0.000. These findings indicate that the increase in innovation resilience varies significantly depending on the enterprise’s state-owned status, accompanied by the positive influence of AI adoption on innovation resilience being more significant among state-owned enterprises.

TABLE 11. Heterogeneity analysis.

5.1.2 Industry heterogeneity analysis

The sample was divided into high-tech enterprises and nonhigh-tech enterprises. The estimated coefficients for the AI interaction term, as illustrated in Columns (3) and (4) of Table 11, exhibit a statistically significant negative sign, and this significance aligns with the 1% level in both subsamples. Consistent with Hypothesis H1, this pattern is in line with prior scholarly work. Further analysis reveals that the results of the intergroup difference test for industry characteristics (high-tech/nonhigh-tech enterprises) yield an empirical p value of 0.000. These findings suggest that high-tech enterprises exhibit a more pronounced promoting effect and that the inherent reason for this may be that new technology industries exhibit higher factor intensity, making enterprises more sensitive to factor mobility, particularly in terms of high-end production factors.

5.1.3 Enterprise size heterogeneity

Using the median grouping method, observations are categorized as either large-scale or small-scale enterprises in terms of whether their total assets exceed the sample median for the given year. Columns (5) and (6) of Table 11 present the corresponding regression results. Column (5) corresponds to large-scale enterprises, where the AI quadratic coefficient is −0.598 and is significant. Column (6) corresponds to small-scale enterprises, the coefficient remains identical at −0.157 and attains statistical significance corresponding to the 1% level. Furthermore, testing for between-group differences in firm size heterogeneity (large-scale enterprises/small-scale enterprises) yields an empirical p value of 0.000. These findings indicate that AI exerts a more pronounced promotional effect on innovation resilience among large-scale enterprises.

5.2 Economic consequences of enterprises' innovative resilience

The following model (Equation 9) is constructed to examine how AI-enabled innovation resilience influences high-quality economic development in enterprises.

O P i , t = β 0 + β 1 A I i , t + β 2 E I R i , t + β 3 A I × E I R + ∑ β C o n t r o l i , t + μ j + θ t + ε i , t (9)

OP denotes the high-quality economic development of enterprises, AI represents AI, and EIR signifies enterprise innovation resilience. Consistent with the variable definitions in Model (1), other variables are set identically. Corresponding regression analysis findings are documented in Table 12. Column (1) shows the effect of AI on high-quality enterprise development after supplementary control variables are incorporated. The results show that the application of AI contributes to enhancing high-quality enterprise development. The results concerning the interactive impact of AI and firm innovation resilience on high-quality development, before and after controlling for a series of factors, are exhibited in Columns (2) and (3). Moreover, the interaction effect term displays positive coefficients, demonstrating that AI empowers enterprise innovation resilience, and this empowerment ultimately contributes to enhancing current-period performance.

TABLE 12. Economic consequences of high-quality economic development of enterprises.

5.3 Economic consequences of employee turnover and organizational inertia

To clarify the underlying logic behind the decline in various corporate capabilities after AI investment exceeds a critical threshold—specifically, the negative effect of the quadratic term of AI investment—we incorporated two supplementary variables, employee turnover rate and organizational inertia, into our empirical analysis. This approach explains the underlying drivers of the downward slope of the inverted U-curve at the data level and further explores the deeper internal causes of the inverted U-shaped relationship between AI and corporate innovation resilience. As shown in the results of columns (1) and (2) in Table 13 below, excessive application of AI promotes employee turnover, confirming that overuse of AI increases employee turnover, which in turn affects corporate innovation resilience. As shown in the results of columns (3) and (4) in Table 13 below, excessive application of AI promotes organizational inertia, confirming that overuse of AI fosters organizational inertia, which in turn affects corporate innovation resilience.

TABLE 13. Economic consequences of employee Turnover.

