Given the long biological cycles in agriculture and the high sensitivity of the sector to natural shocks, substantial endogeneity often arises between the input choices of agricultural enterprises and unobserved productivity shocks. Standard ordinary least squares estimation fails to account for this simultaneity and consequently leads to biased coefficients. Furthermore, the Levinsohn and Petrin semi-parametric approach uses intermediate inputs as a proxy variable, unlike the Olley and Pakes method that requires non-zero investment observations. This approach is particularly suitable for agricultural enterprises where investment behavior is often lumpy and discontinuous. In contrast, the usage of intermediate inputs such as feed and fertilizers is continuous and highly responsive to environmental signals. Therefore, following (Lu and Lian, 2012), this study employs the Levinsohn and Petrin method to ensure the accuracy of TFP estimates in the agricultural context. The calculation proceeds as follows Equation 3–6:

First, assume that the agricultural production function follows a Cobb–Douglas specification:

Where Y it denotes the total output (operating revenue) of agricultural enterprise i in year t , and A it represents its productivity (TFP). K it is capital input, measured by net fixed assets; L it is labor input, measured by the number of employees; and C it is intermediate input, including expenditures on goods purchased and services received. The parameters η , β , and γ denote the output elasticities of capital, labor, and intermediate inputs, respectively. ξ it is the error term.

For empirical estimation, the Cobb–Douglas production function is expressed in logarithmic form as follows:

In traditional OLS regressions, the endogeneity of capital and labor may lead to biased estimates of productivity. The LP method addresses this issue by introducing intermediate inputs C as proxy variables, thereby eliminating the endogeneity bias arising from the correlation between capital, labor, and productivity. Specifically, the LP approach assumes that intermediate input C it can serve as a proxy for unobserved productivity A it , replacing investment data in the estimation. The corresponding regression model is specified as follows Equation 5:

Based on the above model, the output elasticities of the production factors, η , β , and γ , can be obtained. Using these estimated elasticities, the TFP of each enterprise at each point in time is then calculated from the regression results.

Following Yao et al. (2024), this study measures the extent of AI adoption among agricultural enterprises using the natural logarithm of 1 plus the number of AI-related keywords appearing in firms’ annual reports. Importantly, keyword frequency is interpreted not as symbolic or rhetorical emphasis, but as a proxy for substantive AI adoption, because repeated disclosure of highly specific technical terms typically reflects underlying investments in intelligent equipment, data infrastructure, and algorithmic applications rather than short-term strategic communication.

To better capture AI applications specific to the farm sector, we draw on policy documents such as the National Smart Agriculture Action Plan (2024–2028) and the Digital Agriculture and Rural Development Plan (2019–2025) to identify key terms related to agricultural intelligent technologies, such as “intelligent,” “smart,” “drone,” “automation,” “RFID,” “navigation,” “precision,” “sensor,” and “remote sensing.”

Building on the general AI dictionary developed by Yao et al. (2024), we supplement it with 37 agriculture-specific AI keywords to construct a domain-tailored AI dictionary that more accurately reflects the technological characteristics of agricultural enterprises. Table 1 reports the complete list of AI keywords used for measurement.

Selecting an appropriate AI metric for agriculture presents unique challenges. In a regulated disclosure environment, opportunistic or purely strategic disclosure of narrowly defined AI-related technical terms is costly and difficult to sustain over time, which strengthens the link between keyword frequency and actual technological deployment. Unlike manufacturing where automation is often embodied in standardized hardware such as robotic arms that can be easily quantified, agricultural AI adoption increasingly involves intangible assets and service-based solutions. These solutions include algorithm-driven precision breeding and satellite data services which are difficult to capture through traditional fixed-asset statistics. Textual analysis effectively circumvents this measurement bottleneck. By systematically tracking keyword frequencies in annual reports, we can capture the strategic attention and substantive deployment of these soft AI technologies that hardware-based metrics might miss.

Moreover, our approach utilizes textual analysis of annual reports, which offers distinct methodological advantages over alternative measures. Traditional survey data often suffer from the “Halo Effect” (Thorndike, 1920), where subjective responses are inadvertently biased by a firm’s financial performance (Rosenzweig, 2007). In contrast, our metric is theoretically grounded in Signaling Theory (Spence, 1973). In a strictly regulated capital market, the continuous high-frequency disclosure of specific technical terms represents a credible signal of substantive resource commitment, as the reputational and legal costs of false signaling are prohibitive. To further rigorously verify this validity, we conducted extensive cross-checks using patent data as a “hard” output indicator. The consistent results confirm that our text-based metric effectively captures actual AI capabilities. This finding is corroborated by Yao et al. (2024), who similarly validated the reliability of textual indices by demonstrating their strong consistency with external lists of digital economy enterprises, thereby ruling out the concern of mere strategic disclosure.

