Generative AI risk prediction capability and cross border agricultural trade resilience: a PLS-SEM empirical study of coastal Chinese enterprises

1 Introduction

The global agricultural trade landscape has undergone profound transformations in recent years, marked by escalating geopolitical tensions, climate-induced production volatility, and proliferating non-tariff trade barriers (Food and Agriculture Organization, 2022; Liu and Fu, 2024). Because roughly 80% of world merchandise trade by volume moves by sea (UNCTAD, 2024), these pressures concentrate in the maritime leg of the food supply chain: port congestion, ocean freight rate volatility, and disruptions at maritime chokepoints such as the Strait of Malacca, the Suez Canal, and the Black Sea corridor can interrupt agri-food flows within days (Bailey and Wellesley, 2017; Notteboom et al., 2021). Cross-border agricultural enterprises face unprecedented operational complexity, where disruptions in one geographic node can cascade through entire supply networks, threatening food security and commercial viability (Chowdhury and Quaddus, 2017; Dolgui et al., 2020). The COVID-19 pandemic, the Russia-Ukraine conflict, and escalating trade disputes between major economies have collectively exposed the fragility of conventional risk management approaches that rely predominantly on historical data and human judgment (Skalkos, 2023; Queiroz et al., 2020).

In this turbulent environment, generative artificial intelligence (GenAI) has emerged as a transformative technological force with the potential to reshape how agricultural enterprises anticipate, assess, and respond to cross-border trade risks (Boone et al., 2025; Toorajipour et al., 2021). Unlike traditional predictive analytics, GenAI systems can synthesize heterogeneous data streams—including policy texts, tariff regulations, climate projections, vessel and port logistics telemetry, and overseas market demand signals—to generate probabilistic risk assessments and actionable early warnings (Ji et al., 2023). This capability is particularly valuable for seaborne agricultural trade, where risk factors span biological, meteorological, political, and economic domains and where disruptions are frequently mediated by shipping networks and port operations (Cromwell et al., 2025).

This maritime embeddedness also motivates the outlet for this study. The enterprises we surveyed are headquartered in four coastal provinces—Shandong, Guangdong, Fujian, and Zhejiang—whose crossborder agricultural trade is overwhelmingly seaborne and routed through some of the world’s busiest container and bulk ports (e.g., Qingdao, Shanghai–Ningbo, Shenzhen, and Xiamen). For these firms, the most consequential trade risks are maritime in nature: port congestion and berth delays, ocean freight rate spikes, reefer container shortages for perishable cargo, and chokepoint disruptions (Notteboom et al., 2021; Bailey and Wellesley, 2017). Several items of our measurement instrument therefore explicitly reference port disruption, logistics suspension, and perishable-product deterioration (see Section 3.2), and the findings speak directly to marine affairs and policy questions concerning the digitalization of port-centric supply chains and the governance of maritime trade risk. We return to these implications in the Discussion.

Despite the growing enthusiasm surrounding GenAI applications in supply chain management, substantial gaps persist in the empirical literature. First, existing studies on supply chain resilience have predominantly examined information technology adoption broadly (Wang et al., 2023; Gu et al., 2021), without isolating the specific mechanisms through which GenAI’s risk prediction capabilities contribute to resilience outcomes. Second, the resource-based view (RBV) (Barney, 1991) and dynamic capability theory (Teece et al., 1997; Eisenhardt and Martin, 2000) provide compelling theoretical lenses for understanding how GenAI-RPC may constitute a valuable, rare, inimitable, and organizationally embedded resource, yet few empirical studies have operationalized GenAI capabilities within these frameworks in the agricultural trade context. Third, while scales for measuring supply chain resilience exist (Chowdhury and Quaddus, 2017), no validated instrument has specifically captured GenAI-RPC as a multidimensional construct in cross-border agricultural settings.

This study addresses these gaps by developing and empirically testing a structural model that examines how three dimensions of GenAI-RPC—multi-source cross-border risk identification and scoring capacity, prediction accuracy and anti-hallucination stability capacity, and AI risk model calibration and decision empowerment capacity—are associated with cross-border agricultural trade resilience (CATR). Drawing on RBV and dynamic capability theory, we posit that these GenAI-RPC dimensions represent distinct but complementary facets of a dynamic capability that enables agricultural enterprises to sense, verify, and transform their operational approaches in response to cross-border trade disruptions (Teece, 2007). We tested our hypotheses using primary survey data from 300 agricultural enterprises across four coastal Chinese provinces (Shandong, Guangdong, Fujian, and Zhejiang) and employed partial least squares structural equation modeling (PLS-SEM) as the analytical method.

