Uber Says Its Legal AI Agent Cuts Review Time Over 20%

Uber Says Its Legal AI Agent Cuts Review Time Over 20%

Estimated reading time: 7 minutes · Last updated:

Uber says its Legal Redlining Agent, a legal AI agent that runs inside Microsoft Word, has cut average contract review time by more than 20%, and that 91% of its AI-generated decisions are accurate. The tool reads the changes a counterparty has made to a contract, weighs their risk against Uber's legal policies, and proposes whether to accept, reject or modify each clause. It does this by blending retrieval-augmented generation, a record of how Uber's lawyers handled past clauses, and a deterministic rules engine that can overrule the model. Lawyers remain responsible for reviewing and approving what it suggests. Uber has not said how it measured either headline figure, which matters to any legal team hoping to copy the design.

How can AI help streamline repetitive contract review while keeping lawyers in control?

Uber Engineering, on LinkedIn

Key takeaways

  • The claimed result: Uber reports that its Legal Redlining Agent has reduced average contract review time by more than 20%, with 91% accuracy in AI-generated decisions.
  • Lawyers stay in charge: The agent proposes whether to accept, reject or modify a counterparty's edit inside Microsoft Word, but Uber's lawyers review and approve every recommendation.
  • Precedent over playbooks: Uber added a feedback loop recording lawyers' decisions, reasoning and final text, and now narrows thousands of past interactions to roughly 20 candidates per request.
  • Rules beat the model: A deterministic rules engine enforces Uber's non-negotiable legal requirements and overrides the language model whenever a matching rule applies.

What Uber's Redlining Agent Does Inside Microsoft Word

Redlining is the back-and-forth of marking up a contract draft, with each side proposing edits for the other to accept or contest. Uber's agent sits in Microsoft Word as an add-in and works on the edits an outside party has introduced. For each change it works out what the counterparty is trying to achieve, checks the new wording against Uber's legal policies, and suggests one of three moves: accept the clause, reject it, or modify it.

It also drafts a comment explaining its reasoning, so the reviewing lawyer sees why as well as what. Changes that deserve a harder look are flagged rather than quietly resolved. Uber Engineering framed the goal on LinkedIn as streamlining repetitive contract work while lawyers stay in control, and the approval decision remains with the legal team.

The project has been running for some time. Uber started on the first generation in early 2024 and piloted it with its Legal team in 2025. The results it now reports, more than 20% off average review time and 91% decision accuracy, are the company's own figures, and Uber has not published the method behind them.

Why Retrieving the Playbook Was Not Enough

The first version leaned on a standard RAG pattern. Retrieval-augmented generation is a technique in which a system fetches relevant documents and passes them to a language model as context before it answers. In Uber's case, those documents were sections of its legal playbooks.

By Uber's account, that approach ran into three problems. Semantic retrieval did not reliably surface the right passages. The model's negotiation tone missed the mark. And the system had little to go on when a clause fell outside the examples already on file.

Uber's fix was to feed its lawyers' work back into the system. A feedback loop now captures what a lawyer decided, the reasoning, and the text they finally settled on. That is a richer signal than a yes or no on a past recommendation: the agent can retrieve how a comparable clause was actually handled and base its suggestion on that outcome. The change also gives domain experts control over the wording and negotiating style the agent adopts.

How the Agent Narrows Thousands of Precedents to About 20

Every request starts with a large pool. Uber says the system holds thousands of historical interactions, and at runtime it combines similarity search with metadata filtering to shrink that pool to roughly 20 candidates. Similarity search ranks stored items by how close their meaning is to the clause under review, while metadata filtering screens out items that fail set criteria.

A second stage then hands those 20 or so candidates to a language model, which judges whether each is genuinely relevant. The two-step design targets the retrieval accuracy problem that held back the first version: vector search casts the net, and the model does the finer sorting.

Age counts as well. Uber applies exponential decay weighting, meaning a precedent's influence shrinks steadily as it ages without ever reaching zero. The half-life is configurable and currently stands at 365 days, so a decision from a year ago carries about half the weight of one made today, and one from two years ago roughly a quarter. Recent practice dominates, but older examples stay in play.

Where Hard Rules Overrule the Language Model

Uber does not give the model the last word. A deterministic rules engine encodes requirements the company will not negotiate, and when one of those rules matches a clause, it takes precedence over whatever the model recommended. Language models are reserved for work that needs contextual judgment, such as drafting clause recommendations and setting the tone of a response.

A final check closes the loop. Before a response reaches the lawyer, the system verifies it against the rules it retrieved, a guard against output that drifts from policy.

