AI technologies affording the orchestration of ecosystem-based business models: the moderating role of AI knowledge spillover

The paper addresses how AI, as a core technology of the fourth industrial revolution, influences the orchestration of ecosystem-based business models (EBMs) oriented toward the triple bottom line (profit, people, planet). While AI transforms traditional business operations and enables digital infrastructures, its implications for business models, especially EBMs, are underexplored. Grounding in the affordance perspective—where AI is the object, EBM stakeholders are the user context, and effective EBM orchestration is the goal—the study formulates two research questions: (1) How and whether AI technologies afford the orchestration of EBMs? (2) How do AI knowledge spillovers affect the relationship between AI and EBM? The study aims to provide empirical evidence and practical insights for configuring AI-enabled EBMs aligned with UN SDGs.

The theoretical foundation spans three areas. (1) AI and business models (BMs): AI can be integrated across products and value chains, driving BM innovation across sectors (agriculture, healthcare, manufacturing, energy). Despite potential, adoption faces barriers including trust, reliability, fairness, and costs; many firms realize limited benefits, reflecting a productivity paradox. (2) Affordance perspective: Affordances capture action potentials offered by an object (technology) to an actor within a use context. Digital affordances from advanced technologies enable new interactions and coordination among ecosystem stakeholders. (3) EBMs and AI: EBMs are digital-driven, platformized, and stakeholder-diverse ecosystems emphasizing profit, people, and planet. AI and other digital technologies (e.g., blockchain) facilitate cross-boundary collaboration, transparency, and value co-creation but also introduce complexity and resource demands. Hypotheses: H1 posits an inverted U-shaped relationship between AI and EBM. H2a and H2b posit that direct and indirect knowledge spillovers, respectively, moderate this relationship by enabling actors to recognize and exploit digital affordances.

Data: Panel of Chinese A-share listed firms on Shanghai and Shenzhen exchanges from 2014–2021. Filtering removed ST/*ST firms, suspended/terminated firms, and those lacking required data, yielding 3,632 firm-year observations from 454 companies. Data sources: CSMAR (financials, supply chain, AI keywords), and Hexun Score (CSR) for stakeholder and environmental metrics.

Measures: Dependent variable (EBM) constructed via entropy weighting using 13 indicators covering: (a) traditional BM dimensions—value creation (current ratio, debt coverage ratio, capitalization ratio), value delivery (inventory turnover, receivables turnover, total asset turnover), value capture (revenue growth, net profit growth, profit margin); and (b) EBM extensions—people (shareholder, employee, supplier/customer/consumer responsibilities) and planet (environmental responsibility) from CSR. Indicators were standardized (min-max), entropy values and weights computed, and a composite EBM score derived.

Independent variable (AI degree): Word frequency of AI-related terms disclosed in annual reports (e.g., AI, machine learning, deep learning, NLP, robotics, biometrics, autonomous driving, etc.) extracted via CSMAR’s text mining pipeline validated by domain experts.

Moderators: Direct knowledge spillover (DKS) measured as R&D capital stock using the perpetual inventory method with depreciation rate 26.37% (per Xu et al., 2023). Indirect knowledge spillover (IDKS) measured as management costs divided by revenue, capturing indirect internal knowledge diffusion through managerial and organizational communication.

Controls: Firm size (log total assets), intangible asset ratio (intangible/total assets), supply chain concentration (average of top 5 supplier purchase ratio and top 5 customer sales ratio), CEO age, and CEO overseas background (dummy).

Model and estimation: Firm- and year-fixed effects panel regressions with robust inference; Hausman test favored fixed effects (chi-square=147.63, p<0.001). H1 tested by including AI and AI squared. H2a/H2b tested by adding moderator, AI×moderator, and AI^2×moderator. Multicollinearity was low (max VIF 2.42, mean 1.38). Robustness checks: (i) one-year lagged dependent variable (EBM_{t+1}); (ii) sample excluding 2021 (COVID-19 impact). Results were consistent across checks.

