We ran 2,000 historical files through an LLM-based bank statement parser against our human underwriters. It matched on NSF counts 97% of the time and flagged 3x more deposit anomalies. It's not replacing underwriters — it's replacing the first 40 minutes of every file.