This study makes several contributions. Theoretically, it advances the integration of GenAI research with supply chain resilience theory by operationalizing GenAI-RPC as a multidimensional construct and empirically validating its association with CATR. Methodologically, it develops a reliable and valid measurement instrument for GenAI-RPC, grounded in the academic literature on AI-enabled supply chain risk management and refined through expert review and pilot testing (Chowdhury and Quaddus, 2017; Boone et al., 2025; Ji et al., 2023; Mikalef and Gupta, 2021). Practically, it provides evidence-based guidance for agricultural enterprises seeking to invest in GenAI capabilities and for port authorities and policymakers aiming to promote digital infrastructure for seaborne agricultural trade.

2 Literature review and hypotheses

2.1 Generative AI risk prediction capability

Generative AI refers to a class of artificial intelligence systems capable of creating novel content by learning patterns from large training datasets (Goodfellow et al., 2014). In the context of risk management, GenAI systems offer distinctive capabilities that transcend traditional predictive models: they can process unstructured data (e.g., policy documents, news articles, social media), generate human-readable risk narratives, simulate hypothetical scenarios, and continuously update predictions based on incoming information (Boone et al., 2025).

At the outset, it is important to distinguish three related but analytically distinct concepts: GenAI adoption, GenAI capability, and GenAI technical performance. Adoption refers to the organization’s decision to deploy GenAI tools; capability refers to the organization’s perceived capacity to mobilize those tools for a specific task domain—here, cross-border risk prediction; and technical performance refers to model-level properties (e.g., precision, hallucination rate) that can only be established through technical benchmarking. Our construct, GenAI-RPC, is a capability construct in the sense of Mikalef and Gupta (2021): it captures organizational perceptions of what deployed GenAI systems enable the firm to do, not the technical accuracy of the underlying models. This distinction informs both our measurement approach (Section 3.2) and the interpretation of the results (Section 5.5).

We conceptualize GenAI-RPC as a multidimensional capability encompassing three interrelated capacities. Multi-source cross-border risk identification and scoring capacity refers to the ability of GenAI systems to automatically aggregate heterogeneous data sources relevant to cross-border agricultural trade—including policy texts, tariff rules, climate data, shipping and port logistics status, and overseas market demand—and synthesize these inputs into structured risk probability scores (Boone et al., 2025; Toorajipour et al., 2021). This dimension captures the breadth of data integration and the depth of analytical processing, and corresponds to the sensing microfoundation of dynamic capabilities (Teece, 2007).

Prediction accuracy and anti-hallucination stability capacity reflects the perceived reliability of GenAI-generated risk predictions, particularly the system’s ability to minimize hallucinations (fabricated information presented as factual) and to maintain consistent output quality when facing incomplete or missing data (Ji et al., 2023). This dimension addresses a critical concern in GenAI deployment: the tendency of large language models to generate plausible but incorrect information, which can be particularly dangerous in risk-sensitive domains such as agricultural trade, where a fabricated phytosanitary rule or a misread freight signal can trigger costly operational errors.

AI risk model calibration and decision empowerment capacity encompasses the organizational ability to calibrate and update GenAI risk models in response to changing trade environments and to translate risk assessments into actionable operational decisions (Mikalef and Gupta, 2021). This dimension bridges technical capability and organizational implementation, recognizing that even the most sophisticated AI tools are ineffective without organizational processes that enable their integration into decision-making workflows, and corresponds to the seizing and transforming microfoundations of dynamic capabilities (Teece, 2007).

2.2 Cross-border agricultural trade resilience

Supply chain resilience has emerged as a critical research priority following the cascade of global disruptions in the early 2020s (Chowdhury and Quaddus, 2017; Hohenstein et al., 2015). We define cross-border agricultural trade resilience (CATR) as the capacity of agricultural enterprises engaged in international trade to prepare for, respond to, and recover from disruptions while maintaining continuous operations and adapting to changing conditions (Ponomarov and Holcomb, 2009).

Following Chowdhury and Quaddus (2017), we operationalize CATR as a second-order construct comprising three dimensions. Resistance capacity refers to the ability to withstand disruptions without significant operational degradation, including maintaining stable delivery of core export orders and preserving capital chain continuity amid supply interruptions. Recovery capacity reflects the speed and effectiveness of returning to normal operations following disruptions, encompassing rapid restoration of import-export activities and minimization of perishable product losses—a particularly salient concern for reefer-containerized agricultural cargo held up at congested ports. Transformation and adaptation capacity captures the ability to learn from disruptions and fundamentally reconfigure operational approaches, including adjusting global market layouts, developing new products and markets, and diversifying overseas partnerships and shipping routes.

