Latest / 1 More Paper / From Generative AI Adoption to Sustainable Performance Operations Capability Mediation and Environmental Dynamism — He & Yang (2025)
Transcript
- You’re listening to “From Generative AI Adoption to Sustainable Performance Operations Capability Mediation and Environmental Dynamism,” by M. He and S. Yang. Published in 2025. Abstract. Enterprises are rapidly scaling generative artificial intelligence, yet evidence is mixed on whether adoption translates into measurable sustainability gains and under what conditions such gains arise. This study theorizes and tests a capability-based pathway in which adoption improves operations capability, which in turn enhances sustainable performance, with the payoff contingent on environmental dynamism. We assemble a multi-year panel of publicly listed manufacturing firms in China and construct firm-year measures for generative artificial intelligence adoption, operations capability, sustainable performance, and environmental dynamism from public disclosures and administrative sources. Empirically, we estimate two-way fixed-effects models at the firm level, evaluate a moderated-mediation structure with clustered bootstrap inference, and conduct extensive robustness checks and heterogeneity analyses. The results show that adoption is positively associated with higher operations capability—reflected in reliability, responsiveness, and process discipline—and that this capability mediates the link to sustainable performance. The indirect effect is stronger when environmental dynamism is high, indicating that turbulent product and policy environments amplify the value of faster information flows and auditable processes. Findings remain stable across alternative measurements, lag structures, and sample partitions. The study clarifies when and how generative artificial intelligence creates sustainability value: not through technology presence alone, but through the operational capability it enables under dynamic conditions. Introduction. Generative artificial intelligence (GAI) is moving from pilots to enterprise-scale deployment across procurement, production, maintenance, logistics, and customer operations. Beyond content generation, firms increasingly leverage GAI for cognitive automation—summarising unstructured docu-ments, assisting anomaly detection, and supporting planning decisions in near real time. Yet managers still ask a basic question: how does GAI adoption translate into measurable The associate editor coordinating the review of this manuscript and approving it for publication was Yizhang Jiang. firm outcomes, and under what conditions do digital investments deliver superior returns? In manufacturing, where reliability, throughput, and compliance jointly determine performance, technology value hinges not only on adopting tools but also on embedding them into data governance and cross-functional routines. Two gaps motivate our study. First, adoption is not capability. Many digital transformation studies effectively treat technology as a binary presence or as spending levels, assum-ing a linear passage from investment to outcomes. The IT value literature, however, emphasises that complementary assets—process redesign, coordination, and skills—mediate returns. Distinguishing GAI adoption from AI capability is thus essential to understand whether firms merely install tools or actually convert them into replica-ble, organisation-level capabilities. Second, context matters. In turbulent product and factor markets, environmental dynamism (e.g., demand volatility, regulatory change, supply disruptions) can raise the option value of information-rich, reconfigurable processes. If so, the payoff from GAI should be stronger where uncertainty rewards fast learning and reconfiguration, rather than uniform across environments. We respond to these gaps by theorising and testing a capability-based pathway from enterprise GAI adoption to sustainable performance in manufacturing firms. We conceptualise sustainable performance (SUSPERF) as multi-dimensional, integrating environmental outcomes (e.g., energy/material efficiency, emissions intensity) with operational and economic performance relevant to competitiveness. Our central premise is that operations capability (OPSCAP)—the bundle of routines that ensure reliable, responsive, and efficient operations—acts as the transmission mechanism transforming digital inputs into firm-level outcomes. We further posit that environmental dynamism (ENVDYN) amplifies this conversion because volatile contexts increase the marginal value of cognitive automation and fast information flow. Empirically, we assemble a multi-year panel of publicly listed Chinese manufacturing firms and develop two operational measures: a GAI adoption index (GAIINDEX) capturing the presence and breadth of GAI-related initiatives, and a firm-level AI capability measure (GAIC) reflect-ing the embedding of GAI into workflows, data assets, and coordination routines. We then test three questions: (i) whether GAIINDEX and GAIC are associated with higher OPSCAP; (ii) whether OPSCAP mediates the relation between GAI and SUSPERF; and (iii) whether ENVDYN strengthens this indirect (mediated) effect. Identification relies on firm- and year-fixed-effects models to absorb time-invariant firm characteristics and common shocks, with extensive robustness and heterogeneity checks. This design contributes to the IS–operations interface in three ways. First, we separate adoption from capability by constructing practical, firm-level measures (GAIINDEX and GAIC) and validating them on a large panel rather than inferring capability from presence alone. Second, we identify the mechanism linking IS to OM outcomes by positioning OPSCAP as the mediator between GAI and sustainability and by showing that the indirect effect rises with ENVDYN—clarifying how digital technologies become operational advantages and when they matter most. Third, we provide actionable guidance: to convert GAI into measurable sustainability gains, managers should couple technology deployment with process/data governance and time capability building to volatility, treating GAI not as a stand-alone tool but as a catalyst for cognitive, data-rich operations. Our focus on sustainability is deliberate for three rea-sons. First, manufacturers face tightening constraints from energy costs, disclosure, and circularity. Process improvements that save time or reduce defects often also reduce waste and emissions. Second, stakeholders increasingly expect verifiable non-financial outcomes, making sustainability an auditable objective for digital programmes. Third, evalu-ating outcomes beyond narrow financials reduces the risk of overstating transient cost cuts while overlooking durability, resilience, and compliance—capabilities central to long-run competitiveness. Research questions. RQ1: Do enterprise GAI