Ownership is divided along similar lines. Uber's legal team sets the tone and style it wants, while engineers own the prompt structure and the application workflow. Underneath, a Word add-in written in React and TypeScript talks to a Python backend orchestrated with LangGraph, where intent detection, risk assessment and policy lookup run as parallel nodes. OpenSearch holds the vectorized legal rules and the historical feedback, and Redis keeps session state so a lawyer can ask follow-up questions.

Components of Uber's Legal Redlining Agent
Layer Technology Role
Front end React and TypeScript Word add-in Works on counterparty edits inside Microsoft Word
Orchestration Python backend with LangGraph Runs intent detection, risk assessment and policy lookup in parallel
Retrieval store OpenSearch Holds vectorized legal rules and historical lawyer feedback
Session memory Redis Keeps state for follow-up questions
Guardrail Deterministic rules engine Enforces non-negotiable requirements and overrides the model
Interpretation Language models Clause recommendations, negotiation tone, relevance filtering

What could help or hold back the approach

The case for

  • Lawyer feedback compounds: every decision, reasoning note and final edit the loop records adds to the pool of precedent the agent can retrieve.
  • The rules engine and final verification step give Uber a hard limit on what the model can recommend for non-negotiable terms.
  • Splitting ownership lets the legal team adjust tone and style without engineers rewriting prompt structure.

The case against

  • Uber's 91% accuracy figure leaves about 9% of AI-generated decisions that a lawyer must catch.
  • Neither the 20% time saving nor the accuracy figure comes with a published measurement method.
  • A 365-day half-life discounts older precedent that may still reflect valid policy, unless the setting is tuned.

What to be careful about

  • Uber has not disclosed how it measured the more-than-20% reduction or the 91% accuracy, so other teams cannot compare like with like.
  • Roughly 9% of AI-generated decisions fall outside the reported accuracy rate, and catching them depends on lawyers reviewing closely rather than approving by habit.
  • The rules engine only protects terms someone has encoded; a non-negotiable requirement missing from the rules falls back to model judgment.
  • Precedent retrieval reflects past lawyer choices, so inconsistent historical decisions could feed inconsistent recommendations.

The bottom line

Uber's headline numbers are self-reported, and the 91% accuracy figure still leaves about 9% of AI-generated decisions for a lawyer to catch. The design is the more portable lesson. Uber found that retrieving policy text was not enough, and that the stronger signal was how its own lawyers had actually resolved similar clauses, weighted toward recent practice and fenced in by rules the model cannot override. The company is not treating this version as finished. Its engineers are exploring agentic harnesses including Claude Code and OpenCode, along with skills that structure and automate workflows across Uber's agentic AI work, and plan to study legal ontologies and knowledge graphs to give the agent a semantic map of legal concepts and their relationships.

What to watch

  • Watch whether Uber publishes how it measured the more-than-20% cut in review time and the 91% accuracy rate; no date has been set.
  • Watch for results from Uber's planned work on legal ontologies and knowledge graphs, which it says should give the agent a more consistent basis for interpreting legal concepts; no timeline has been given.
  • Watch whether the agentic harnesses Uber is exploring, including Claude Code and OpenCode, end up in the redlining workflow; Uber has announced no schedule.

Frequently asked questions

No. The Legal Redlining Agent recommends whether to accept, reject or modify a clause and drafts an explanatory comment, but Uber's lawyers remain responsible for reviewing and approving its recommendations.

What does a 365-day half-life mean for Uber's contract precedents?

Uber weights past lawyer decisions with exponential decay, and the half-life is currently set to 365 days. A decision made a year ago counts for about half as much as a fresh one, and one from two years ago roughly a quarter, so recent practice leads without older examples disappearing.

How long did Uber take to build its redlining agent?

Uber began developing the first generation in early 2024 and piloted it with its Legal team in 2025. It reworked the original playbook-only retrieval along the way, adding a feedback loop built on lawyers' decisions and final edited text.



Share:

Categories

Newest course every month

Advertise your offline course to a wider audience with our landing page.

You May Also Like

Uber says its legal AI agent, a Word add-in built on rules and RAG, cuts average contract review time by...
  • November 16, 2026
UAE AI liability for harmful AI advice falls mainly on the company using it, lawyer Ashish Mehta argues, citing the...
  • November 15, 2026
AI privilege waiver is not automatic after Heppner, Norton Rose Fulbright lawyers argue, citing Morgan, Assini and a Texas Business...
  • November 13, 2026