The evolution of AI-driven sustainable finance: A bibliometric and systematic review of ESG performance and investment strategies

Abstract

The exponential growth of Artificial Intelligence (AI) has revolutionized sustainable finance, transforming how investors evaluate Environmental, Social, and Governance (ESG) criteria. However, a significant gap persists regarding the practical integration and transparency of these complex algorithmic models into reliable investment signals, as traditional approaches are often hindered by data lags and greenwashing risks. To address this evolving challenge and bridge the divide between technical innovation and portfolio application, this study employs a rigorous hybrid research design. The proposed study integrates a bibliometric mapping of 388 Scopus-indexed documents published between 2020 and February 2026 alongside an in-depth systematic thematic review of 39 high-impact, implementation-focused studies. The analysis identifies a transformative shift in model dominance driven by data structure: tree-based ensemble models currently dominate structured corporate sustainability assessment (observed in 8 key studies), while Transformer and Large Language Model (LLM) architectures are emerging to extract narrative signals from unstructured text. Additionally, Graph Neural Networks (GNNs) are increasingly adopted to map hidden, non-linear dependencies in credit risk and complex supply chains. Furthermore, the evaluation highlights a critical need for model interpretability; while Explainable AI (XAI) frameworks like SHAP and LIME are increasingly utilized to meet strict regulatory transparency requirements and clarify ESG predictions, many existing models still fail to provide actionable validation for investors. From a practical perspective, this synthesis offers a strategic roadmap for fund managers. By translating raw ESG signals into decision-grade intelligence, it facilitates the integration of transparent, XAI-verified outputs into active portfolio optimization, risk management, and the mitigation of greenwashing risks, directly supporting strategies like best-in-class screening and thematic tilting. Finally, the integrity and scholarly reliability of these insights are reinforced by the strict adherence to the PRISMA protocol, ensuring the reproducibility and long-term validity of the synthesized evidence.