adoption (GAIINDEX) and the associated AI capability (GAIC) improve operations capability (OPSCAP) in manufacturing firms? RQ2: Does OPSCAP mediate the effect of GAI on sustainable performance (SUSPERF)? RQ3: Is the indirect (GAI → OPSCAP → SUSPERF) effect stronger under environmental dynamism (ENVDYN)? Preview of approach. We construct firm-level GAIINDEX and GAIC from public disclosures and verifiable sources, operationalise OPSCAP with indicators of process reliability, responsiveness, and coordination, and measure SUSPERF via validated proxies aligned with environmental and operational outcomes. Baseline specifications employ firm- and year-fixed effects with clustered robust standard errors; mediation is evaluated via the product-of-coefficients logic in a panel setting, and moderation is assessed by inter-acting ENVDYN with the mediator pathway. Robustness includes alternative operationalisations (e.g., lag structures, proxy sets), outlier diagnostics, and subsample analyses by baseline automation, region, size, and ownership. Roadmap. Section II develops the theoretical framework and hypotheses. Section III describes data, measures (GAIINDEX, GAIC, OPSCAP, SUSPERF, ENVDYN), and empirical strategy. Section IV presents baseline, mediation, and moderation results, followed by robustness and heterogeneity analyses. Section V discusses theoretical and managerial implications. Section VI concludes with limitations and avenues for future research. Contributions. This study makes three contributions: We conceptually distinguish generative AI adoption (GAIINDEX) from embedded generative AI capability (GAIC), clarifying why capability—not mere adoption— drives value creation. We theorize and empirically test operations capability (OPSCAP) as the mechanism linking generative AI to sustainable performance (SUSPERF). We establish a second-stage moderated mediation, showing that environmental dynamism (ENVDYN) amplifies the OPSCAP → SUSPERF payoff. II. THEORETICAL BACKGROUND AND HYPOTHESES A. FROM ADOPTION TO CAPABILITY IN DIGITAL TRANSFORMATION Much of the digital-transformation literature treats technology as an input—presence/absence or spending levels— implicitly assuming a linear passage from adoption to outcomes. However, IS value and dynamic-capabilities research emphasises that firms capture value when digital tools become repeatable organisational abilities embedded in data governance, workflows, and coordination routines. Following this view, we distinguish enterprise-level GAI adoption (GAIINDEX) from AI capability (GAIC): the for-mer reflects the presence and breadth of GAI initiatives; the latter indicates the extent to which GAI has been routinised and scaled across processes. In manufacturing set-tings, moving from adoption to capability typically requires cross-system data integration, role redesign (human-in-the-loop validation), and consistent risk/governance practices— the ingredients that convert projects into capabilities. B. OPERATIONS CAPABILITY AS THE TRANSMISSION MECHANISM Operations capability (OPSCAP) denotes the routines that enable reliable, responsive, and efficient processes—e.g., schedule adherence, first-pass yield, changeover agility, supplier coordination, and exception handling. IS–OM studies suggest that digital investments improve performance through process capabilities rather than despite them. With GAI, the mechanism is particularly salient: cognitive automation can summarise engineering change orders, flag anomalies, surface supplier risks, and draft proce-dures, but realised impact depends on how outputs are absorbed by routines—who acts, how quickly, through which workflow, with what escalation. We therefore expect GAIINDEX and GAIC to strengthen OPSCAP via (i) richer and more accessible information (retrieval plus synthesis), (ii) lower decision latency (faster triage and routing), and (iii) denser coordination (standardised prompts/templates embedded in cross-functional workflows). Stronger OPSCAP should, in turn, improve sustainable performance (SUSPERF) because fewer defects, less rework, and smoother changeovers co-move with better resource/energy efficiency and lower emissions intensity. C. SUSTAINABLE PERFORMANCE AS AMULTI-DIMENSIONAL OUTCOME We conceptualise SUSPERF as a multi-dimensional outcome combining environmental metrics (resource/energy efficiency, emissions intensity) with operational and economic dimensions aligned with manufacturing competitiveness. Evidence indicates that operational excellence often co-delivers environmental benefits. Stakehold-ers increasingly demand verifiable non-financial outcomes, making sustainability a salient, auditable target for digital programmes. Framing outcomes as SUSPERF reduces over-attribution of short-run cost fluctuations to technology and highlights resilience and compliance—capabilities central to long-run advantage. D. ENVIRONMENTAL DYNAMISM AS A BOUNDARY CONDITION Environmental dynamism (ENVDYN)—demand volatility, regulatory churn, supplier instability—compresses decision windows and raises the cost of delay. Under higher ENVDYN, the option value of rapid information processing, codified response playbooks, and reconfigurable routines is larger. GAI-enabled cognitive automation can help firms process weak signals faster, shorten escalation paths, and recon-figure procedures at lower marginal cost; thus the marginal return to OPSCAP should be higher when ENVDYN is high. Conversely, in stable environments the same capability investments may exhibit diminishing returns. E. CONCEPTUAL POSITIONING AND OVERVIEW We propose a capability transmission chain from GAIINDEX/GAIC → OPSCAP → SUSPERF, with ENVDYN moderating the OPSCAP → SUSPERF link— i.e., a second-stage moderated mediation. To make the logic transparent, Fig. 1 depicts the conceptual model, and TABLE 1 summarises constructs, operationalisations, and hypothesis mapping in a compact format suitable for readers and reviewers. F. HYPOTHESES DEVELOPMENT GAI adoption/capability → OPSCAP. As GAI is embedded across procurement, production, maintenance, and logistics, micro-routines such as retrieve-and-summarise, triage-and-route, and draft-and-validate become standardised and scal-able, strengthening coordination and execution quality. H1: GAI adoption (GAIINDEX) is positively associated with OPSCAP. H1a: AI capability (GAIC) is positively associated with OPSCAP. OPSCAP → SUSPERF. Higher OPSCAP reduces scrap, rework, changeover delays, and unplanned downtime, which translate into improved resource/energy efficiency and lower emissions intensity. H2: OPSCAP is positively associated with SUSPERF. Mediation (GAI → OPSCAP → SUSPERF). If GAI creates value primarily through routinised execution, OPSCAP should mediate the GAI–SUSPERF relation-ship. H3: OPSCAP mediates the effect of GAI (GAIINDEX/GAIC) on SUSPERF. Moderated mediation by ENVDYN. When ENVDYN is