Transforming Financial Data Extraction With AI for a Global Professional Services Firm

The Challenge

Our client is one of the largest global professional services networks in the world. They operate in numerous industries and have a significant global presence.

They had specific challenges related to the extraction and analysis of information from a vast collection of 10-K and 8-K reports. The main need was to extract specific fields from documents in 10-K Exhibit 10, which are mostly legal agreements and contracts with different fields of interest. They used to perform this task manually, consuming substantial time and resources.

The first challenge was the lack of a dataset for document classification. Additionally, the initial interaction of the system lacked a test dataset, which was necessary for achieving high precision in the extracted fields. Finally, there was a need to design an efficient validation mechanism for the extracted data.

Phase 1: Classification

We utilized Azure OpenAI to generate a classification dataset, in which a series of documents were labeled according to their specific types. This dataset was then used to train the classification model, using embeddings and Random Forest classification. This algorithm categorizes new, unseen documents with a high confidence level.

Phase 3: UAT

In this stage we performed iterative User Acceptance Testing (UAT), testing our solution with its final users: the business team. Their deep business knowledge helped us determine whether the precision of the extraction was high enough or if it could be improved. We built a user interface (UI) embedded in Microsoft Teams that allows users to operate the extracted fields, manually editing them or even modifying the prompts to refine the extraction process. Once the required data is ready, it can be exported to a CSV file and made available for analyzing and sharing as needed.