A data-driven framework for AI adoption efficiency assessment using hybrid DEA and machine learning methods

Despite growing academic interest in artificial intelligence (AI) as a driver of economic transformation, the efficiency with which EU member states translate R&D investment and ICT human capital into measurable AI adoption outcomes remains insufficiently explored. This study proposes a data-driven computational framework that integrates data envelopment analysis (DEA) with machine learning (ML) to assess and explain the efficiency of AI adoption across EU countries during 2021–2024. In the first stage, an output-oriented VRS DEA model is combined with smoothed bootstrap bias correction. In the second stage, extreme gradient boosting (XGBoost) and random forest models, with Shapley additive explanations (SHAP)- based decomposition, are employed to identify key predictors of efficiency, while bootstrapped truncated regression provides complementary statistical inference for robustness validation. Average bias-corrected efficiency increased over time, reflecting the diffusion of AI technologies across the EU. However, the results reveal substantial heterogeneity in the efficiency of AI adoption across EU member states. SHAP analysis identifies GDP per capita, regulatory quality, and the share of the ICT sector as the most important predictors of AI adoption efficiency. Country-grouped cross-validation indicates limited out-of-country predictive generalisability, while sensitivity analyses confirm broad stability of the leading-predictor structure. The truncated regression yields no statistically significant covariate effects, highlighting the inferential uncertainty associated with the small country-year sample. The study demonstrates the usefulness of combining bootstrap-corrected DEA, XGBoost, random forests, and SHAP for interpretable efficiency analysis in rapidly evolving technological environments.