Hybrid implementation
outperforms LLM-first approach
Challenge
This leading asset manager processes large volumes of financial documents, including fund-of-fund notices and vendor invoices, that trigger payments on behalf of investment funds. The firm wanted to automate the extraction of invoice and payment data and feed it directly into its workflow platform. But because these documents often contain complex tables, the solution needed to deliver highly accurate data extraction. They couldn’t risk mistakes in payment; the data must be absolutely correct.
Capturing this information with precision initially proved difficult. One previous automation effort ultimately required more than 20,000 hours of manual document review. The firm explored building its own document processing capability based on emerging generative AI models, which proved challenging to control with reasonable cost and timeframe. Ultimately, they decided to evaluate purpose-built solutions, specifically those that could meet the reliability, consistency, and accuracy requirements of financial workflows.
