Generative AI for Simplified ESG Reporting in Financial Services

Tools

Zhang, Hao and Bowden, James and Cummins, Mark and Black, Findlay (2025) Generative AI for Simplified ESG Reporting in Financial Services. University of Strathclyde, Glasgow.

Full text not available in this repository.Request a copy

Abstract

We demonstrate the potential for Generative AI to simplify Environmental, Social, and Governance (ESG) reporting in financial services. Banking and financial institutions are required to comply with ever more stringent and demanding ESG related compliance requirements. A lack of mandatory, universally enforceable sustainable finance standards and guidelines makes effective ESG reporting across industries and countries difficult for financial institutions. Vast amounts of data processing are required, spanning structured quantitative numerical data and unstructured qualitative textual data. Generative AI has the potential to deliver an innovative solution to this ESG reporting challenge through identifiable capabilities in decision support, including document summarisation; data visualisation; individual and multiple company analytics; and customised report generation. Furthermore, several technical features allow organisations to customise Generative AI systems to meet bespoke business requirements and information technology constraints. These technical features include response speed and agility; multiple version choice and algorithmic support; user friendly interfaces; scalability and upgradability. In the use case demonstration, we show how a Large Language Model (LLM) can be used to generate responses to a set of common analyst questions pertaining to ESG using single and multiple annual report sources. This use case brings to life the potential for Generative AI in simplifying compliance in respect of ESG reporting. We then bring together LLM and cutting-edge large Vision Model (LVM) capability to move from text-based prompting to verbal-based prompting for the ESG reporting exercise. We show that this integrated language-vision approach leads to enhancements in performance compared to a sole LLM approach. Indeed, we demonstrate that placing emphasis on key words within the verbal prompts generates more targeted responses from the LLM.

ORCID iDs

Zhang, Hao ORCID logoORCID: https://orcid.org/0009-0002-1021-7476, Bowden, James ORCID logoORCID: https://orcid.org/0000-0002-0419-1882, Cummins, Mark ORCID logoORCID: https://orcid.org/0000-0002-3539-8843 and Black, Findlay;
CORE (COnnecting REpositories)