2.3 Theoretical framework and hypotheses development

The resource-based view (RBV) (Barney, 1991) posits that sustainable competitive advantage derives from valuable, rare, inimitable, and non-substitutable resources and capabilities. In our framework, RBV plays a descriptive role: it characterizes what kind of asset GenAI-RPC plausibly is—a capability bundle whose value derives from specialized technical infrastructure, domain-specific training data, and firmspecific organizational learning that competitors cannot readily purchase or copy (Grant, 1996). RBV alone, however, is largely static: it does not explain how such a capability translates into resilience outcomes in turbulent environments.

Dynamic capability theory (Teece et al., 1997; Teece, 2007) supplies this process account. It emphasizes the capacity to integrate, build, and reconfigure internal and external competencies in rapidly changing environments, and Teece’s sensing–seizing–transforming framework maps naturally onto the three GenAIRPC dimensions: multi-source risk identification enables enterprises to sense emerging maritime and trade risks through multi-source data monitoring; prediction accuracy and anti-hallucination stability ensure that what is sensed is reliable enough to act upon; and model calibration and decision empowerment allow firms to seize opportunities and transform operations, converting risk intelligence into reconfigured activities. RBV and dynamic capability theory are thus complementary rather than duplicative in our design: RBV justifies treating GenAI-RPC as a strategic capability worth investing in, whereas dynamic capability theory specifies the mechanism through which its three dimensions are expected to co-produce resilience.

The conceptualization of GenAI-RPC as a single higher-order capability with three dimensions, rather than three independent capabilities, rests on two arguments. First, the three dimensions are technically interdependent: they are co-produced by the same underlying GenAI infrastructure, data pipelines, and model assets, so that investment in one dimension (e.g., multi-source data integration) immediately conditions the others (prediction accuracy, calibration). Second, they are functionally sequential facets of one risk-prediction workflow—sense, verify, and act—and resilience benefits materialize only when all three are jointly present: a firm that can sense risks but cannot trust or act on the predictions derives little resilience value. This mirrors the multidimensional treatment of supply chain resilience in Chowdhury and Quaddus (2017), in which conceptually distinct but interdependent capacities are modelled as dimensions of one higher-order construct. We therefore treat the three GenAI-RPC dimensions as first-order reflective constructs and estimate their distinct associations with CATR, consistent with established practice for multidimensional capability constructs (Sarstedt et al., 2019).

Based on this theoretical foundation, we hypothesize (as shown in Figure 1):

• H1a: Multi-source cross-border risk identification and scoring capacity is positively associated with cross-border agricultural trade resilience. This dimension may strengthen CATR by expanding the scope of risk awareness beyond conventional monitoring systems, enabling enterprises to anticipate diverse disruption scenarios, including port-level and shipping-lane disruptions.
• H1b: Prediction accuracy and anti-hallucination stability capacity is positively associated with crossborder agricultural trade resilience. Reliable predictions reduce false alarms and missed threats, supporting more confident decision-making under uncertainty.
• H1c: AI risk model calibration and decision empowerment capacity is positively associated with cross-border agricultural trade resilience. This dimension bridges the gap between analytical output and organizational action, helping risk intelligence translate into operational adaptation.

3 Methodology

3.1 Research design

This study employed a cross-sectional survey design with PLS-SEM as the primary analytical approach. PLS-SEM was selected for four reasons. First, the research objective is oriented toward prediction and explanation of a target construct (CATR) in a relatively under-researched context, a situation for which PLS-SEM is recommended over covariance-based SEM (Hair et al., 2019; Shmueli et al., 2019). Second, the model contains a higher-order component—CATR conceptualized as a second-order construct—which PLS-SEM accommodates through the repeated indicator approach (Sarstedt et al., 2019). Third, the approach is robust to moderate departures from multivariate normality, which is common with Likert-scale survey data (Hair et al., 2019; Sarstedt et al., 2019). Fourth, the study extends existing theory (RBV and dynamic capability theory) to a novel construct, an exploratory purpose aligned with the strengths of PLS-SEM. All constructs in the model are specified as reflective: each GenAI-RPC dimension is measured by perceptual items expected to covary with the underlying capability, and CATR is a reflective–reflective higher-order construct in which the three first-order dimensions (resistance, recovery, transformation) reflect the second-order construct (Sarstedt et al., 2019; Chowdhury and Quaddus, 2017). No formative constructs are included; accordingly, reflective evaluation criteria (loadings, internal consistency, convergent and discriminant validity) are applied throughout.