Table 2 summarizes the disclosure of artificial intelligence (AI)–related information in the annual reports and Management Discussion and Analysis (MD&A) sections of agricultural listed companies from 2007 to 2023. The data show that the share of farming firms mentioning AI in their annual reports increased from 14.63% in 2007 to 78.34% in 2023, while the proportion disclosing AI in MD&A sections rose from 9.76 to 61.78% over the same period. This pronounced upward trend reflects the rapid expansion of AI applications within the agricultural sector.

Moreover, during 2007–2023, approximately 57.75% of agricultural enterprises disclosed AI-related information in their annual reports—substantially higher than the average disclosure rate reported by Yao et al. (2024) for all listed firms. This discrepancy underscores the importance and urgency of investigating the development, adoption, and implications of AI in agriculture.

This study incorporates several control variables to account for firm-level characteristics that may influence total factor productivity. Firm size (SIZE) is measured as the natural logarithm of total assets. Firm age (AGE) is calculated as the natural logarithm of the observation year minus the firm’s establishment year. Leverage (LEV) is defined as the ratio of total liabilities to total assets. Profitability (PRO) is measured as the ratio of net profit to total assets. Operating capability (OPER) is captured by the ratio of operating revenue to total assets. The number of board members measures board size (BOARD), while the proportion of independent directors (INDEP) is defined as the ratio of independent to total directors. The largest shareholder’s shareholding ratio measures ownership concentration (TOP). Tobin’s Q (TOBIN) is calculated as the ratio of the market value of assets to their book value. CEO duality (DUAL) is a dummy variable equal to 1 if the CEO simultaneously serves as board chair and zero otherwise. R&D investment (RD) is measured as the ratio of R&D expenditure to total assets.

Table 3 reports the descriptive statistics of the main variables. Between 2007 and 2023, the mean TFP of agricultural listed companies is 8.217, with a standard deviation of 1.143. The mean value of AI adoption is 1.170, with a standard deviation of 1.170, indicating substantial heterogeneity across firms in both productivity levels and AI deployment.

Although the instrumental variable approach helps mitigate endogeneity arising from reverse causality, omitted variables, and measurement error, it cannot fully address sample-selection bias resulting from the non-random adoption of artificial intelligence (AI) by firms. To further strengthen the robustness of the findings, this study therefore employs propensity score matching (PSM) as an additional empirical check. Using the control variables from the baseline regression as matching covariates, firms are matched using 1:1 nearest-neighbor matching with replacement. The balance-test results indicate that the absolute standardized differences in firm characteristics between the treatment and control groups decline substantially, and t-tests fail to reject the null hypothesis of no significant differences in covariate means across the two groups, confirming that the matching quality is satisfactory.

Column (3) of Table 5 reports the regression results based on the matched sample. After controlling for potential self-selection bias, the estimated coefficient on AI remains positive and statistically significant at the 1% level, providing further support for the robustness of the main findings.

To further verify the effect of AI on firms’ TFP, this study recalculates TFP using the Olley–Pakes (OP) method and ordinary least squares (OLS), replacing the baseline TFP measure with these alternatives as a robustness check. Columns (1) and (2) of Table 6 report the regression results based on the OP- and OLS-derived TFP measures, respectively. In both cases, the coefficient on AI remains positive and statistically significant at the 1% level, indicating that AI consistently promotes TFP in agricultural enterprises. These results reaffirm the validity of Hypothesis 1.

To further assess the robustness and reliability of the core explanatory variable, this study replaces the original AI measure with two alternative indicators following Yao et al. (2024): (i) the logarithm of the frequency of AI-related keywords appearing in the MD&A section of firms’ annual reports (AI1), and (ii) the logarithm of the number of AI-related patents granted to the firm (AI2). Columns (3) and (4) of Table 6 report the corresponding regression results. Regardless of whether MD&A keyword frequency or the number of AI patents is used as the proxy for AI adoption, the estimated coefficients remain positive and statistically significant at the 5% level. These findings provide additional confirmation of the strong robustness of the conclusion that AI enhances the total factor productivity of agricultural enterprises.

