Simplifying Double Materiality Assessment via IPA
Challenge to overcome
The requirement to perform and disclose the result of a “double materiality assessment” (DMA) was introduced in the European Union by the Corporate Sustainability Reporting Directive in 2022. In Switzerland, the Code of Obligations (Art. 964a-c) contains a similar, although less precise and prescriptive, expectation. A number of Swiss companies also perform DMAs as part of their sustainability strategy activities.
The DMA is (as of now) a mostly qualitative procedure meant to assess how a company affects ESG components and, in turn, how the ESG components can affect said company. Performing a DMA requires to first determine a series of impacts, risks and opportunities (IRO) and then go through any available information source and perform an assessment of factors such as Severity or Likelihood of each. As the list of IROs can be very large, the process can be very time consuming for the teams performing it.
AI application / use case
The time intensity of the process for humans made it clear that intelligent automation, typically via a Large Language Model (LLM), could provide valuable support in improving the sourcing of information from the documents as well as speeding up the process. The solution relies on the following steps:
- Defining a list of all IROs to be assessed, with the relevant factors (e.g. Severity, Scope and Likelihood) for each of those IROs.
- Drafting the related questions that a human team member would ask themselves when screening underlying documentation.
- Recording these questions in an Excel template, optimizing the wording of the question via prompt optimization, so that they are tailored to be sent to an LLM rather than a human operator.
- Selecting and providing to the LLM a list of sources (reports and other documents) identified as relevant for the assessment.
- Writing a Python script, that goes through the Excel file, formulates the questions for each impact, communicates with the LLM, retrieves the answer and populates them in the respective cells in the Excel.
The final result is a populated Excel file with a sheet containing the list of all IROs, assessments of each factor for each of those IROs (score and rationale for the score) and references to the documents which were relevant in assessing them.
Use case key beneficiaries
☐ Relationship Managers
☐ Portfolio Managers
☒ Research teams, macroeconomists
☒ Control functions
☐ Support functions (HR, CFO, …)
☐ Other: Advisors, auditors
Benefits of AI use case for financial services sector
This Intelligent Process Automation requires human intervention in certain steps (as do almost all successful AI use cases) but allows for significant efficiency gains in the whole process, freeing up resources for other tasks.
Supporting technology
The solution is based on an LLM accessible through the Azure AI Foundry. The Python script extracts the questions
from the Excel file, sends them to the LLM, and records the responses in the file. The solution integrates a Retrieval Augmented Generation (RAG) component, i.e. the LLM navigates the files it had been provided by the user to generate answers. This has the effect of reducing hallucinations and ground responses in source content.
