Uber says its Legal Redlining Agent has cut average contract review time by more than 20%, with 91% accuracy in AI-generated decisions. The tool works inside Microsoft Word, where it identifies changes proposed by outside parties, assesses legal risks and recommends whether to accept, reject or revise contract clauses.
The agent also drafts comments explaining its suggestions and flags changes that need closer attention. Uber lawyers review and approve every recommendation; the system does not make the final decision. Uber has not explained how it measures the reported accuracy figure, so it is not clear what kinds of decisions or errors that number covers.
Why the first version needed more than playbooks
Uber began building the tool in early 2024 and piloted it with its Legal team in 2025. Its first version used retrieval-augmented generation, or RAG: a language model looks up relevant material—in this case, legal playbooks—to help answer a question. Uber says that approach struggled to find the right examples, match the tone of negotiations and handle situations not covered by existing guidance.
The company then added a feedback loop that stores lawyers’ decisions, reasoning and final edited wording. Uber says its store contains thousands of past interactions. When a new clause comes in, the system uses a search and filters to narrow the examples to about 20 candidates, then uses another language model to check their relevance. A weighting system favors newer decisions, with its half-life set to 365 days, while retaining older examples.
Rules keep some decisions out of the model’s hands
Uber combines those examples with a rules engine that enforces non-negotiable legal requirements and can override the model’s recommendation. The system also checks its drafted response against retrieved rules. In Uber’s design, models handle context-sensitive tasks such as interpreting intent and matching negotiation style, while fixed rules protect requirements that should not be negotiated.
The company says lawyers can see the rules and past feedback behind a suggestion, and can change the wording or decision before responding to a counterparty. Uber is also exploring legal knowledge graphs—structured maps of legal concepts—and tools for organizing AI workflows. It has not said when those explorations might lead to changes in the agent.
Why it matters
For Uber lawyers, the tool may reduce time spent on repeatable contract edits and leave more time for complex judgment calls; Uber reports a 20% time reduction, but does not explain how it measured accuracy. For companies considering similar tools, Uber’s approach keeps lawyers responsible for approving recommendations, though the material does not establish whether the results would transfer to other legal teams.