AI-ENABLED ESG INTELLIGENCE: A SYSTEMATIC REVIEW OF NLP AND PREDICTIVE ANALYTICS FRAMEWORKS FOR AUTOMATED SUSTAINABILITY REPORTING (2020–2025)

Resumen

Environmental, social, and governance (ESG) reporting is undergoing a fundamental transition from voluntary narrative disclosure to standardised, mandatory, and audit-ready reporting, driven by the convergence of multiple international and regional regulatory frameworks. This shift has generated substantial demand for AI-enabled ESG intelligence systems capable of processing heterogeneous data sources and producing decision-relevant, traceable, and assurance-ready outputs. However, ESG data remains characterised by significant structural fragmentation, definitional inconsistency across rating methodologies, and the absence of standardised taxonomy mappings conditions that create material barriers to supervised learning approaches reliant on consistent ground-truth labelling [9, 12]. This paper presents a PRISMA 2020-compliant systematic review of AI-based ESG reporting automation research for the period January 2020 to February 2025, encompassing 19 studies selected from 412 records initially retrieved across Scopus, Web of Science, IEEE Xplore, ACM Digital Library, SSRN, and Google Scholar. The review is structured across four functional stages: (i) data acquisition and harmonisation; (ii) NLP-based extraction and classification of ESG statements across environmental, social, and governance disclosure pillars; (iii) predictive analytics for forward-looking risk and disclosure gap forecasting; and (iv) governance, assurance readiness, and reporting design. The regulatory analysis encompasses the EU Taxonomy, CSRD/ESRS, ISSB IFRS S1/S2, and the Saudi Capital Market Authority (CMA) ESG Disclosure Guidelines (2023). Domain-adapted transformer models have demonstrated consistent performance improvements in ESG text classification and entity extraction benchmarks, with F1 score gains of 5–15 percentage points over general-purpose language model baselines [14, 16]. Notwithstanding these advances, ensemble and temporal models demonstrate superior performance for ESG risk prediction tasks, particularly when integrating structured ESG metrics with textual sentiment features. Systemic challenges persist, including the absence of verified ground-truth ESG labels, regulatory domain shift, and limited supply chain reporting data. The synthesis produces a deployable seven-step methodological blueprint validated against ESRS, ISSB, CSRD, and Saudi CMA design constraints, with direct applicability to Saudi Vision 2030 sustainability governance, the Public Investment Fund (PIF) portfolio reporting agenda, and Tadawul-listed entity disclosure obligations.