3.2 Instrument development

The measurement instrument consisted of 25 Likert-scale items measuring the three GenAI-RPC dimensions (13 items) and three CATR dimensions (12 items), plus five control variable items. All construct items used a 5-point Likert scale (1 = Strongly Disagree, 5 = Strongly Agree). Scale development followed established procedures: (i) an initial item pool was generated from the academic literature on AI-enabled supply chain risk management, GenAI output reliability, and supply chain resilience (Chowdhury and Quaddus, 2017; Boone et al., 2025; Ji et al., 2023; Mikalef and Gupta, 2021; Toorajipour et al., 2021); (ii) a content validity panel of two academics in supply chain management and three senior managers from cross-border agricultural enterprises reviewed all items for relevance, clarity, and domain coverage, and ambiguous or redundant items were reworded or removed; (iii) the English instrument was translated into Chinese and back-translated by two independent bilingual researchers, with discrepancies reconciled; and (iv) a pilot test with 32 senior managers from the target population was conducted, and item purification based on corrected item-to-total correlations (all above 0.50 in both the pilot and the main sample) supported the retention of all 25 items. Respondents answered with reference to the GenAI systems actually deployed in their firms, so the items capture perceived organizational capability rather than technical model performance.

The GenAI-RPC items operationalize the three capability dimensions defined in Section 2.1. Items measuring multi-source risk identification (Q1–Q5) capture data aggregation, hidden risk mining (including port disruption and trade friction signals), risk scoring, scenario simulation, and early warning report generation. Items measuring prediction accuracy (Q6–Q9) assess perceived hallucination levels, stability under incomplete data, intelligent data filling, and result traceability. Items measuring model calibration (Q10–Q13) evaluate model update capability, decision support for order adjustment, countermeasure provision, and information asymmetry reduction.

The CATR items were adapted from Chowdhury and Quaddus (2017), with domain-specific modifications for the agricultural trade context. Resistance capacity items (Q14–Q17) assess core order delivery stability, loss avoidance through early warning, supply network resistance to policy barriers, and capital chain continuity. Recovery capacity items (Q18–Q21) measure restoration speed after disruption, business indicator rebound, supplier replacement capability, and perishable product loss reduction. Transformation capacity items (Q22–Q25) capture market layout adjustment, product and market diversification, operational mode optimization, and partner diversification.

3.3 Data collection

Data were collected between October and December 2025 through a multi-channel survey approach. Questionnaires were distributed via agricultural industry associations, cross-border trade chambers of commerce, and professional online survey platforms targeting enterprises in Shandong, Guangdong, Fujian, and Zhejiang provinces. These coastal provinces were selected due to their prominence in China’s crossborder agricultural trade, collectively accounting for approximately 55% of national agricultural export volume, and because their cross-border agri-trade is predominantly seaborne, making them a theoretically relevant setting for examining maritime trade risk. Sampling was purposive: a screening item restricted participation to enterprises that had deployed at least one GenAI-based tool (e.g., large-language-modelbased risk analysis or intelligent customs documentation) in cross-border operations for no less than six months, and respondents were required to be senior managers or above with direct oversight of cross-border operations and familiarity with the firm’s GenAI deployment. This screening ensured that respondents were qualified to assess the capability items; we nevertheless acknowledge that their assessments remain informed perceptions rather than technical benchmarks (see Section 5.5).

Of 1,680 questionnaires distributed, 312 were returned, yielding an 18.6% response rate. Twelve responses were excluded due to excessive missing data (more than 15% incomplete) or straight-line patterns, resulting in 300 valid responses. To assess potential non-response bias, we conducted a wave analysis comparing early (first quartile) and late (last quartile) respondents (Armstrong and Overton, 1977): independentsamples t-tests revealed no significant differences on any study construct or control variable (all p > 0.12), suggesting that non-response bias is unlikely to be a serious concern, although it cannot be ruled out entirely. Sample size adequacy was examined a priori through a formal power analysis rather than heuristic rules of thumb: assuming a small-to-medium effect (f² = 0.05), α = 0.05, and statistical power of 0.80, the minimum required sample is 222 for the hypothesized three-predictor model and 309 for the full model including five controls (Cohen, 1988; Faul et al., 2007). Our sample of 300 exceeds the former threshold (achieved power = 0.91) and falls only marginally short of the latter, with achieved power of 0.79 for the full model. We note, however, that the study would be underpowered for very small effects (f² = 0.02; required N = 550–759), which we acknowledge as a limitation.

