Uber Legal Redlining Agent cuts review time over 20%

Uber dates the first generation of its Legal Redlining Agent to work that began in early 2024 and continued through 2025, before today's agent harnesses were widely available. The company frames the problem as a steady stream of contract negotiations containing substantial repeatable redline work — thousands of contracts a year — with delays rippling across sales, onboarding, and launches. Its answer was to automate the repeatable parts without replacing a lawyer's judgment, and to make the agent learn from the lawyers themselves.

The system went through four iterations. The first was RAG (retrieval-augmented generation, where the model looks up relevant documents before answering), which stumbled on semantic similarity, tone, and generalization. The second added a self-learning feedback loop that captured each lawyer's decision, expected action, comments, and final modified text. The third tackled tone with an extra model call, and the breakthrough came when the in-house Legal team, not engineering, owned the specific language defining voice and style. The fourth moved beyond accept-or-reject calls into drafting counter-proposals, supported by a rules database tied to the original contract templates.

A few design choices recur. A thin Word add-in handles document interaction while a back end orchestrates the AI, and a hybrid engine splits non-negotiable policies — handled by a deterministic rules engine that can override the model — from negotiable nuances handled by the probabilistic feedback loop. OpenSearch holds two indices: rules vectorized by target sentence, and feedback vectorized by the counterparty's intent. Over time Uber weights older interactions less, using a configurable half-life currently set to 365 days, so the system tracks the team's current positions rather than stale precedents. Looking ahead, Uber says it is exploring agentic harnesses such as Claude Code and OpenCode and, separately, legal ontologies and knowledge graphs.