higher, the marginal value of cognitive automation and codified response playbooks is larger, magnifying the indirect effect from GAI to SUSPERF via OPSCAP. H4: The indirect (GAI → OPSCAP → SUSPERF) effect is stronger under higher ENVDYN. G. IDENTIFICATION IMPLICATIONS The model yields three empirical implications. First, ana-lysts should separate adoption from capability to avoid conflating technology presence with routinised use. Second, mediation tests should consider temporal ordering— GAI preceding OPSCAP changes and OPSCAP preceding SUSPERF—implemented via panel designs with firm and year fixed effects and, where feasible, lags. Third, moderation and moderated-mediation tests should explicitly interact ENVDYN with the mediator pathway to capture second-stage moderation. H. POSITIONING WITHIN IS–OM AND SUSTAINABILITY LITERATURES Our theorisation connects and extends three streams. IS value and dynamic capabilities argue that digital investments yield heterogeneous outcomes via capability differences. We bring GAI explicitly into this conversation and operationalise both adoption and capability at scale. Operations and manufacturing systems research links IT integration to reliability and responsiveness; we extend this to cognitive automation, specifying OPSCAP as the mechanistic bridge to performance. Sustainability and green operations show that operational excellence co-delivers environmental outcomes; we integrate this by modelling SUSPERF as the outcome and ENVDYN as the boundary condition that shapes realised gains. Beyond these streams, very recent studies document how generative-AI usage and AI capability improve sustainable supply-chain and environmental performance, and how environmental dynamism conditions the value of digital transformation and innovation in sustainability contexts,,. Our study complements this work by using a multi-year panel of listed manufacturing firms, an operations-capability mediation design, and an ENVDYN-contingent transmission mechanism. Summary. Technology presence is not capability. Firms must embed GAI into routines to build OPSCAP, and environmental dynamism determines the payoff’s magnitude. This yields four testabl. III. MATH A. DATA, SAMPLE, AND SCOPE We build a multi-year panel of publicly listed Chinese manufacturing firms over 2018–2024. Starting from all A-share firms classified in two-digit manufacturing industries on the Shanghai and Shenzhen stock exchanges, we drop financial and utility firms, firms with missing key variables, firms under ST (special treatment) status, and extreme outliers in the main variables (see §3.6 for filters). Our unit of analysis is the firm–year. After these steps, the working sample comprises 420 firms and 3,024 firm–year observations in the baseline specifications; due to lagging and occasional miss-ing regressors, the estimation samples in Tables 6–8 range between 2,680 and 3,024 firm–years. Firm-level information is assembled from four sources: (i) narrative and quantitative disclosures in annual reports and ESG reports, which we use to code GAIINDEX and GAIC; (ii) operational key performance indicators such as through-put, defect and rework rates, on-time delivery, and inventory turns, which enter OPSCAP; (iii) financial statements and balance sheet data, which provide the control variables; and (iv) third-party market and policy data used to construct industry-level demand, price and regulatory indicators for ENVDYN. All variables are merged into a single panel at the firm–year level. Our unit of analysis is the firm-year. Following IMDS conventions, we summarise construct definitions in TABLE 1 and provide the full variable dictionary, codebook and sources in Supplementary Tables S1–S2. Table 2 reports descriptive statistics and correlations. On average, GAIINDEX and GAIC exhibit moderate dispersion, and their correlation is about 0.55, suggesting related but non-redundant constructs. OPSCAP and SUSPERF are positively correlated (ρ ≈ 0.45), consistent with our theorisation that operational capability co-delivers sustainable performance. ENVDYN is only weakly correlated with GAIINDEX/GAIC, reducing concerns about multicollinearity. B. MEASURES Tables 3–5 summarise the operationalisation of all constructs, including the exact proxy definitions, time windows and data sources used for each component of GAIINDEX, GAIC, OPSCAP, SUSPERF and ENVDYN. Detailed construct definitions, coding rubrics, and additional examples are pro-vided in Supplementary Tables S1 and S2. (a) GAI adoption (GAIINDEX). We implement a content-validated indicator of enterprise-level GAI adoption. Using firms’ annual/ESG reports and official disclosures, we code the presence and breadth of GAI initiatives (e.g., pilots, vendor integrations, production deployments, governance protocols). The index aggregates binary/band scores across adoption facets with equal weights; alternative weights and cut-offs appear in robustness (Supplementary Table S3). This measure captures presence, not depth (Brynjolfsson and McAfee ). (b) AI capability (GAIC). To separate adoption from capability, GAIC captures the extent to which GAI is embedded in workflows, data governance and coordination routines, (Teece ). We code evidence of: (i) cross-system data integration and lineage; (ii) human-in-the-loop validation standards; (iii) reusable prompt/templates and playbooks; (iv) risk/compliance artefacts. GAIC is scaled to; details and inter-coder reliability are in Supplementary Table S2. GAIC is conceptually distinct from GAIINDEX; empirical correlation and VIF diagnostics are reported in §3.6. (c) Operations capability (OPSCAP). We operationalise OPSCAP via normalised indicators of reliability (e.g., first-pass yield, defect/scrap inverses), responsiveness (e.g., order fulfilment/on-time), and coordination (e.g., inventory turns, schedule adherence). A principal-component or standardised-sum composite is used in the baseline; alterna-tives are shown in Supplementary Table S3. (d) Sustainable performance (SUSPERF). We build a composite outcome reflecting environmental (resource/energy efficiency, emissions intensity) and operational sustainability dimensions aligned with manufacturing competitiveness. Components are z-scored and averaged (directionally aligned so higher = better). (e) Environmental dynamism (ENVDYN). ENVDYN captures the extent to which firms operate in turbulent product–market and regulatory environments. Following prior work on environmental turbulence and supply volatility, we construct ENVDYN at the 2-digit industry–year level as a composite of three normalised components: demand volatility, policy flux, and supply turbulence. Formally, for industry j in year t we define: ENV DYNjt = (DEM VOLjt + POL FLUXjt + SUPTURBjt)/3, where each component is an industry–year z-score. DEM VOLjt is a rolling (three-year) standard deviation of year-on-year industry