Uber's Legal Redlining Agent cuts review time over 20%

Uber says it started building the first generation of the Legal Redlining Agent in early 2024 and continued through 2025, before today's agent harnesses were widely available. The team began with the business problem rather than the model: a steady stream of contract negotiations involving substantial repeatable redline work. That framing shaped a key design choice, building the assistant as a Microsoft Word add-in rather than a standalone web application, on the reasoning that asking lawyers to copy-paste between tools would kill adoption.

The first attempt was a RAG system built on Legal team playbooks and negotiation examples. Uber reports it ran into three problems: imprecise semantic similarity, tone that swung between overly defensive and excessively positive, and effectively random responses on scenarios the playbooks did not cover. That pushed the team toward a self-learning loop capturing the lawyer's decision, expected action, comments and final modified text. Uber describes this granular data as a turning point, letting the system map a counterparty's intent to the lawyer's precise language. Later iterations added a dedicated tone-modulation step and an agentic workflow for drafting counter-proposals, plus a rules database for frequently modified template language.

Under the hood, Uber describes a thin-client architecture: the Word add-in handles document interaction, while a Python back end orchestrates the AI work, using OpenSearch vector stores as the memory layer and Redis for session state. A hybrid decision engine separates non-negotiable policies, handled by a deterministic rules engine, from negotiable nuances handled by the probabilistic feedback loop. Uber says it is now exploring agentic harnesses such as Claude Code and OpenCode, along with legal ontologies and knowledge graphs.