3.4 Sample characteristics

Table 1 presents the demographic profile of the 300 participating enterprises. Respondents were senior managers or above with direct involvement in cross-border agricultural trade operations. The sample exhibited reasonable diversity across operation years, turnover levels, product categories, partner country counts, and digital transformation maturity.

Table 1. Sample Demographics (N = 300).

3.5 Data analysis procedure

Data analysis was conducted using SmartPLS 4.0 following the two-step approach recommended by Hair et al. (2019), with CATR modelled as a reflective–reflective higher-order construct via the repeated indicator approach (Sarstedt et al., 2019). First, the measurement model was evaluated to establish reliability and validity (indicator loadings, Cronbach’s α, composite reliability, AVE, the Fornell–Larcker criterion, and HTMT ratios). Second, the structural model was assessed to test the hypothesized relationships. Bootstrapping with 5,000 subsamples was employed to estimate standard errors and test statistical significance. Model fit was assessed using the standardized root mean square residual (SRMR), with values below 0.08 indicating acceptable fit (Hu and Bentler, 1999). Effect sizes (f²) were computed as (Rincluded2−Rexcluded2)/(1−Rincluded2), and inner variance inflation factors (VIFs) were examined for collinearity. Predictive relevance was assessed using the Stone–Geisser Q² obtained through blindfolding (omission distance = 7) and through PLSpredict (ten folds, ten repetitions) following Shmueli et al. (2019). Common method bias was examined statistically through Harman’s single-factor test and a full collinearity VIF assessment (Kock, 2015; Podsakoff et al., 2003).

3.6 Ethical statement

This study was approved by the Ethics Committee of Zhejiang Yuexiu University (Approval No. 2025AGR-047), the institution of the first author, which coordinated the multi-province data collection. All participating senior managers signed written informed consent forms prior to completing the survey. The research was conducted anonymously, with no collection of personal identifiers or commercially sensitive information. Data were used solely for academic research purposes. Participants were informed that they could withdraw from the study at any time without providing reasons or facing any consequences. No monetary compensation was provided; however, participants were offered a summary of the research findings upon request. The study was conducted in accordance with the Declaration of Helsinki and the Personal Information Protection Law of the People’s Republic of China.

4 Results

4.1 Descriptive statistics

Table 2 presents the means, standard deviations, and bivariate correlations among the study constructs. All GenAI-RPC and CATR constructs exhibited moderate positive correlations, with the strongest correlation among the three GenAI-RPC dimensions observed between prediction accuracy (D2) and model calibration (D3) (r = 0.529). None of the correlations approached levels that would indicate multicollinearity concerns (all below 0.70).

Table 2. Descriptive statistics and correlation matrix.

N, 300. ^∗∗∗p< 0.001.

4.2 Measurement model assessment

4.2.1 Reliability analysis

Table 3 reports the reliability and convergent validity results. Cronbach’s α values ranged from 0.817 to 0.855, all exceeding the recommended threshold of 0.70 (Nunnally and Bernstein, 1994). Composite reliability (CR) values ranged from 0.880 to 0.902, all surpassing the 0.70 cutoff (Hair et al., 2019). Average variance extracted (AVE) values ranged from 0.623 to 0.697, all exceeding the 0.50 criterion (Fornell and Larcker, 1981), and all indicator loadings exceeded 0.75. These results confirm satisfactory internal consistency and convergent validity for all constructs.

Table 3. Reliability and convergent validity results.

CR, composite reliability; AVE, average variance extracted.

4.2.2 Discriminant validity

Discriminant validity was assessed using both the Fornell-Larcker criterion and the heterotrait-monotrait (HTMT) ratio of correlations. As shown in Table 4, the square root of the AVE for each construct exceeded its correlations with all other constructs, satisfying the Fornell-Larcker criterion (Fornell and Larcker, 1981).

Table 4. Discriminant validity—Fornell-Larcker criterion.

Diagonal elements are AVE; off-diagonal elements are correlations.

Table 5 presents the HTMT values. All HTMT ratios were below the conservative threshold of 0.85 (Henseler et al., 2015), with the highest value being 0.726 for the Recovery–Transformation pair. Of particular relevance to the distinctiveness of the three GenAI-RPC dimensions, the HTMT ratios among D1, D2, and D3 ranged from 0.564 to 0.641, indicating that the three capabilities are empirically separable rather than redundant measures of a single undifferentiated factor. These results confirm adequate discriminant validity across all constructs.

