Building an AI to Read Bank Contracts Like Bankers Do

The author built an AI system that reads bank contracts the way bankers do, rather than how customers typically read them. The system uses specialized agents to extract clauses, scan for risks, analyze cross-contract impacts, and detect contradictions.

Details

The author, a former banker, noticed an information asymmetry where banks used internal scoring grids and LLMs to evaluate customer data before any human review, while customers still signed contracts without fully understanding them. To address this, the author built an AI system that reads bank contracts the way bankers do, rather than how customers typically read them linearly. The system uses four specialized agents running in parallel: a Clause Extractor that maps the document structure, a Risk Scanner that scores clauses against a library of known adverse patterns, a Cross-Contract Analyzer that checks clauses against the borrower's other agreements, and a Contradiction Detector that surfaces discrepancies between the contract and the borrower's understanding. This approach, inspired by the 4D analytical framework, is more reliable than a single general-purpose model. The system was able to identify issues like margin ratchet clauses, cross-default links to unrelated contracts, callable provisions triggered by unmonitored ratios, and termination asymmetries that customers were unaware of. The author notes a technical limitation where LLMs can hallucinate on numerical conditions without properly resolving defined terms, which requires explicit mapping of defined terms before interpretation.