OpenAI Turns Contracts into Searchable Data

As contract review work surged, OpenAI faced the problem that existing manual methods were no longer scalable. To solve this, the finance and engineering teams directly built a contract data agent.

This agent operates through the following 3-step process:

• Ingest data: Various file formats, including PDFs, scans, and photos containing handwritten notes, are unified into a single pipeline.
• Inference with prompting: Using the RAG (Retrieval-Augmented Generation) approach, contracts are parsed into structured data. Rather than simply extracting text, the agent finds and reasons over relevant content and provides justification.
• Review: Finance experts review the structured outputs and annotations generated by the agent. The agent identifies non-standard clauses and provides justification to assist expert judgment.

As a result of introducing this system, review time was cut in half, and thousands of contracts can now be processed without a linear increase in headcount. Another advantage is that the outputs are stored as tables in a data warehouse, making them easy to query.

This architecture is expanding beyond contracts to various financial operations such as procurement, compliance, and month-end closing. This demonstrates a new operating model in which AI does not replace human decisions, but rather automates repetitive tasks so that experts can focus on judgment.