Companies’ reports, often include valuable non-financial information related to environmental, social, and governance (ESG) factors. Extracting this non-financial data is vital to asset managers and investors aiming to support sustainable projects. However, the manual extraction of ESG information is complicated by the absence of standardized ESG-specific sections and the diverse presentation formats of such data. This study evaluates and compares various AI-driven approaches, including semantic similarity retrieval (SSR) methods and supervised learning models, to automatically extract and classify ESG-related content from these reports. Our findings reveal that supervised methods such as FinBERT-ESG (Financial Bidirectional Encoder Representations from Transformers) generally achieve higher accuracy, particularly in identifying environmental data, and require significantly more resources and complexity, with only marginal gains over simpler SSR methods. When combining the outputs of multiple models, we observe a slight improvement in overall performance, suggesting that ensemble or hybrid approaches may enhance robustness. Furthermore, the results indicate that model performance is inconsistent across ESG categories, with notable difficulties in detecting governance aspects due to the nuanced and overlapping nature of the content. These insights underscore the necessity for developing more context-aware AI models that can better capture the intricacies of ESG information, improving both the precision and efficiency of data extraction. This study offers practical insights into the strengths and limitations of current AI methodologies and highlights areas for future advancement in automated ESG data processing for sustainable finance.