Hur effektivt kan AI användas för att extrahera ESG-relaterade nyckeltal från årsrapporter?

As ESG (Environmental, Social and Governance) reporting becomes more mandated by EU regulations such as SFDR and CSRD, many organizations face challenges in obtaining, structuring and verifying this data. This report explores the feasibility of using locally deployed large language models (LLMs) to extract ESG-related KPIs from company annual reports in XHTML format with some data being present within the iXBRL tags. The study was carried out in collaboration with Captor, a fund management company, which aims to replace the dependence on costly third-party ESG data providers, whose data can be outdated at times. A Python based pipeline was developed with the use of the Ollama framework to run three open-sourced Large Language Models (LLMs) Mistral, Deep Seek and Command on Apple silicon hardware. Through Retrieval Augmented Generation (RAG) and semantic chunking the XHTML files were parsed and JSON files containing scope 1,2 and 3 as well as revenue and LEI codes were outputted. The overall accuracy in extracting data from the preprocessed data was shown to be between 79 and 91%. However, the observation that extracting from iXBRL tags had a promising 100% accuracy compared to extracting from free text. The results suggest that small-scale local AI deployment can support accurate and scalable ESG data extraction, supporting compliance with EU regulations. The project also sheds light on broader challenges within reporting, including inconsistent XHTML formatting and regulatory frameworks that variate highly between financial and non-financial reporting. Inspired by an interview conducted with the company as well as Mario Draghi’s 2024 EU report, the thesis also argues the importance for a centralization system for ESG as well as financial data, directing focus on the build up of data infrastructure. These insights could contribute to technical advancements and act as a catalyst for innovation within data disclosure.