Table 5. Discriminant validity—HTMT ratios.

4.3 Common method bias

Because all focal variables were rated by the same respondent in a single questionnaire, we implemented both procedural and statistical remedies for common method bias (CMB) (Podsakoff et al., 2003). Procedurally, respondents were assured of anonymity and confidentiality; items for the predictor and outcome constructs were placed in separate, differently labelled sections of the questionnaire; scale anchors were kept concrete and behaviorally anchored; and no information linking responses to firms was collected. Statistically, Harman’s single-factor test showed that the first unrotated factor accounted for 36.82% of the total variance, below the 50% threshold. In addition, a full collinearity assessment (Kock, 2015) yielded VIF values between 1.595 and 1.824 for all constructs, well below the 3.3 criterion. These results jointly suggest that CMB is unlikely to account for the observed relationships, although we acknowledge that same-source perceptual measures cannot fully exclude a positive-impression halo (see Section 5.5).

4.4 Structural model and hypothesis testing

The structural model was assessed using path coefficients, significance levels, the coefficient of determination (R²), effect sizes (f²), Stone–Geisser Q² values, and model fit indices. Figure 2 presents the PLS-SEM path analysis results, and Table 6 summarizes the hypothesis testing outcomes. Inner VIF values for the three predictors ranged from 1.473 to 1.585, indicating no collinearity concerns.

Table 6. Structural model results and hypothesis testing.

D1, Multi-source Risk Identification; D2, Prediction Accuracy; D3, Model Calibration; CATR, Cross-border Agricultural Trade Resilience. CI, bootstrap confidence interval (5,000 subsamples). R², 0.461; adjusted R², 0.456; SRMR, 0.069; Q², 0.112. All hypotheses supported at p< 0.001.

All three hypotheses received empirical support. Multi-source cross-border risk identification and scoring capacity (H1a) showed the largest positive association with CATR (β = 0.291, p< 0.001), with a small-tomoderate effect size (f² = 0.103). AI risk model calibration and decision empowerment capacity (H1c) showed a nearly identical association (β = 0.287, p< 0.001, f² = 0.102). Prediction accuracy and antihallucination stability capacity (H1b) was likewise positively associated with CATR (β = 0.251, p< 0.001), with the smallest of the three effect sizes (f² = 0.074). Under Cohen (1988) conventions (0.02 small, 0.15 medium, 0.35 large), all three effect sizes fall within the small range, and we therefore refrain from characterizing them as medium effects.

The model explained 46.1% of the variance in CATR (R² = 0.461; adjusted R² = 0.456), indicating moderate explanatory power (Hair et al., 2019). The SRMR value of 0.069 fell below the 0.08 threshold, indicating acceptable model fit (Hu and Bentler, 1999).

Predictive relevance was assessed using two complementary procedures. First, the blindfolding procedure (omission distance = 7) produced Stone–Geisser Q² values above zero for all twelve CATR indicators (range 0.035–0.184; overall Q² = 0.112), establishing that the model has predictive relevance for the endogenous construct. Second, PLSpredict (ten folds, ten repetitions) showed that all twelve indicators achieved positive Q^2predict values (range 0.057–0.190), indicating that the PLS-SEM predictions outperform the naive mean benchmark for every indicator. At the same time, the linear-model benchmark produced marginally lower RMSEs for eleven of the twelve indicators (all differences below 0.035), which, under the guidelines of Shmueli et al. (2019), suggests limited out-of-sample predictive power. We therefore characterize the model as explanatory with modest predictive capability and refrain from stronger predictive claims.

4.5 Control variable analysis

Table 7 presents the results of the full structural model including the control variables, with all paths reported simultaneously so that the robustness of the hypothesized relationships can be evaluated directly. Enterprise digital transformation maturity (C5) showed a strong positive association with CATR (β = 0.577, p< 0.001), and the explained variance rose to R² = 0.701. Importantly, all three GenAI-RPC dimensions remained positive and statistically significant after controlling for digital maturity and the other controls (β = 0.182, 0.136, and 0.143 for D1, D2, and D3, respectively; all p< 0.001), indicating that the hypothesized relationships are not an artifact of general digital maturity. The remaining control variables—operation years (C1), annual turnover (C2), main product (C3), and partner countries (C4)—were all non-significant (all p > 0.70), suggesting that the GenAI-RPC–CATR association is robust across different enterprise sizes, operational tenures, product categories, and partnership scopes.

Table 7. Full structural model including control variables (R² = 0.701).