6 Conclusions and implications

6.1 Conclusions of the study

This study employs panel data from Chinese listed companies between 2010 and 2024 and uses a dynamic capability framework to analyze the effect of AI on corporate innovation resilience. AI exhibits an inverted U-shaped relationship with corporate innovation resilience, whereby it initially increases initially before decreasing, a pattern that holds after robust validity checks. Mechanism analysis indicates that AI enhances innovation resilience by influencing learning and absorption capacity, coordination and integration capacity, and technological innovation capacity. Heterogeneity analysis reveals that the resilience-enhancing effect of AI is more pronounced for state-owned enterprises, high-tech firms, and large-scale companies. The research further indicates that environmental dynamism constitutes a boundary condition for AI’s influence on innovation resilience, positively moderating the relationship between AI and innovation resilience. Artificial intelligence plays a crucial role in enhancing innovation resilience, which significantly promotes high-quality corporate development. In addition, by analyzing other variables such as “employee turnover” or “corporate inertia,” we can explain why performance declines once AI investment exceeds a certain threshold (i.e., the negative effects of AI²), thereby revealing the underlying causes of this downward trend.

6.2 Practical implications

This study also proffers several practical implications. First, the foundational enabling role of AI should be strengthened to solidify the underlying enhancement of enterprises’ innovation resilience. On the one hand, enterprises should adhere to the “principle of moderation” in AI investment, avoiding both under investment and over investment that could lead to technological lock-in. On the other hand, enterprises should proactively strengthen foundational AI support by increasing their investments in computing infrastructure, deepening the accumulation and governance of data assets, tackling core large-model technology R&D, and continuously solidifying their AI technological foundations to build a robust base for enhancing innovation resilience. Simultaneously, machine learning and deep learning technologies should be leveraged to establish AI-based dynamic risk early-warning mechanisms for real-time monitoring and risk assessment.

Second, dynamic corporate capabilities should be cultivated to activate AI-driven innovation resilience mechanisms. For example, enterprises should leverage AI technologies to efficiently identify, screen, and internalize external technical information and market signals, thus accelerating the effective integration and transformation of explicit and tacit knowledge. Simultaneously, AI-based organizational learning platforms should be established; internal education, training, and knowledge sharing should be prioritized; and enterprises should utilize AI to optimize internal learning mechanisms and continuously enhance their adaptability and innovation resilience. Furthermore, human‒machine collaborative decision-making mechanisms should be developed through intelligent collaboration platforms, and smart algorithms should be employed to optimize internal coordination across R&D, manufacturing, and marketing processes, while also enhancing external collaboration with suppliers, partners, and platform users.

Third, differentiated policy support should be implemented to build a dynamically adaptive external environment for innovation resilience. Government departments should formulate targeted support policies on the basis of enterprise scale, ownership structure, and technological characteristics. High-tech enterprises, for instance, require strengthened fiscal guidance, tax incentives, and risk compensation mechanisms to support breakthroughs in frontier and core technologies, whereas large-scale enterprises should be encouraged to leverage their role as “chain leaders” in industrial ecosystems and spearhead collaborative innovation ecosystems. Moreover, state-owned enterprises must break free from traditional institutional constraints and vigorously advance market-oriented reforms to fully unleash the potential of AI to enhance innovation resilience.

Fourth, excessive AI application may trigger hidden risks of internal stratification and discriminatory barriers within enterprises. As companies increase their reliance on Al, “reasonable accommodations' (ensuring human intervention, guaranteeing in reskilling, and designing human-in-the-loop mechanisms) are essential for employees left behind by technology and stakeholders who suffer disadvantages due to biases in Al models. Instead of operating a company solely on Al optimization logic, it should be suggested that establishing a safety net (buffer) to protect diverse knowledge will ultimately lead to maintaining the organization’s innovation resilience. From the perspective of managerial implications, it is critical to implement flexible safeguard measures including employee reskilling and diversified knowledge protection to defuse the risk of organizational segmentation and sustain corporate innovation resilience.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Jane G. Payumo, Michigan State University, United States

Reviewed by: Yasuko Kawahata, Rikkyo University, Japan

Muhammad Qamar Rasheed, Beijing Institute of Technology, China