A comprehensive systematic literature review assessing the application of artificial intelligence across environmental, social, and governance financial intelligence has determined that advances in computational model architectures have significantly outpaced the data infrastructures required to support them. Published in the journal Discover Artificial Intelligence, the investigation was conducted by researchers Miehleketo Mathebula, Abiodun Modupe, Vukosi Marivate, and Elda Du Toit. The authors examined 70 peer-reviewed journal articles published between 2017 and 2025 to determine whether prevailing operational limitations in sustainable finance analytics stem from inherent deficiencies in computational models or from underlying data fragmentation and institutional disparities. Their primary finding reveals that the principal constraints restricting dependable analytical deployment arise from unstandardised disclosure environments, incompatible metadata schemas, and proprietary data dependencies rather than computational power or algorithmic complexity.
The methodological framework followed Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines across four major interdisciplinary databases, specifically IEEE Xplore, ScienceDirect, SpringerLink, and MDPI. An initial search executed on January 7, 2026, captured 1,623 candidate records across the computer science, financial analytics, and corporate sustainability literature. Following the removal of 57 duplicate records and 78 preliminary exclusions based on publication type, non-English text, or temporal mismatch, 1,488 records underwent title and abstract screening. Of these, 1,357 records were excluded for failing to satisfy thematic criteria, leaving 131 reports sought for retrieval. Thirteen full texts could not be retrieved, yielding 118 complete articles for formal eligibility assessment. Forty-eight additional studies were subsequently eliminated due to insufficient computational integration or weak ESG linkages, producing a final corpus of 70 peer-reviewed journal articles evaluated across four primary research questions.
Methodological quality across the reviewed corpus was appraised using an explicit six-criterion appraisal framework yielding a maximum cumulative score of ten points. The evaluation rubric allocated three points to the rigour of model validation, two points to engagement with ESG reporting standards, two points to relevance within emerging-market or data-constrained settings, and one point each to bias mitigation, methodological transparency, and explainability mechanisms. Based on these cumulative metrics, 49 studies were classified as high quality, scoring between 7 and 10 points, while 21 studies were designated as medium quality, scoring between 4 and 6 points, with all investigations scoring below 4 points excluded. To ensure consistency, two independent reviewers performed the extraction and scoring, achieving an inter-rater agreement Cohen’s kappa coefficient of 0.82, which confirms substantial consensus across both study selection and qualitative appraisal tiers.
Bibliometric mapping revealed a marked acceleration in academic output toward the end of the observation window, with 2024 and 2025 accounting for 17 and 18 publications respectively, representing 50 percent of the total reviewed corpus. Prior to this expansion, early exploratory work between 2017 and 2019 generated only 12 publications in total, followed by moderate activity between 2020 and 2023 ranging from 3 to 7 papers annually. Publication activity was heavily concentrated in interdisciplinary technology and sustainability venues, led by IEEE Access with 14 articles, Sustainability with 11 articles, and Artificial Intelligence Review contributing 5 articles, alongside specialised finance journals such as Finance Research Letters. Citation metrics demonstrated a right-skewed distribution across the corpus, characterized by an overall mean citation count of 149 and a median of 42, reflecting the temporal lag in citation accrual for investigations published after the rapid market adoption of generative models in 2023.
The geographical and institutional distribution of the evidence base revealed pronounced regional imbalances and fragmented international collaboration networks. Asian institutions authored the largest volume of publications across structured, unstructured, and multimodal data categories, followed closely by European and North American researchers, whereas African institutions contributed eight studies. Country-level co-authorship networks comprising 26 nations and 32 collaborative links demonstrated a modularity score of 0.490 and a network density of 0.098, confirming moderate clustering around a few dominant nodes. China recorded the highest degree centrality of 0.280 and betweenness centrality of 0.273, followed by the United Kingdom with degree centrality of 0.240 and betweenness of 0.198, and the United States with degree centrality of 0.200 and betweenness of 0.126. South Africa produced 11 documents but exhibited a total link strength of only 2, demonstrating that research within certain emerging economies remains largely domestic rather than globally integrated.
Computational paradigms across the reviewed corpus displayed an ongoing structural transition from traditional tabular machine learning toward contextual natural language processing and generative language modeling. Traditional machine learning represented the single most frequent technique with 40 occurrences, followed by deep learning and neural network architectures with 36 occurrences, and natural language processing with 26 implementations. Transformer-based models including BERT and FinBERT recorded 14 occurrences, while general large language models accounted for 15 occurrences. Hybrid computational approaches, retrieval-augmented generation systems, and explainable artificial intelligence frameworks each recorded 12 occurrences, alongside 11 sentiment analysis implementations. Less frequent techniques included predictive time-series forecasting with 10 occurrences, topic modelling with 9, clustering with 7, knowledge graph representations with 7, generative artificial intelligence with 6, and Internet of Things integrations with 5 occurrences.