sales growth in the focal 2-digit manufacturing industry, capturing product- demand volatility. POL FLUXjt measures the intensity of relevant policy change by counting national and provincial policies related to manufacturing, environment, digitalisation, and AI that affect the industry in year t, and then normalising this count by its three-year moving average. SUPTURBjt captures supply-side turbulence as the rolling standard deviation of key input price indices and/or supplier concentration measures for the same industry, again standardised by industry–year. All three components are winsorised at the 1st/99th percentiles and standardised before aggregation so that higher values consistently indicate more turbulent external environments. The equal-weighted composite is our baseline ENVDYN measure; Table 5 provides a compact summary of these operational definitions. As robustness checks, we also re-estimate our models using single-dimension substitutes (demand-only, policy-only, supply-only) and alternative composite weights, as well as text-based uncertainty indices; the main coefficients of interest remain qualitatively unchanged (Supplementary Table S3). (f) Controls. Firm size (log assets), age, leverage, prof-itability (ROA), ownership (SOE indicator), export intensity, R&D intensity, and industry-year trends. We include firm fixed effects (μi) and year fixed effects (τ t) throughout to absorb time-invariant heterogeneity and common shocks. All continuous variables are winsorised at the 1st/99th percentiles; logarithms are used where skewed. C. CONSTRUCT OPERATIONALISATION TABLES To enhance transparency and reproducibility, we summarise in this subsection how the main constructs are operationalised in the empirical analysis. Tables 3–5 provide compact construct operationalisation tables for generative AI adoption (GAIINDEX), AI capability (GAIC), operations capability (OPSCAP), sustainable performance (SUSPERF), and environmental dynamism (ENVDYN). Each table reports the construct, underlying facets or indicators, coding rules, data sources, and illustrative disclosure snippets from our sample firms. Table 3 documents the content-validated indices for GAIINDEX and GAIC. For GAIINDEX, we code whether firms explicitly report deploying generative AI in key value-chain domains (e.g., R&D and design, production and scheduling, supply-chain planning, customer service and marketing). Each facet is coded as a binary or ordered indicator based on the presence and specificity of deployment statements in annual and ESG reports. GAIC separates adoption from capability by capturing whether firms describe dedicated data/AI talent, MLOps and engineering pipelines, governance structures, and cross-functional routines that embed AI in operations, in line with dynamic capability arguments. Table 4 summarises the three-dimensional operations capability composite (OPSCAP). We distinguish reliability, responsiveness, and coordination by mapping firm- and industry-level performance ratios (e.g., on-time delivery, order fulfilment rates, inventory turns) into standardised z-scores and then averaging them into a single index. The table clarifies how each indicator is constructed and how higher values consistently reflect stronger operations capability. Table 5 covers the sustainable performance composite (SUSPERF) and the environmental dynamism index (ENVDYN). SUSPERF aggregates several environment-related intensity measures (e.g., energy use per unit of output, emissions per unit of revenue, waste intensity), with sign conventions chosen so that higher values indicate better environmental performance. ENVDYN combines three normalised components—demand volatility, policy flux, and supply turbulence—capturing the extent to which manufacturing firms operate in turbulent environments. For each component, the table reports the specific proxy, time window, and data source. Across Tables 3–5, the last column provides representative snippets from firms’ disclosures (translated where necessary) to make the coding rules concrete. These tables condense the richer coding rubrics in the Supplement and are intended to make our operationalisation choices as transparent as possible to readers and replicators. D. EMPIRICAL STRATEGY AND MODELS We test the capability transmission chain—GAIINDEX/ GAIC → OPSCAP → SUSPERF—with firm- and year-fixed-effects panel models, then evaluate second-stage moderation by ENVDYN. Baseline (effect of GAI on OPSCAP): OPS CAPit = α1GAI INDEXit −1 + α′ 1GAICit −1 + β1Xit + μi + τt + εit Outcome (OPSCAP on SUSPERF, with GAI): SUS PERFit = γ1OPS CAPit-1 + γ2GAI INDEXit-1 + γ ′ 2GAICit-1 + β2Xit + μi + τt + uit Second-stage moderation (ENVDYN on OPSCAP → SUSPERF): SUS PERFit = δ1OPS CAPit −1 + δ2ENV DYNit + δ3OPS CAPit −1 × ENV DYNit + δ′ 4GAICit −1 + β3Xit + μi + τt + εit We lag mediators/antecedents by one year to respect temporal ordering; alternative lags appear in robustness (§3.6). Standard errors are clustered at the firm level; two-way clustering by firm and year is reported in S4. As shown in Fig. 1, our empirical design implements the capability transmission chain GAIINDEX/GAIC → OPSCAP → SUSPERF and evaluates second-stage moderation by ENVDYN. A compact mapping of constructs and hypotheses appears in TABLE 1; full definitions and data sources are listed in Supplementary Tables S1–S2. E. MEDIATION AND MODERATED MEDIATION For mediation (H3), we follow the product-of-coefficients approach: the indirect effect is ˆα1 × ˆγ1 (and ˆα′ 1 × ˆγ ′ 1 when using GAIC). We compute bootstrapped (clustered) confidence intervals and report delta-method checks in S4. For moderated mediation (H4), we estimate the index of moderated mediation based on Equation: ˆα1 × ˆδ1 + ˆδ3 · ENV DYN. We evaluate conditional indirect effects at low/mean/high ENVDYN (±1 SD) and provide Johnson–Neyman bands in the Supplement. F. IDENTIFICATION AND VALIDITY CONSIDERATIONS Our identification rests on within-firm variation, absorbing time-invariant firm attributes (culture, baseline automation) via μi and macro shocks via τ t. We address three concerns: Reverse timing.Reverse timing and temporal ordering. Our baseline models respect temporal ordering by lagging GAIINDEX/GAIC and OPSCAP by one year relative to SUSPERF, so that generative-AI adoption and capability build-up at t-1 predict operations capability at t-1 and sustainable performance at t, rather than the reverse. To probe the sensitivity of this design, we re-estimate Eqs. – using two-year lags of GAIINDEX/GAIC and OPSCAP (t-2). Coefficients remain positive and statistically significant, albeit somewhat attenuated, consistent with a gradual capability-formation process rather than instan-taneous shocks. As a falsification, we reverse the timing and regress OPSCAP{i,t-1} and SUSPERF{i,t-1} on future GAIINDEX{i,t} and GAIC{i,t}, controlling for current outcomes and covariates. The ‘‘future’’ GAI terms are small and insignificant, suggesting that our main effects are unlikely to be driven by reverse causality or pure anticipation. Lead–lag placebo tests based on future OPSCAP{i,t+1} and SUSPERF{i, t+1} likewise show no evidence of pre-trends prior to observed changes in GAIINDEX/GAIC (Supplementary Table S4). Omitted time-varying shocks. We include rich controls and industry-year trends; results are robust to adding province-year shocks (e.g., policy dummies) and firm-specific trends (Supplementary Table S4). Measurement error. GAIC and GAIINDEX are multi-facet constructs with coder guidelines; we report inter-coder reliability (κ) and show that results hold under alternative codings/weights (Supplementary Table S2–S3). While panel FE with lags cannot prove causality, our design and diagnostics align with IS–OM practice for mechanism testing under observational data. G. ROBUSTNESS, DIAGNOSTICS, AND HETEROGENEITY We report a compact set of robustness tables in the main text and move extensions to the Supplement. (a) Alternative operationalisations. GAIINDEX/GAIC: alternative weighting schemes; binary vs multi-level codings. OPSCAP: alternative composites (PCA vs equal weights; excluding one dimension at a time). SUSPERF: alternative environmental proxies (e.g., intensity vs absolute), normalisation strategies. ENVDYN: alternative volatility constructs (policy-only, demand-only; text-based uncertainty). (b) Estimation checks. Different lag structures (t−2 mediators); system GMM sensitivity (S4). Two-way clustered SEs; wild-cluster bootstrap p-values. Winsorisation at 1/99 vs 2/98; excluding crisis years. (c) Multicollinearity and leverage. VIFs < threshold; condition indices reported. Influence diagnostics (DFBETAs, Cook’s D); results robust to excluding top-leverage firms. (d) Subsamples (heterogeneity). Baseline automation (high vs low), ownership (SOE vs non-SOE), size, region. Effects are generally stronger where baseline processes are complex and environments more turbulent, consistent with H4. (e) Placebo and falsification. Replace OPSCAP with an unrelated capability proxy; effects attenuate to null. Shuffle ENVDYN across industries within year; moderated mediation disappears. (f) Temporal and placebo checks. To further assess temporal robustness, we implement a set of lead–lag specifications. First, we re-estimate the main fixed-effects models using two-year lags of GAIINDEX /GAIC and OPSCAP (t−2) and obtain qualitatively similar patterns, with somewhat smaller coefficients as expected when more distant histories are used. Second, we estimate lead specifications where future OPSCAP{i,t+1} or SUSPERF{i,t+1} are regressed on current GAIINDEX /GAIC, controlling for current outcomes and controls. These ‘‘placebo’’ coefficients are small and statistically insignifi-cant, indicating no evidence of anticipatory effects or spuri-ous pre-trends. Third, reversing the timing by letting future GAIINDEX/GAIC predict past OPSCAP or SUSPERF yields null results. Together, these temporal checks, sum-marised in Table 9 and detailed in Supplementary Table S4, reinforce the interpretation that improvements in operations capability and sustainable performance follow, rather than precede, GAI adoption and capability building. Summary. robustness and heterogeneity appear in TABLE 9 (main text) with details in Supplementary Tables S5–6. IV. UNITS. A. DESCRIPTIVES AND CORRELATIONS We begin with descriptive statistics and pairwise correlations. Means, standard deviations and selected correlations appear in Table 2. All continuous variables are standardised as described in §3.2 and winsorised at the 1st/99th percentiles. Pairwise correlations are moderate; VIFs are below conven-tional thresholds, indicating that GAIINDEX and GAIC are related but empirically distinct (see §3.6 diagnostics). Correlations between OPSCAP and SUSPERF are positive, consistent with the theorised transmission mechanism. B. TEST OF H1: GAI → OPSCAP TABLE 6 reports firm– and year–fixed-effects estimates for Equation. Across specifications that progressively add controls and industry–year trends, GAIINDEX shows a positive and statistically significant association with OPSCAP. When GAIC is entered alongside GAIINDEX, both covary positively with OPSCAP, and the model fit improves. The pattern is robust to alternative scalings of GAIINDEX/GAIC (binary vs multi-level; reweighted facets; see Table 7). In our preferred firm and year fixed-effects specification, a one-SD increase in GAIINDEX (t−1) is associated with a 0.12-SD increase in OPSCAP (β = 0.12, SE = 0.04, 95% CI [0.04, 0.20]; Table 6, col. 4). Interpretation: embedding GAI across procurement, production, maintenance and logistics (captured by GAIC) is associated with stronger routines and coordination, beyond the presence/breadth captured by adoption. These results support H1 and H1a. C. TEST OF H2: OPSCAP → SUSPERF TABLE 7 presents Equation. OPSCAP (t–1) is positively associated with SUSPERF, controlling for GAIINDEX/GAIC (t–1), firm size, age, leverage, ownership, export/R&D intensities and fixed effects. The magnitude is economically meaningful: moving from lower to higher OPS capability is associated with better resource/energy efficiency and lower emissions intensity, consistent with the ‘‘operational excellence co-delivers environmental benefits’’ view. This finding supports H2. The estimated effect is sizeable: in the two-way FE model, OPSCAP(t−1) → SUSPERF is β = 0.20 (SE = 0.05), with a 95% CI [0.10, 0.30] (Table 7, col. 4). D. TEST OF H3: MEDIATION (GAI → OPSCAP → SUSPERF) We evaluate mediation using the product-of-coefficients approach (see §3.4). TABLE 7, Panel B reports indirect effects and bootstrapped (clustered) confidence intervals. The indirect effect of GAIINDEX on SUSPERF via OPSCAP is positive and significant; the same holds for GAIC. Direct effects attenuate once OPSCAP is included, consistent with partial mediation. Delta-method checks deliver similar inferences; placebo mediation with an unrelated capability proxy collapses to null. Taken together, results support H3: GAI affects sustainability through operations capability. E. TEST OF H4: MODERATION BY ENVDYN; CONDITIONAL INDIRECT EFFECTS TABLE 8 estimates Equation by adding ENVDYN and its interaction with OPSCAP(t−1). The interaction term is positive and significant, indicating that the payoff to operations capability is larger when environmental dynamism is higher. We visualise the conditional slope of OPSCAP across ENVDYN in Fig. 2: the OPSCAP → SUSPERF gradient steepens at higher ENVDYN. Quantitatively, in the firm- and year-fixed-effects model, OPSCAP(t−1) remains strongly predictive of SUSPERF (β = 0.17, SE = 0.05, 95% CI [0.07, 0.27]) and the interaction OPSCAP(t−1) × ENVDYN is also positive (β = 0.09, SE = 0.03, 95% CI [0.03, 0.15]; Table 8, col. 4). We then compute conditional indirect