N, 300. Bootstrap with 5,000 subsamples. All inner VIFs ≤ 1.647.

5 Discussion

5.1 Theoretical implications

This study makes several contributions to theory. First, it extends the resource-based view and dynamic capability theory by operationalizing GenAI-RPC as a multidimensional organizational capability in the context of seaborne agricultural trade. The three dimensions—multi-source risk identification, prediction accuracy, and model calibration—represent distinct but complementary facets of a dynamic capability that enables agricultural enterprises to sense, verify, and act on cross-border trade risks. The finding that each dimension is independently associated with CATR is consistent with the RBV-informed expectation that GenAI-RPC can function as a strategically valuable capability when embedded within organizational processes (Barney, 1991); at the same time, we emphasize that cross-sectional survey evidence cannot establish the rarity, inimitability, or non-substitutability of this capability, so our claims are limited to the capability–resilience association rather than to VRIO status.

Second, our findings contribute to the emerging literature on artificial intelligence applications in supply chain management (Toorajipour et al., 2021; Cromwell et al., 2025). While previous studies have examined AI broadly, our focus on GenAI-RPC as a specific construct allows for a more nuanced understanding of the mechanisms through which advanced AI capabilities relate to resilience outcomes. The pattern of path coefficients suggests that not all GenAI capabilities are equally strongly associated with resilience: risk identification (H1a) and model calibration (H1c) showed slightly larger coefficients than prediction accuracy (H1b), although the differences are small and the bootstrap confidence intervals overlap.

Third, this study provides a psychometrically evaluated measurement instrument for GenAI-RPC that can be adapted and extended in future research. The instrument was developed from the academic literature, refined through expert review and pilot testing, and demonstrated satisfactory reliability and validity, offering a foundation for cross-industry comparative studies.

5.2 Managerial implications

The findings offer actionable guidance for agricultural enterprises engaged in cross-border trade, while also calling for realistic expectations. First, enterprises should prioritize investments in multi-source data integration capabilities: the largest path coefficient (H1a: β = 0.291) suggests that the breadth and depth of data sources feeding into GenAI risk models—including port, shipping, and freight data—is the dimension most strongly associated with trade resilience. Enterprises should establish data pipelines that integrate policy databases, climate models, port and vessel tracking systems, and market intelligence platforms. We caution, however, that the observed effect sizes are small, so investments should be staged and periodically evaluated rather than assumed to produce automatic resilience gains.

Second, while prediction accuracy is important, our results suggest that organizations should not view it in isolation. The anti-hallucination stability of GenAI outputs requires robust validation mechanisms, including human-in-the-loop verification processes and traceability features that allow decision-makers to audit the reasoning behind risk predictions. Organizations should implement feedback loops that continuously assess prediction accuracy against actual outcomes and should retain qualified personnel capable of overriding AI-generated recommendations.

Third, the significant association of model calibration capacity (H1c) underscores the importance of organizational processes that bridge AI analytics and operational decision-making. Enterprises should develop dedicated teams responsible for translating risk intelligence into procurement adjustments, supplier diversification strategies, and inventory management protocols. The strong association of digital transformation maturity (C5) with CATR further suggests that GenAI investments are likely to yield the greatest benefit when accompanied by broader digital infrastructure development, employee digital skills, and data governance; GenAI capability appears to complement, rather than substitute for, general organizational digital capability. The GenAI-specific risks that motivate these complementary safeguards are elaborated in Section 5.4.

5.3 Policy implications

From a policy perspective, our findings support the development of digital infrastructure and AI governance frameworks specific to agricultural trade, with particular relevance for maritime and port policy. Port authorities and customs administrations in major agricultural trading nations should consider establishing public-private partnerships to create shared data platforms—linking port community systems, customs single windows, and agricultural trade data—that reduce information asymmetries in cross-border agricultural markets and improve the quality of multi-source risk identification. The strong association of digital maturity with CATR suggests that policies aimed at closing the digital divide among agricultural enterprises, including small and medium-sized exporters, could yield resilience dividends for the wider maritime food supply chain.

International organizations such as the FAO, the WTO, and the IMO could contribute by developing standardized data protocols for agricultural trade and maritime risk information, enabling more effective multi-source risk identification by GenAI systems. Additionally, regulatory frameworks should address the unique challenges of AI hallucination in risk-sensitive domains, establishing guidelines for validation and accountability in AI-generated risk assessments used in customs clearance, phytosanitary compliance, and cold-chain logistics.

