Enterprise
Uber's Legal AI Agent Cuts Contract Review Time 20% Using RAG
The Legal Redlining Agent combines retrieval-augmented generation with deterministic rules and lawyer feedback to automate contract analysis while keeping attorneys in control.
Key takeaways
- Uber's Legal Redlining Agent reduced contract review time by over 20% with 91% accuracy while keeping lawyers in final approval roles.
- The system evolved from basic RAG to incorporate historical lawyer feedback, using similarity search to retrieve relevant past decisions with exponential decay weighting.
- A deterministic rules engine overrides AI recommendations when non-negotiable legal requirements apply, combining flexibility with hard constraints.
- The architecture uses LangGraph orchestration, OpenSearch for vectorized storage, and parallel workflow nodes for intent detection and risk assessment.
- Uber is exploring legal ontologies and knowledge graphs to provide semantic foundations for more consistent legal reasoning.
Uber has deployed an AI agent that reduces contract review time by more than 20 percent while maintaining 91 percent accuracy in its recommendations, according to the company. The Legal Redlining Agent operates as a Microsoft Word add-in that analyzes proposed contract changes, evaluates legal risks, and suggests policy-aligned responses—with lawyers retaining final approval authority.
The agent identifies edits from external parties, interprets their intent, compares changes against internal legal policies, and recommends whether to accept, reject, or modify specific clauses. It generates explanatory comments for each recommendation and flags changes requiring closer attorney scrutiny. Uber began developing the system in early 2024 and piloted it with its legal team in 2025.
Architecture combines multiple AI techniques
The system integrates retrieval-augmented generation (RAG), historical lawyer feedback, and deterministic rules. A React and TypeScript Word add-in connects to a Python backend orchestrated through LangGraph. Parallel workflow nodes handle intent detection, risk assessment, and policy lookup simultaneously.
OpenSearch stores vectorized legal rules and historical feedback, while Redis maintains session state for follow-up questions. The architecture separates concerns: legal teams define tone and style preferences, while engineers maintain prompt structure and application workflow.
Learning from lawyer decisions
Uber's initial RAG implementation retrieved relevant sections from legal playbooks but struggled with semantic retrieval accuracy, negotiation tone, and edge cases not covered by existing examples. The company added a feedback loop capturing lawyers' actual decisions, reasoning, and final edited text.
This shift enabled the agent to retrieve examples of how similar clauses were previously handled rather than relying solely on accept/reject signals. At runtime, similarity search and metadata filtering narrow thousands of historical interactions to approximately 20 candidates. A second language model-based filtering stage evaluates their relevance.
Exponential decay weighting—currently set to a 365-day half-life—prioritizes recent decisions while retaining older examples. This approach gives domain experts control over language and negotiation style as organizational preferences evolve.
Rules override model recommendations
A deterministic rules engine enforces non-negotiable legal requirements and overrides model recommendations when a matching rule applies. Language models handle tasks requiring contextual interpretation, including clause recommendations and negotiation tone. A final verification step checks generated responses against retrieved rules.
Why it matters
Uber's approach addresses a core tension in enterprise AI: automating repetitive work without removing human judgment from high-stakes decisions. By combining retrieval, learning from expert feedback, and hard-coded guardrails, the system demonstrates how organizations can deploy AI agents that augment rather than replace specialized knowledge workers. The 20 percent time reduction translates to meaningful productivity gains across thousands of contracts, while the 91 percent accuracy rate suggests the system has moved beyond proof-of-concept to production reliability.
Uber is exploring agentic harnesses including Claude Code and OpenCode, and plans to investigate legal ontologies and knowledge graphs to represent legal concepts and relationships for more consistent interpretation. The details were first reported by InfoQ.