Uber’s internal Legal Redlining Agent reviews proposed contract edits in Microsoft Word and recommends whether to accept, reject or modify a clause. Uber says the tool cut average review time by more than 20% and that its AI-generated decisions reached 91% accuracy; Uber’s lawyers review the suggestions and decide what to use. Uber’s engineering team describes the system and its reported results.
For contract review, automate the repeatable first pass while keeping people responsible for the decision.
How the agent uses past decisions and fixed rules
The agent combines retrieval-augmented generation (RAG)—a way of bringing relevant records into an AI system’s response—with examples from earlier negotiations and rules set by Uber Legal. It examines a proposed edit, assesses its risk and can suggest a response with an explanatory comment.
Uber says its feedback records include the contract context, the other party’s proposed edits, the lawyer’s decision and reasoning, and the final edited text. That gives the system more to draw on than a simple tally of accepted or rejected suggestions. Similarity searches and metadata filters narrow thousands of past interactions to about 20 candidates; a language model then filters those for relevance. Uber describes a configurable weighting system with a 365-day half-life, which reduces the influence of older feedback over time.
Fixed rules handle requirements Uber treats as non-negotiable: a matching rule can override a model recommendation. The feedback-based approach is used for matters that allow judgment, such as negotiation tone and strategy. The two mechanisms give the system different roles: rules enforce defined boundaries, while prior decisions inform suggestions where the response can vary.
Lawyers retain the decision
The agent proposes actions; Uber’s lawyers review, approve, reject or revise them before a response is incorporated into the contract negotiation. The system is designed to assist with repeatable, high-volume work, leaving legal judgment with the lawyers rather than making the final contract decision itself.
That distinction matters because a recommendation is not the same as an approved change. The agent can flag a clause for closer attention and draft a comment, but a lawyer determines whether the proposed response fits the negotiation.
What Uber reports about performance
Uber reports a reduction of more than 20% in average contract-review time and 91% accuracy in AI-generated decisions. Those are company-reported results for its internal tool; the figures do not establish outcomes across other legal teams or contract settings.
Uber says development of the first generation began in early 2024, followed by a pilot with Uber Legal in 2025. The company describes four major development iterations, including changes after the first approach struggled with matching the meaning of clauses, producing suitable negotiation tone and handling situations absent from its examples.
Building the tool into Word
Working inside Microsoft Word puts the agent in the document workflow, but Uber also describes engineering problems in the Office JavaScript API used for that integration. Some operations were slow, some tracked-change document structures caused crashes, and deleted tracked text could return as an empty string.
Uber says it used caching, defensive loading and separate text-state handling to address those issues. The add-in uses React and TypeScript; a Python backend orchestrates the workflow through LangGraph, with OpenSearch for retrieval and Redis for session state.