Measuring AI Automation Exposure for Sustainable Workforce Transitions: A Two-Mechanism Decomposition

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Artificial intelligence (AI) automates work through two mechanisms: codification-based automation, which applies rule-based systems to structured tasks, and learning-based automation, which applies machine learning and generative AI to pattern- and language-intensive tasks. An account of task-based exposure must represent both, yet each existing measure captures a single mechanism or conflates the two. Using publicly available O*NET data, we decompose exposure into a Routine Task Intensity (RTI) measure for codification and a Machine Learning Capability Index (MLCI) for learning, combined into a composite AI Exposure Index (AEI). The two components correlate weakly, so no single index can stand in for both. Validating against O*NET’s incumbent-reported degree of automation across two periods (2011–2019 and 2020–2025), we find realized automation remains dominated by codification, while the learning channel registers only recently and faintly. The two components also interact: each predicts realized automation most strongly where the other is absent, so entering them jointly with their interaction outperforms either component alone and the composite. Because forecasts built on such indices inform where retraining and income support are directed, accurate mechanism attribution bears on decent work and inequality. The decomposition offers a more complete, transparent, and reproducible account of task-based exposure than any single-mechanism index.