Uber says an internal AI assistant has reduced average contract-review time by more than 20% while reaching 91% accuracy on its generated decisions. The Legal Redlining Agent works inside Microsoft Word, identifies edits from outside parties and recommends whether lawyers should accept, reject or revise each clause.
The first version retrieved relevant passages from legal playbooks, but Uber found that it struggled with semantic matches, negotiation tone and cases not covered by examples. The company added a feedback loop that stores lawyers’ decisions, reasoning and final language. At runtime, search and metadata filters narrow thousands of earlier interactions to about 20 candidates, followed by another model-based relevance check. More recent decisions receive greater weight.
A deterministic rules engine handles non-negotiable requirements and can override a model recommendation. Language models interpret context and draft negotiation language, while a final check compares the response with retrieved rules. Lawyers still review and approve the output, so the reported speed gain does not represent autonomous legal decision-making.
Uber’s figures are company-reported rather than independently audited. The architecture nevertheless shows a practical division of labor: retrieval supplies precedent, models handle ambiguous language, and fixed software rules reserve control over policies that must not be negotiated away.