Scaling AI in Legal: Building Uber’s Redlining Agent

This post details the development and iteration of Uber's Legal Redlining Agent (LRA). It describes the initial RAG approach and its limitations (inaccurate similarity, inappropriate tone, lack of generalization). It then elaborates on the iterative improvements, focusing on a feedback loop that captures lawyer decisions, expected actions, and final modified text, enabling the agent to learn user preferences and policy nuances. The post also highlights the refinement of the agent's tone through a three-part prompt anatomy managed by the legal team. The system utilizes Uber's GenAI Gateway and has achieved significant reductions in contract review time and high accuracy in AI-generated decisions.