AI Adoption and Firm Innovation: Evidence from Firm-Level Predictive Models

Abstract

Building on financial and firm-performance literature that uses firm-level indicators to model business outcomes, this article examines whether artificial intelligence (AI) adoption indicators can predict subsequent firm-level innovation outcomes. It addresses the growing need to understand AI adoption as a dynamic firm-level economic process connected to productivity, investment behaviour, R&D activity, and patent-based innovation. The study constructs a harmonized firm-year panel by integrating AI employment and AI-related software-use data with firm financial statements, industry classifications, productivity measures, patent information, and trademark registrations. The modelling sample focuses on the top 50 ever-adopting firms. Innovation is measured primarily through total patent counts. Three predictive approaches are compared using a chronological training-validation-test split: a dynamic firm fixed-effects model, pooled ridge regression, and XGBoost. The models incorporate lagged AI variables, lagged controls, prior innovation outcomes, moving-average terms, and centered time controls. The results show that pooled ridge regression achieves the strongest predictive performance on the held-out test sample, with the lowest RMSE and MAE and the highest R². XGBoost ranks second, outperforming the firm fixed-effects benchmark but not the regularized linear model. The findings suggest that coefficient shrinkage is particularly effective when modelling correlated lagged predictors in firm-level innovation data. The article contributes by offering a leakage-aware predictive comparison of AI adoption and patent-based innovation using firm-level panel data, rather than claiming a causal effect. The findings may help firms, analysts, and policymakers in evaluating how AI employment, AI software use, R&D intensity, investment, and prior innovation histories can inform innovation forecasting and strategic decision-making.