Beyond Prediction: Bridging the AI Intervention Gap in Pavement Maintenance through Evidence-based Decision-making: A Critical Review

Road agencies are investing in artificial intelligence to inspect pavements, forecast deterioration and plan maintenance, yet the value of these tools ultimately depends on whether they improve the interventions that are funded and delivered. This critical narrative review examines the gap between predictive performance and intervention quality in pavement maintenance, termed here the artificial intelligence intervention gap. Literature was identified through structured searches of multidisciplinary and regional scholarly indexes, supplemented by citation tracking and institutional sources, and was appraised for design adequacy, validation practice, causal identification and decision relevance. The synthesis is organised along a decision chain that runs from sensing and condition assessment, through deterioration prediction and estimation of treatment effects, to optimisation, implementation and outcome feedback. The evidence is strongest for automated distress detection and short-term condition prediction, where deep learning and ensemble models frequently achieve high accuracy on benchmark or single-agency data. Confidence weakens progressively along the chain. Performance claims rarely survive transfer across regions, survey technologies or time periods, validation designs often ignore spatial and temporal dependence, and explanations derived from post hoc attribution methods are commonly interpreted as causal mechanisms. Treatment effects are usually inferred from observational records in which maintenance was assigned according to condition, so selection bias and endogeneity remain largely unaddressed outside a small econometric and causal-inference literature. Reinforcement learning and metaheuristic planning studies report cost savings, but most are trained and evaluated within simulators built from the same associational models, and few compare recommended policies with realised outcomes. Implementation evidence, including digital twins and language-model assistants, is at an early, largely demonstrative stage. The review argues that pavement maintenance requires an evidence-based decision-making standard in which intervention effects are explicitly identified, uncertainty is propagated into decisions, decision quality rather than prediction error is the primary evaluation criterion, and agencies monitor the outcomes of recommended treatments. Priorities include benchmark datasets with recorded interventions, quasi-experimental and pragmatic field evaluations of treatments, decision-focused model training and prospective comparison of algorithmic and conventional programmes.