effects (moderated mediation). TABLE 8, Panel B summarises the indirect (GAI → OPSCAP → SUSPERF) effect at low/mean/high ENVDYN (±1 SD), with bootstrapped CIs. The indirect effect is statistically and economically stronger at higher ENVDYN, and Johnson–Neyman bands (reported in Figure S1) indicate the range of ENVDYN values for which the indirect effect is reliably positive. These results support H4 and align with the theorised second-stage moderated mediation. F. ROBUSTNESS AND ESTIMATION CHECKS We summarise the robustness and estimation alternatives in Table 9, with full results reported in Supplementary Table S5. Heterogeneity analyses and additional validity checks are reported in Supplementary Table S6. Alternative operationalisations. Reweighting GAI facets, using binary adoption, or alternative GAIC codings yields similar signs and significance. OPSCAP composites based on PCA vs equal weights agree closely; excluding one OPS dimension at a time does not overturn results. Alternative SUSPERF proxies (intensity vs absolute; alternative nor-malisations) preserve the main inferences. Lag structure and dynamics. Using t–2 mediators and/or outcomes yields consistent signs with expected attenuation. Distributed-lag summaries remain positive for the cumulative OPSCAP channel. Estimation alternatives. Two-way clustering by firm and year; wild-cluster bootstrap p-values; and (as a sensitivity) system GMM all point to the same qualitative conclusions (see §3.6 and Supplementary Table S4). Multicollinearity and leverage. VIFs and condition indices are within norms. Removing top-leverage observations by Cook’s D/DFBETAs does not change signs or significance. Placebo and falsification. Replacing OPSCAP with an unrelated capability proxy yields no mediation; shuffling ENVDYN within year–industry cells removes the moderation pattern. G. HETEROGENEITY TABLE 10 explores whether effects vary with baseline conditions: Baseline automation/IT intensity. Effects are stronger among firms starting with higher process complexity (automation-intensive segments), consistent with greater scope for cognitive automation and coordination gains. Ownership. Patterns are present for both SOE and non-SOE groups; ENVDYN amplification is more pronounced among non-SOEs (potentially due to tighter budget con-straints and faster decision cycles). Size and region. Larger firms show slightly stronger GAIC → OPSCAP links (scale/fixed-cost absorption), while ENVDYN amplification is visible across regions with stronger effects in more volatile coastal markets. These slices are descriptive but consistent with the theory: where turbulence and complexity are higher, the capability pathway from GAI to sustainability is more valuable. H. ECONOMIC SIGNIFICANCE To complement statistical significance, we translate our standardized estimates into practical benchmarks that map to underlying operational and environmental metrics (e.g., defect and rework rates, on-time delivery, and resource/energy intensity). Because all continuous variables are winsorised at the 1st and 99th percentiles and then standardised within industry–year, coefficients can be interpreted as effects in standard-deviation units along the GAI → OPSCAP → SUSPERF chain. From Table 6, a one-standard-deviation increase in GAIINDEX (roughly moving from a lower-quartile to an upper-quartile adopter within an industry–year) is associated with about a 0.13 standard-deviation increase in OPSCAP (95% CI [0.05, 0.21]). From Table 7, a one-standard-deviation increase in OPSCAP(t-1) is in turn associated with about a 0.20–0.22 standard-deviation increase in SUSPERF (e.g., 0.21 with 95% CI [0.11, 0.31]). Taken together, these estimates imply an indirect effect of GAI on SUSPERF on the order of 0.03 standard deviations for a one-standard-deviation change in GAIINDEX (≈0.027; 95% CI [0.01, 0.05]), and about 0.06–0.07 standard deviations for a shift from the 10th to the 90th percentile of GAI adoption, once we account for both paths and dynamic adjustment (Supplementary Table S4) (95% CI approxi-mately [0.02, 0.12]). While modest in absolute terms, these magnitudes are non-trivial when viewed against the cross-sectional dispersion in SUSPERF: in our sample the difference in average SUSPERF between the bottom and top terciles of firms is around 0.8–1.0 standard deviations, so a realistic GAI upgrade closes roughly 7–10% of the within-industry sustainability-performance gap (implied range based on the indirect-effect CI: approximately 2–14%). The moderated mediation results in Table 8 imply that these gains are substantially larger in more dynamic environments. The conditional indirect effect of GAI on SUSPERF via OPSCAP is about 0.03 standard deviations at low ENVDYN (−1 SD), 0.05–0.06 at medium ENVDYN (around the mean), and roughly 0.07 or more at high ENVDYN (+1 SD), with corresponding 95% clustered-bootstrap confidence intervals reported in Supplementary Table S4 Panel D. In other words, moving from a low- to a high-dynamism industry roughly doubles the sustainability payoff to a given GAI upgrade transmitted through operations capability. This pattern is consistent with the idea that when demand, policy and supply conditions are more turbulent, firms benefit more from faster information flow and cognitive automation, so that improvements in OPSCAP translate into disproportionately larger gains in sustainable performance. Overall, the estimated effect sizes indicate that enterprise GAI adoption and capability are not only statistically significant but also economically relevant drivers of sustainable performance. I. ANCILLARY ANALYSES. We conduct three additional checks (reported in Supplementary Table S6): Alternative timing: contemporaneous vs lagged specifications; results are stronger under lagged designs, as expected if routines take time to absorb GAI. Nonlinearities: testing for diminishing returns by adding quadratic terms for GAIC shows no strong evidence of crowd-ing at typical levels; very high GAIC is rare. Common shocks and policy episodes: including province– year controls for major policy changes does not alter the main inferences. J. SUMMARY OF FINDINGS Across fixed-effects panel models, mediation tests and moderated-mediation analyses, the evidence supports all four hypotheses. GAI adoption and capability are associated with stronger operations capability (H1/H1a); operations capability is positively associated with sustainable performance (H2); OPSCAP mediates the GAI → SUSPERF link (H3); and the indirect effect is stronger under environmental dynamism (H4). The results are robust to alternative operationalisations, lag structures, estimators and diagnostics, and the magnitudes are economically meaningful. V. DISCUSSION AND IMPLICATIONS. A. OVERVIEW OF THE CONTRIBUTIONS Our findings show that enterprise GAI adoption and AI capability are associated with stronger operations