• Descriptive statistics: Mean EBM=0.015 (min 0.003, max 0.308). Mean AI mentions=8.93 (min 1, max 259). DKS and IDKS show substantial dispersion across firms.
• Inverted U-shaped relation (H1): Model 3 shows AI coefficient=0.085 (p<0.10) and AI^2 coefficient=-0.150 (p<0.01), supporting an inverted U-shape between AI and EBM. The turning point occurs at AI≈40.53 within the observed range (1–259). Thus, EBM performance initially increases with AI but declines beyond moderate levels.
• Moderation by direct knowledge spillover (H2a): Model 5 shows DKS×AI=-0.311 (p<0.01) and DKS×AI^2=0.138 (p<0.05). The computed β1β4−β2β3<0 indicates the turning point shifts left as DKS increases. The inverted U flattens with higher DKS and can become U-shaped at high DKS, indicating strong direct spillovers dampen and eventually reverse the curvature.
• Moderation by indirect knowledge spillover (H2b): Model 7 shows IDKS×AI=0.139 (p<0.01) and IDKS×AI^2=-0.137 (p<0.05). β1β4−β2β3>0 implies the turning point shifts right with higher IDKS. The inverted U becomes steeper at high IDKS levels, enhancing benefits at low-to-moderate AI and accentuating penalties at very high AI.
• Controls: Supply chain concentration is negatively associated with EBM (weak significance). Other controls are generally insignificant.
• Model fit: R^2 improves with nonlinear and interaction terms (up to ~0.342).
• Robustness: Using lagged EBM and dropping 2021 both preserve the inverted U and the moderating patterns. For lagged EBM: AI^2≈-0.110 (p<0.01), DKS×AI≈-0.422 (p<0.01), DKS×AI^2≈0.198 (p<0.01), IDKS×AI≈0.062 (p<0.10), IDKS×AI^2≈-0.090 (p<0.01). Excluding 2021 yields similar magnitudes and significance (e.g., AI≈0.096, p<0.05; AI^2≈-0.169, p<0.01).

Findings confirm that AI affords the orchestration of EBMs up to a point: at low-to-moderate levels, AI’s digital affordances enhance value creation, delivery, and capture across diverse stakeholders, improving EBM outcomes. Beyond a threshold, escalating costs, complexity, capability gaps, and integration challenges outweigh benefits, yielding diminishing and negative returns—consistent with the affordance lens and the productivity paradox literature. Knowledge spillovers are pivotal: direct spillovers (R&D-based) shift the turning point left and flatten the curve, indicating earlier onset of decreasing returns as intense R&D exchanges may create coordination overloads or misalignment with broader ecosystem needs. Indirect spillovers (managerial/organizational knowledge flows) shift the turning point right and steepen the inverted U, suggesting that broader, diffuse knowledge diffusion helps stakeholders absorb and exploit AI’s affordances more effectively for longer before hitting complexity limits. Together, the results illuminate mechanisms by which knowledge contexts shape how AI supports EBM orchestration, offering an empirically grounded link between AI, knowledge spillovers, and sustainable, ecosystem-oriented business practices aligned with SDGs.

The study provides large-scale empirical evidence that AI’s relationship with ecosystem-based business models is inverted U-shaped, with both direct and indirect knowledge spillovers moderating this curvature in distinct ways. This advances BM and affordance theories by quantifying how digital affordances translate into ecosystem orchestration and how knowledge environments condition that translation. Practically, firms should calibrate AI investment to avoid overcomplexity, foster indirect knowledge diffusion mechanisms to extend AI’s beneficial range, and manage direct R&D-driven spillovers to prevent overload. Future research can generalize across countries and sectors, refine EBM and AI measures, and explore micro-level mechanisms and governance designs that optimize knowledge flows for sustainable EBM outcomes.

The sample is limited to Chinese A-share listed firms (2014–2021), potentially constraining generalizability across geographies and firm types. Measurement choices (e.g., AI via word frequency, EBM via entropy-weighted indicators, IDKS via management cost ratio) may introduce construct and proxy limitations. The time horizon is relatively short for structural transformations. Future work should expand datasets internationally, test alternative measures and identification strategies, and explore longitudinal and causal designs to validate and deepen these findings.