Delivery and mobility / Ride-hailing and delivery
Uber: Reviewing contract redlines and drafting responses for the legal team
- Company
- Uber
- Country
- United States
- Adoption stage
- In operation
- Source published
- Date basis
- The date the source was published. It can differ from the date adoption started.
- How the source was checked
- Read the full source text
The work problem
According to Uber's official engineering blog, Uber negotiates thousands of contracts every year, and its legal team reviews and responds to each set of edits (redlines) made by the other party. Uber explained that this process is time-consuming and repetitive and can easily become a bottleneck that delays deals, client onboarding and launches. At the same time, it saw similar clauses, edits that recur across clients and consistent patterns of legal reasoning, so it judged that the repeatable parts could be automated to give lawyers more time for complex judgments.
Technology and data
Uber built the first generation of its Legal Redlining Agent (LRA) from early 2024 through 2025 and piloted it with its legal team in 2025, folding their feedback back into the product. LRA runs as an add-in inside Microsoft Word, where lawyers already work. It finds the changes made by the other party in a document, interprets their intent, recommends accepting, rejecting or modifying each change based on the legal team's policies and guidelines, and generates a comment justifying the recommendation for the other party along with a risk level. Lawyers review each suggestion one by one and accept, reject or modify it before it goes into the response. The team first used a RAG approach that retrieved negotiation playbooks and fed them to an LLM, but retrieval was imprecise and the tone was wrong, so it moved to a feedback structure that collects and reuses lawyers' decisions, expected actions, comments and final modified text. At runtime, the agent selects about 20 documents from thousands of past records using metadata filters, filters them once more with an LLM to use as the basis for its decision, and gives more weight to recent decisions. In-house lawyers manage the prompt that adjusts tone themselves, and when a rules database, in which lawyers record Uber's position and fallback positions for template sentences, returns a matching rule, a final LLM step checks that the output follows that rule. The back end uses LangGraph to run intent detection, risk assessment and policy lookup in parallel, stores rules and past feedback in OpenSearch, and continuously checks retrieval quality and rule adherence with LLM-based scoring evaluations.
Results
According to Uber's official engineering blog, since LRA was launched with the legal team, average contract review time has fallen by more than 20%, and AI-generated decisions have been 91% accurate. The blog also quoted a lawyer who said the tool saved real time, especially with its comment suggestions. LRA was also part of the submission with which Uber's legal organization won Most Innovative Legal Department of the Year at the ALM Legalweek Leaders in Tech Law Awards on March 9, 2026.
Limits and open questions
The blog does not disclose the number of contracts processed, the measurement period or how the 91% accuracy was measured. Expansion to other lines of business is described as a future plan.
Sources
- Scaling AI in Legal: Building Uber's Redlining Agentuber.com, Accessed
Compiled from public sources. These are not results from ATF Works customers.