Building Ubers Redlining Agent
- Source
- uber.com
- Author
- Uber Engineering 2
- Date
Uber Engineering published an account of building a legal redlining agent, a tool that proposes edits to contracts, and placing it inside Microsoft Word where lawyers already work. The post covers how the team framed the problem, how it earned the trust of lawyers who are trained to distrust unreviewed changes, and how it improved the agent using their feedback. It closes with a retrospective on how the team would rebuild the system with the agent harnesses available today, which is a useful signal of how quickly the underlying tooling has changed. The broader lesson is about deployment rather than models. In high-stakes work, an agent is adopted when it fits the existing workflow and when experts can see and correct what it did. Teams planning agents for legal, finance or compliance work can use this as a reference for sequencing trust before automation.
Context
Uber's engineering team describes a first-generation Legal Redlining that learns from lawyers' corrections. The assistant works as an add-in inside Microsoft Word, examines counterparty edits, and proposes responses for a lawyer to accept, reject, or modify. Its engineering account says it saves the lawyer's decision, the original text and redline, comments, and any final modified wording, so later suggestions can draw on relevant past negotiations.
That stored feedback becomes for future model calls, with an exponential decay weighting that favors recent decisions. Lawyers also maintain a rules database describing Uber's positions and fallback language. Uber says this lets the tool adapt to the team's preferences without manual model . Lawyers still review each suggestion before it goes into the response to the counterparty.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
- fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
Checking sign-in…
Loading comments…