SAGE-AIP: A Scale-Adaptive Gated Ensemble Framework for Predicting AI Adoption and Productivity Across Multiple Data Scales

Artificial intelligence (AI) is frequently associated with productivity improvement, yet empirical conclusions depend on where AI exposure is measured, how productivity is defined, and which model is used. This study develops a multi-scale comparative framework using three datasets: 145,000 AI-adoption records from 10 countries and eight industries; 2,074 observed country-year records from 233 country or regional units (2015–2023); and 150 employee-level observations. Robust regression, country fixed effects, dynamic fixed effects, and difference-GMM sensitivity analysis are combined with a proposed Scale-Adaptive Gated Ensemble Framework for AI Adoption and Productivity (SAGE-AIP), which is benchmarked against six established regressors using temporal holdouts for the two longitudinal datasets and a train–test split for the employee data. The adoption dataset shows almost no systematic structure (R2 = 0.00018), and SAGE-AIP does not improve on simpler models there (test RMSE 28.976 vs. 28.900 for Elastic Net). In the global panel, automation intensity is strongly associated with AI-attributed productivity (β = 19.882, SE = 0.434), and SAGE-AIP attains the lowest test RMSE (4.448; R2 = 0.544). At the employee level, recorded AI-tool information remains positively associated with productivity (β = 0.905, SE = 0.220), but Random Forest outperforms SAGE-AIP (test RMSE 0.435 vs. 0.583). SAGE-AIP is therefore not uniformly superior; its value is scale-specific, which supports evaluating models across data scales rather than relying on a single dataset or model class.