capability (OPSCAP); that OPSCAP improves sustainable performance (SUSPERF); that the GAI → SUSPERF link is mediated by OPSCAP; and that the indirect effect is stronger when environmental dynamism (ENVDYN) is higher (see Tables 6–8; Fig. 2). Together with the conceptual model (Fig. 1) and construct mapping (TABLE 1), these results advance the IS–OM conversation by specifying how digital investments convert into performance and when the conversion is most valuable. B. THEORETICAL IMPLICATIONS From adoption to capability. We clarify that technology presence is not capability. By separating GAI adoption (GAIINDEX) from routinised capability (GAIC), the study aligns with dynamic-capability logic and explains why firms with similar adoption statuses realise heteroge-neous outcomes. The evidence that GAIC adds explanatory power beyond adoption supports the view that value cre-ation depends on embedded routines, data governance and coordination. OPSCAP as the transmission mechanism. The mediation results formalise operations capability as the mechanistic bridge linking IS to OM outcomes. Rather than treating performance gains as an opaque mechanism, we show that information richness, shorter decision latency and denser coordination are the channels through which GAI pays off. This mechanism view provides a unifying lens for disparate observations about analytics, automation and digital workflows in manufacturing. Contextualising digital value with ENVDYN. By demonstrating second-stage moderation—ENVDYN amplifies the OPSCAP → SUSPERF link—we position turbulence as a boundary condition on digital returns. This refines ‘‘IT value under uncertainty’’ arguments: the marginal value of codified, information-rich routines rises with volatility, consistent with option-value reasoning and contingency theory. Sustainability as an operations outcome. Modelling SUSPERF (rather than narrow financials) shows that digital capability can co-deliver environmental and operational benefits. This adds an operations-based microfoundation to sustainability claims and complements macro/strategy narra-tives about green performance. C. MANAGERIAL IMPLICATIONS 1) CAPABILITY BUILDING (TURN PILOTS INTO OPSCAP). Budget for the capability bundle—not one-off pilots: invest in data lineage/quality, workflow and prompt libraries, role redesign, and compliance artifacts so that GAI becomes routinised capability (GAIC) rather than isolated experi-mentation. KPI: share of GAI use embedded in workflows; reuse rate of templates/playbooks; data-quality/lineage coverage. Codify operational micro-routines: institutionalise ‘‘retrieve-and-synthesise,’’ ‘‘triage-and-route,’’ and ‘‘draft-and-validate’’ routines with clear SLAs and escalation rules so cognitive outputs reliably enter execution. KPI: exception closure time; SLA compliance; rework/defect rate; handoff count. Design the GAI portfolio for reuse: manage initiatives as (i) platform layers (data, access control), (ii) reusable assets (prompts/templates/workflows), and (iii) line-of-business apps; prioritise modular components over bespoke builds. KPI: cross-unit reuse ratio; pilot-to-product conversion rate; portfolio ROI dispersion. 2) GOVERNANCE (MAKE GAIC RELIABLE AND AUDITABLE). Instrument performance and controls: track process latency (detect–decide–act), coordination density (handoffs, routing, closure), and defect/energy/emissions indicators as lead-ing/lagging metrics; maintain prompt/version libraries, risk registers, and audit trails to institutionalise GAIC. KPI: time-to-decision; audit-trail completeness; override rate and outcomes; energy per good unit; scrap intensity. Scaling Under Volatility (allocate where ENVDYN makes returns larger). Sequence investments by dynamism: prioritise capability building and scaling in business units/plants facing higher demand, policy, or supply turbulence where the OPSCAP → SUSPERF payoff is larger; use calmer areas for low-risk learning. KPI: scaling share in high-ENVDYN units; time-to-policy-update; response lead time under shocks. D. BOUNDARY CONDITIONS AND GENERALISABILITY Industry and process structure. The mechanisms should gen-eralise to settings where processes are codifiable and latency matters (e.g., discrete manufacturing, regulated services). In highly tacit, creativity-dominant tasks, returns may require different complementary assets (e.g., knowledge curation, human factors). Data and governance maturity. GAIC presupposes mini-mum viable data governance and integration. Firms lacking lineage/quality controls may observe weaker or delayed gains because model outputs cannot be absorbed by routines at scale. Institutional context. The study uses publicly listed man-ufacturers within one national context. While the capability logic is portable, institutional differences (reporting norms, regulatory pressure) could shift the levels of adoption/capability and the salience of the sustainability channel. Our evidence suggests three levers that connect GAI to operations: Information richness and accessibility. Retrieval-augmented synthesis reduces search and comprehension costs, enabling faster, broader situational awareness across procurement, production and maintenance. Decision latency reduction. Triage and routing standardise who acts and when; this compresses the detect–decide–act cycle and reduces variability. Coordination density. Shared templates and playbooks align cross-functional responses, raising the probability that detection leads to consistent action. Each lever can be instrumented with measurable KPIs, giving managers a diagnostic dashboard to steer capability building rather than relying on headline adoption. The positive OPSCAP → SUSPERF link indicates that classic operations gains—fewer defects, smoother changeovers, shorter queues—co-deliver environmental benefits (resource/energy efficiency; emissions intensity). This provides a micro-operational pathway for meeting disclosure and compliance requirements without treating sustainability as a separate, add-on programme. In practice, managers can embed environmental KPIs (e.g., energy per good unit, scrap intensity) directly into the OPSCAP dashboard so that process improvements reveal their environmental coin. Phase 1 (0–3 months): baseline and pilots. Audit data lineage/quality; map latency and escalation paths. Stand up 2–3 thin-slice GAI pilots tied to measurable latencies (e.g., engineering change triage, supplier incident routing). Establish prompt/version libraries and governance artifacts. Phase 2 (3–9 months): embed and scale. Convert successful pilots into micro-routines with SLAs; integrate with workflow tools/MES/ERP. Roll out OPSCAP KPIs and dashboards; instrument reliability/responsiveness metrics and link them to energy/scrap indicators. Prioritise scaling in high-ENVDYN lines or plants. Phase 3 (9–18 months): optimise and harden. Refactor bespoke