A RAG Approach to Automate Materiality Analysis

Materiality assessment is a critical but resource-intensive process in sustainability reporting, requiring organisations to identify and prioritise the environmental, social, and economic topics most relevant to their stakeholders. This thesis investigates the feasibility of automating this process using large language models (LLMs), proposing a retrieval-augmented generation (RAG) pipeline that identifies and ranks GRI topics from sustainability documents without manual intervention. The pipeline processes documents by splitting them into chunks, identifying relevant GRI topics through structured prompting, and ranking these topics by materiality through pairwise LLM comparison, producing a stakeholder materiality matrix as its final output. The pipeline was compared against two similarity search baselines and a human-annotated ground truth. The pipeline achieves a Pearson score of 0.87 and a Kendall Tau score of 0.80 overall, consistently outperforming both similarity search variants at every level of the hierarchy, demonstrating strong alignment with the ground truth and confirming the viability of LLM-based automation as a promising foundation for supporting real-world materiality assessment.