Explainable artificial intelligence for understanding digital technology sophistication and firm productivity in developing and emerging economies

Abstract Digital technologies increasingly shape how firms coordinate internal processes, engage markets, and organise commercial transactions, yet evidence remains limited on whether technology sophistication across distinct business functions is consistently associated with firm productivity and generalises across heterogeneous developing and emerging economies. This study examines business administration, marketing, and sales technology sophistication in relation to relative firm labour productivity using harmonised data from 9664 establishments across Bangladesh, Brazil, Ethiopia, Georgia, Ghana, India, Kenya, Senegal, and Vietnam. Productivity was measured as within-country-wave standardised log sales per permanent full-time employee. The expanded specification accounted for managerial sector experience, foreign ownership, firm size, firm age, workforce tertiary education, export status, capital intensity, and country and sector heterogeneity. Fixed-effects econometric estimation was combined with Extreme Gradient Boosting (XGBoost), geographically grouped cross-validation, a linear predictive benchmark, out-of-fold SHapley Additive exPlanations (SHAP), and leave-one-country-out validation. All three technology dimensions exhibited positive conditional associations with productivity in the preferred specification, although business administration technology sophistication showed the most consistent relationship across robustness analyses, while marketing and sales technology associations were not statistically significant under survey weighting. Capital intensity emerged as the dominant XGBoost predictive feature, while business administration technology sophistication remained the most important technology-related feature. Under grouped country-wave validation, the fixed-configuration XGBoost model achieved an out-of-fold R 2 of 0.152 compared with 0.150 for the linear benchmark, indicating little aggregate predictive advantage from the evaluated nonlinear specification. Leave-one-country-out validation produced pooled R 2 values of 0.148 for XGBoost and 0.143 for the linear benchmark, while country-specific XGBoost performance ranged from − 0.027 to 0.379. The two geographically restrictive validation designs therefore indicate limited and heterogeneous geographic predictive transfer, with little overall predictive advantage from the more flexible learner. These findings show that function-specific technology sophistication contains predictive information about relative sales-based labour productivity but is embedded within a broader configuration of firm resources and does not transfer uniformly across national settings. By combining econometric estimation, predictive benchmarking, geographically restrictive validation, and out-of-fold explainability, the study distinguishes conditional association, predictive importance, and geographic transportability without treating explainable machine learning as evidence of causality.