The adoption of recycled and low-carbon materials in asphalt pavements is constrained by the lack of reliable decision support during early-stage mix design, when laboratory performance data are scarce or unavailable. Existing machine-learning approaches largely assume abundant historical data and therefore struggle to guide mixture proportioning for novel or locally sourced recycled constituents. This paper presents AI2GreenPave, a data-scarcity–native artificial intelligence framework for sustainable asphalt mix design that integrates agent-based modelling, knowledge-based engineering, explainable machine learning, into a unified decision-support system. AI2GreenPave operates by (i) explicitly encoding mix-design rules, specifications, and sustainability constraints in a knowledge base; (ii) applying a BorutaSHAP-driven feature-selection framework to structure the design space under limited data; and (iii) using agent-based simulation to generate mechanistically informed indicators related to aggregate packing and volumetric feasibility that enable early screening of infeasible or high-risk mixtures prior to extensive laboratory testing. Machine-learning models trained on public Marshall datasets are then incrementally refined as new experimental data become available, with Shapley Additive Explanations (SHAP)-based explanations ensuring transparency and traceability. Unlike conventional data-hungry workflows, AI2GreenPave is designed to function before large experimental datasets exist, reducing reliance on trial-and-error testing while maintaining compliance with engineering heuristics and specifications. The proposed architecture demonstrates how hybrid AI systems can accelerate the adoption of recycled materials in asphalt pavements by providing reliable, interpretable guidance under real-world data constraints.