In categorising data modalities, the synthesis established five operational groupings comprising corporate disclosures, policy texts, structured firm indicators, governance data, and financial sentiment streams. Corporate disclosures encompassed sustainability reports, regulatory 10-K filings, and earnings call transcripts, which provided rich contextual narratives but exhibited severe cross-jurisdictional vocabulary divergence. Structured indicators formed the most directly comparable modality, including accounting ratios, panel datasets, carbon emissions metrics, and commercial ESG ratings. However, these structured datasets introduced acute reproducibility barriers due to pervasive reliance on licensed, proprietary databases with opaque rating methodologies. Financial sentiment streams drawn from news and social platforms offered dynamic market signals, yet suffered from temporal noise and text brevity, while policy and sustainable development goal corpora introduced cross-border semantic ambiguities that hindered automated extraction pipelines.
The authors identified five core financial intelligence task formulations operationalised across the literature, including disclosure extraction, performance scoring, greenwashing detection, sustainable reporting intelligence, and climate-aligned finance. Automated indicator extraction converted unstructured sustainability narratives into quantitative disclosures, while scoring tasks benchmarked corporate practices against historical ratings. Investigations into greenwashing deployed regression and text-mining models to detect incongruities between corporate environmental claims and objective operational performance, with several studies documenting that artificial intelligence adoption correlates with reduced greenwashing through enhanced internal disclosure controls. Nevertheless, most examined tasks were executed in complete isolation on curated benchmark corpora, rarely demonstrating operational integration with live compliance verification, supervisory auditing pipelines, or multi-source investment workflows.
Synthesised constraints were classified into five persistent structural barriers: restricted data access, fragmented reporting standards, machine-readability deficits, governance risks, and technical infrastructure bottlenecks. Overlapping and conflicting international frameworks, notably the Global Reporting Initiative, the Sustainability Accounting Standards Board, and the Task Force on Climate-related Financial Disclosures, generate inconsistent corporate reporting structures that undermine cross-border algorithmic generalisability. Furthermore, explainability and accountability components were almost universally applied post hoc via attribution methods like Shapley Additive Explanations or Local Interpretable Model-agnostic Explanations, rather than embedded directly into initial system designs. This widespread reliance on interpretive overlays fails to deliver the verifiable data provenance, auditability, and regulatory assurance necessary for high-stakes financial environments.
To contextualise these operational frictions, the review synthesised an evidence-informed conceptual architecture comprising preprocessing, semantic standardisation, contextual modelling, task inference, and governance oversight. The researchers illustrated this framework through a practical walkthrough of a South African corporate sustainability report prepared under the King IV governance code. Because King IV principles emphasize integrated thinking, stakeholder inclusivity, and ethical leadership, its narrative concepts do not align neatly with international schema taxonomies like GRI or SASB. Consequently, semantic misalignments introduced during document ingestion propagate through vector retrieval, embedding, and inference stages, yielding distorted scoring or risk classifications. The authors stressed that this architecture represents an unvalidated conceptual design blueprint rather than a deployed software solution, demonstrating that technological interventions cannot compensate for absent semantic crosswalks.
Ultimately, the systematic review establishes that the critical challenges facing sustainable financial intelligence are fundamentally infrastructural and institutional rather than computational. The authors urge the research community to discontinue treating commercial ESG ratings as objective ground-truth labels and to cease relying on post-hoc explainability to compensate for unverified data provenance. Progress in automated financial intelligence will require concerted global investments in open-access benchmark datasets, standardized multilingual corpora, and jurisdiction-aware reporting frameworks. By shifting strategic focus away from standalone model complexity and toward rigorous data quality, ontology alignment, and embedded governance mechanisms, the field can build verifiable analytical pipelines that perform reliably across both established and emerging market disclosure environments.
Subject of Research: Systematic review of artificial intelligence methods and data constraints in ESG financial intelligence
Article Title: A systematic review of artificial intelligence methods and data constraints in ESG financial intelligence with implications for emerging markets
Article References: Mathebula, M., Modupe, A., Marivate, V., & Du Toit, E. (2026). A systematic review of artificial intelligence methods and data constraints in ESG financial intelligence with implications for emerging markets. Discover Artificial Intelligence, 6(1), Article 1935. https://doi.org/10.1007/s44163-026-02647-5
Image Credits: AI Generated
DOI: 10.1007/s44163-026-02647-5
Keywords: artificial intelligence, ESG reporting, financial intelligence, systematic literature review, machine learning, deep learning, transformer models, large language models, data infrastructure, greenwashing detection, sustainability disclosures, emerging markets, explainable AI, King IV, semantic interoperability, PRISMA methodology, corporate governance, metadata fragmentation