To further test the robustness of the empirical results, this study adjusts the sample to account for potential confounding factors. First, agricultural firms engaged in food processing are excluded. Relative to other agricultural enterprises, food-processing companies differ markedly in market demand structures, production processes, and innovation trajectories, and the incentives and mechanisms underlying their adoption of artificial intelligence (AI) may not be comparable. Second, observations from 2007 and 2008 are removed. On the one hand, the Accounting Standards for Business Enterprises were newly implemented in 2007, and disclosure practices were not yet fully standardized. On the other hand, the adoption of AI in agriculture is subject to a diffusion lag, making early years less representative of actual technology use.

Columns (5) and (6) of Table 2 report the estimation results after excluding food enterprises and after dropping the 2007–2008 observations, respectively. In both cases, the coefficient on AI remains positive and significant at the 1% level, indicating that the baseline findings are robust to alternative sample specifications.

To examine the labor-structure adjustment mechanism, this study follows the approach of Autor et al. (2003). Routine-skill labor (ROU) is measured as the number of employees with a junior-college degree or below, divided by total employment. Non-routine skill labor (NONROU) is calculated as the number of technical and R&D personnel relative to the total number of employees. These two variables are then incorporated into Model (2) as mechanism variables for regression analysis.

Column (1) of Table 7 shows that the coefficient on artificial intelligence is significantly negative, indicating that AI adoption reduces agricultural enterprises’ dependence on routine, low-skilled labor. Column (2) reports a significantly positive coefficient on AI, suggesting that AI increases the demand for non-routine, high-skilled labor. Taken together, these results demonstrate that adjustments in labor structure constitute an essential channel through which AI enhances the total factor productivity of agricultural enterprises.

To examine whether artificial intelligence (AI) enhances total factor productivity (TFP) through strengthened collaborative R&D, this study follows the approach of Long et al. (2023). Two indicators are constructed: whether a firm’s technological patents in a given year are jointly filed with research institutions (COOPER1), and whether they are jointly filed with other firms (COOPER2). These variables are then incorporated into Model (2) as mechanism variables.

Testing the theoretical channels proposed in Figure 1 (Section 2), the empirical results in Table 7 reveal a significant structural divergence within the collaborative R&D mechanism.

First, in Column (4), the coefficient on AI is significantly positive, indicating that AI effectively promotes technological exchange and resource sharing among enterprises. Drawing on the Transaction Cost Theory (Coase, 1937), we posit that AI technologies, exemplified by intelligent matching algorithms and automated quality grading systems, drastically reduce information asymmetry and search costs between upstream and downstream enterprises. This digital connectivity allows for immediate efficiency gains in resource integration and market-based coordination.

Second, column (3) of Table 7 reports the regression results for the collaborative R&D mechanism. Interestingly, unlike the significant positive effects observed in labor structure optimization and resource sharing, the coefficient on artificial intelligence is statistically insignificant regarding university-industry collaborative R&D. It is important to emphasize that this insignificant result should not be interpreted as a failure or ineffectiveness of AI technologies themselves. Rather, it reflects deep-rooted structural and institutional constraints that currently limit the role of AI in fostering university-industry R&D collaboration within the agricultural sector.

This insignificant finding offers a critical insight into the structural limitations of agricultural digitization. While AI effectively reduces information asymmetry in standardized market transactions (as seen in Section 4.4.1), it faces substantial friction when penetrating the upstream innovation chain. University-industry collaboration in agriculture is heavily shaped by institutional arrangements, incentive misalignment, and governance mechanisms, which cannot be easily altered by technological adoption alone. The translation of laboratory-based agronomic breakthroughs into commercial applications involves long biological validation cycles and complex field-testing requirements-a gap often referred to as the “Valley of Death” (Markham, 2002).

Current AI applications in agriculture primarily focus on application-layer efficiency (e.g., automated harvesting) rather than the deep-layer scientific validation required for collaborative R&D. As a result, AI tends to complement downstream production and management activities more effectively than upstream knowledge co-creation processes. Consequently, a “digital mismatch” exists where digital tools have optimized operational management but have yet to effectively bridge the cognitive and institutional gaps between academic research and agricultural production. This mismatch highlights that the limited impact of AI on university-industry R&D collaboration arises from the organizational and institutional environment of agricultural innovation systems, rather than from technological constraints per se.

This underscores the urgent need for further policy support to refine mechanisms for transforming scientific research outputs, thereby fully unleashing the potential of AI to advance agrarian modernization.