builds into reusable components; auto-mate audit trails. Expand moderated-mediation diagnostics organization-wide (monitor indirect effects by ENVDYN quartiles). Tie capability milestones to budgeting and performance management. E. LIMITATIONS AND FUTURE RESEARCH Measurement refinement. GAIC and GAIINDEX rely on disclosure-based coding. While we report inter-coder reliability and robustness to alternative codings, direct instrumen-tation (system logs; workflow traces) would offer sharper measurement. Future work could fuse digital exhaust with survey-based capability audits. Causal identification. Fixed-effects and lag structures mitigate but do not eliminate endogeneity concerns. Quasi-experimental designs—policy shocks, staggered adop-tions, or instrumented availability—could tighten causal claims. Granularity and heterogeneity. We model firm-level composites. Future work can move to process-level panels (e.g., line/plant granularity) or study human–AI allocation choices that moderate capability conversion. Outcome scope. We focus on integrated sustainability outcomes. Additional lenses—safety incidents, compliance breaches, or customer-level reliability—could enrich the performance map and stress-test the mechanism. By distinguishing adoption from capability, elevating operations capability as the transmission mechanism, and situating returns within environmental dynamism, this study offers an integrated account of how and when GAI creates value in manufacturing. The practical message is straight-forward: invest in capability building and embed cognitive automation into routines, and do so where volatility makes speed and coordination most valuable. Theoretical and empirical extensions can deepen causal understanding and broaden generalisability, but the core implication is robust—capabilities convert digital signals into sustainable operational gains. VI. GUIDELINES FOR GRAPHICS PREPARATION AND SUBMISSION. 1) This study examines how and when enterprise generative-. AI (GAI) creates value in manufacturing by distinguishing adoption from capability and by elevating operations capability (OPSCAP) as the transmission mechanism to sustainable performance (SUSPERF). Using a multi-year firm-level panel and fixed-effects designs, we show that (i) GAI adoption (GAIINDEX) and AI capability (GAIC) are each associated with stronger OPSCAP; (ii) OPSCAP is positively related to SUSPERF; (iii) the GAI → SUSPERF link is mediated by OPSCAP; and (iv) the indirect effect is stronger under higher environmental dynamism (ENVDYN). These findings corroborate a capability-centric view of digital transformation and contextualise digital returns within turbulence. 2. Theoretical contribution. We separate technology presence from routinised capability, showing that the lat-ter adds explanatory power beyond adoption. We specify OPSCAP—via information richness, lower decision latency and denser coordination—as the mechanism connecting IS investments to OM outcomes, and we identify ENVDYN as a boundary condition that amplifies capability payoffs. 3. Managerial implications. Managers should budget for the capability bundle—data governance and lineage, reusable prompts/templates, workflow integration, role redesign and compliance artefacts—rather than isolated pilots. Sequenc-ing matters: capability building should be prioritised where volatility is high, because ENVDYN raises the marginal value of coordinated, information-rich routines. Organisa-tions can operationalise these insights by tracking a small set of leading indicators (process latency, exception closure, handoffs) alongside sustainability KPIs (scrap and energy/emissions intensity). 4. Limitations and future research. Our identification relies on observational panel variation; quasi-experimental designs (e.g., policy shocks or staggered deployments) could strengthen causal claims. Disclosure-based measures of GAIINDEX/GAIC can be refined with digital exhaust (system logs, workflow traces) and process-level audits. Extending the analysis to plant/line granularity, to other industries with different knowledge intensities, and to additional outcomes (safety, compliance incidents, customer-level reliability) would broaden external validity. 5) In sum, capabilities convert digital signals into sustain-. able operational gains. By embedding GAI into routines and aligning investments with environmental turbulence, firms can translate digital adoption into resilient, auditable and scalable performance improvements. SUPPLEMENTARY MATERIALS Table S1. Constructs, operationalisations and measurement details for GAIINDEX, GAIC, OPSCAP, SUSPERF, ENVDYN and controls. Table S2. Additional coding examples and variable dictio-nary (extended items and rubrics for GAIINDEX, GAIC, OPSCAP and SUSPERF). Table S3. Environmental dynamism (ENVDYN): component definitions, single-dimension substitutes and robustness checks. Table S4. Temporal ordering, identification checks, and effect-size translations.Panels A–C report reverse-timing regressions, lead–lag placebos, and alternative lag structures. Panel D reports effect-size translations with 95% confidence intervals. Table S5. Robustness checks and estimation alterna-tives: alternative operations-capability and sustainable-performance measures, two-way clustering, alternative fixed-effects specifications. Table S6. Heterogeneity and additional validity checks: subsample analyses (firm size, industry, digitalisation), and the three additional checks discussed in the main text. Conceptualization: Mintian He; Methodology: Mintian He and Shuili Yang; Software: Mintian He; Validation: Mintian He and Shuili Yang; Formal analysis: Mintian He; Investi-gation: Shuili Yang; Resources: Shuili Yang; Data curation: Mintian He; Writing—original draft: Mintian He; Writing— review and editing: Mintian He and Shuili Yang; Visu-alization: Mintian He; Supervision: Shuili Yang; Project administration: Shuili Yang. All authors have read and agreed to the published version of the manuscript. INSTITUTIONAL REVIEW BOARD STATEMENT Not applicable. The study uses publicly available firm-level secondary data and does not involve human participants or animals. INFORMED CONSENT STATEMENT Not applicable. DATA AVAILABILITY STATEMENT The codebook and variable construction details are provided in the Supplement. Data access may be subject to third-party restrictions; any proprietary identifiers are masked in accordance with disclosure rules. ACKNOWLEDGMENT. The authors thank editors and anonymous reviewers for con-structive feedback. Any remaining errors are their own. CONFLICTS OF INTEREST ABBREVIATIONS (OPTIONAL) GAI — Generative AI (adoption index) GAIC — Generative-AI capability (routinized) OPSCAP — Operations capability SUSPERF — Sustainable performance (integrated outcome) ENVDYN — Environmental dynamism FE — Fixed effects SE — Standard error CI — Confidence interval.