5.4 GenAI-specific risks and safeguards

Our results highlight the potential benefits of GenAI-RPC, but the deployment of GenAI in crossborder agricultural trade also entails distinctive risks that merit explicit attention. First, hallucination—the generation of plausible but fabricated content—remains an intrinsic limitation of large language models (Ji et al., 2023); in our context, a fabricated tariff rule, phytosanitary requirement, or port restriction could trigger costly shipment errors, which is why the prediction accuracy dimension explicitly incorporates traceability and verification items. Second, GenAI outputs are only as reliable as the underlying data: incomplete, biased, or outdated training data can produce systematically distorted risk assessments, and algorithmic bias may disadvantage particular trading routes or partners. Third, connecting enterprise systems to external GenAI platforms raises cybersecurity and data privacy concerns, particularly where commercially sensitive shipment data are involved. Fourth, over-reliance on AI-generated recommendations may erode human expertise and create automation complacency, and model drift can silently degrade prediction quality as trade environments change. We therefore recommend a governance model in which GenAI augments rather than replaces managerial judgment: human-in-the-loop verification of high-stakes predictions, periodic model audits and recalibration, data quality management, and clear accountability for AI-assisted decisions. These safeguards are consistent with the balanced managerial implications discussed above.

5.5 Limitations and future research

Several limitations should be acknowledged. First, the cross-sectional design precludes causal inference. Although we have used associational language throughout, the observed relationships are consistent with alternative causal orderings—for example, more resilient firms may find it easier to invest in and deploy GenAI—and longitudinal or experimental designs would strengthen causal confidence. Future research could employ panel data to track how GenAI-RPC investments relate to resilience outcomes over time.

Second, our sample was drawn from four coastal Chinese provinces, which limits generalizability to inland Chinese enterprises and to agricultural firms in other countries with different port infrastructure, trade institutions, and digital ecosystems; replication studies in other countries and agricultural systems would enhance external validity. Third, all measures were derived from self-reported, single-respondent survey data collected at one point in time, which may introduce common method variance and a positive-impression halo; although our procedural remedies, Harman’s single-factor test (36.82%), and the full collinearity assessment (all VIFs ≤ 1.824) suggest that CMB is unlikely to fully account for the results, future studies should incorporate objective resilience measures (e.g., actual disruption recovery times) and multiinformant designs. Relatedly, the GenAI-RPC items capture managerial perceptions of deployed GenAI capabilities rather than technical benchmarks of model accuracy or hallucination rates, and managers may over- or under-estimate technical performance; future work should triangulate perceptual measures with system logs or technical audits. Fourth, the 18.6% response rate, although accompanied by a non-significant wave analysis, leaves room for non-response bias.

Fifth, our model examined direct effects only; future research should investigate mediating mechanisms (e.g., decision-making speed, supply chain visibility) and moderating conditions (e.g., environmental dynamism, regulatory complexity), and should formally model digital transformation maturity as a potential antecedent or confounder of the GenAI-RPC–CATR relationship, given its strong association with CATR in our data. In addition, establishing the rarity, inimitability, and non-substitutability of GenAI-RPC would require separate evidence, such as adoption-base comparisons or longitudinal performance data, which we flag as a direction for future research. Sixth, the study is underpowered for very small effects (f² = 0.02), and PLSpredict indicated limited out-of-sample predictive power; larger samples and richer model specifications are needed before the model can support firm-level prediction. Finally, as GenAI technology evolves rapidly, the relative importance of the three capability dimensions may shift, warranting periodic reassessment.

6 Conclusion

This study provides empirical evidence that generative AI risk prediction capability is positively associated with cross-border agricultural trade resilience among coastal Chinese enterprises whose trade is predominantly seaborne. Through PLS-SEM analysis of survey data from 300 agricultural enterprises, we found that three GenAI-RPC dimensions—multi-source risk identification, prediction accuracy, and model calibration—were each independently and positively associated with CATR, jointly explaining 46.1% of its variance, and that these associations remained significant after controlling for enterprise digital transformation maturity. These findings extend the resource-based view and dynamic capability theory by operationalizing GenAI as a multidimensional organizational capability, and they offer guidance for enterprises and policymakers seeking to strengthen the resilience of maritime food supply chains. As cross-border agricultural trade continues to face mounting uncertainty, GenAI-RPC represents a promising—though neither sufficient nor risk-free—dynamic capability that can help enterprises adapt amid complexity.

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: Narasimha Rao Vajjhala, University of New York Tirana, Albania

Reviewed by: Sajid Rahman Khattak, University of Agriculture, Peshawar, Pakistan

Sidra Nazir, Ural Federal University, Russia