As climate transparency requirements intensify, ESG teams within companies are struggling to keep pace. In Switzerland and beyond, the production of reliable indicators still relies heavily on manual processes or surveys—approaches that are often costly and difficult to scale. This tension between growing regulatory demands and limited resources calls for a rethinking of measurement methods.
In a recent white paper published by Enterprise for Society (E4S), Professors Eric Jondeau and Fabio Alessandrini at HEC Lausanne (Unil), together with Nathan Delacrétaz, HEC Phd today at Quanthome, explore an alternative based on artificial intelligence: automatically extracting ESG information from companies’ annual reports.
Leveraging existing data rather than collecting new information
The proposed approach relies on language models capable of analyzing large volumes of text. Instead of directly surveying companies through questionnaires, it leverages information already disclosed in corporate reports.
These models are trained to identify and structure data according to a defined ESG framework, transforming narrative content into quantifiable indicators. This method taps into a widely available yet still underutilized source of information.
A scalable solution
The results show that automated extraction achieves around 80% accuracy in evaluating 36 ESG criteria. This approach makes it possible to analyze a large number of companies quickly, including those that do not traditionnaly publish ESG reports.
The study also finds that lighter models, deployed locally, can perform as well as more expensive solutions. Depending on the use case, some models are better suited for rapid screening, while others meet stricter compliance requirements.
Limitations to consider
This approach nevertheless depends on the quality of the information disclosed in reports. Undisclosed or insufficiently detailed elements remain beyond reach, and models may occasionally miss important nuances. Their use therefore requires careful validation and rigorous calibration.
By automating the extraction of unstructured data, this method significantly reduces the manual workload that still dominates ESG analysis, while paving the way for broader and more consistent company coverage.
