AI Aulis

AI NEWS IN ENGLISH · 17 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

← Issue 17 November 2026Paper

Uber: legal team’s AI agent cut average contract review time by more than 20 per cent

AI-generated contentWritten by AI

infoq.com, techgig.com · 2 sources · Sources published: 16–17 November 2026

Uber says an AI agent that works in Word for its legal team has cut average contract review time by more than 20 per cent. The agent suggests responses to the other party’s changes, but lawyers make the decisions, and fixed rules override the language model’s suggestion when a rule applies to a clause.

The technology site InfoQ reported on the contract agent used by Uber’s legal team on 16 November, drawing on the company’s own blog post. TechGig repeated the details the next day in almost the same words. The figures are therefore Uber’s own, and neither publication has verified them.

The Legal Redlining Agent runs as a Microsoft Word add-in. It identifies changes proposed by the other party to a contract, interprets their intent and compares them with the positions of Uber’s legal department. It then uses a comment to recommend, with reasons, whether to accept, reject or modify the clause. It flags changes that need closer review separately. Lawyers review and approve the recommendations.

Uber says average review time has fallen by more than 20 per cent, with 91 per cent accuracy in AI-generated decisions. Neither InfoQ nor TechGig gives the baseline for review time or explains how accuracy was measured.

Lawyers’ past decisions guide the recommendations

Uber began development in early 2024 and piloted the agent with its legal team in 2025. The first version retrieved relevant passages from the legal department’s negotiation guidelines and gave them to a language model as the basis for its answer (retrieval-augmented generation, RAG). InfoQ reported that the system struggled with retrieval relevance, negotiating tone and situations for which the guidelines had no example.

Uber later added a feedback loop that records lawyers’ decisions, their reasons and the final edited text. This gives the agent not just data on whether past recommendations were accepted but also examples of how similar clauses were handled.

Similarity search and metadata filtering narrow thousands of past cases to about 20 candidates, and a language model then assesses their relevance. The system favours recent decisions through exponential decay weighting, with a configurable half-life currently set to 365 days. A year-old example therefore gets half the weight of a new one, but older examples stay in use.

The legal team defines the target tone and style. Software engineers maintain the prompt structure and the application workflow.

Fixed rules override the language model

A deterministic rules engine, which always gives the same output for the same input, enforces non-negotiable legal requirements. If one of its rules applies to a clause, the rule overrides the model’s recommendation. Language models handle tasks that require interpreting context, such as clause suggestions and negotiating tone. Finally, the system checks the responses against the retrieved rules.

The Word add-in is built with React and TypeScript, and LangGraph orchestrates the steps of the Python backend. Parallel steps identify the intent of a change, assess risk and retrieve the relevant positions. OpenSearch stores rules and feedback for retrieval. Redis keeps session data in memory for follow-up questions.

Uber is also exploring harnesses for its AI agents, including Claude Code and OpenCode. A harness is a program that controls an agent, gives the model its tools and runs its steps. The agent’s development team also plans to investigate whether knowledge graphs could represent legal concepts and the relationships between them.

Report an error in this story