Legora’s agent stack for lawyers automates contract review at scale with no customer data access
Legora’s agent system drafts contracts, reviews thousands of documents at once, and coordinates work between lawyers and clients—all while keeping customer data confidential and running exclusively on Azure.
1 source · cross-referenced
- Legora’s agent automates drafting, edits Word documents, and reviews thousands of contracts simultaneously for legal teams.
- The company evaluates agents without access to customer data, simulating legal work via custom RL environments and hiring legal engineers with law and CS backgrounds.
- Legora classifies agent failures into 10–20 known modes and runs cheap LLM checks behind feature flags, with all infrastructure on Azure due to enterprise customer restrictions.
- The agent stack relies on four primitives—projects, agents, Tabular Review, and lists—and has scaled to 100 engineers, with legal engineers configuring products and forward deployed engineers handling integrations.
- Legora estimates models already handle ~60% of legal work well, but complex million-token analysis still needs more intelligence; the company post-trains open-weight models for narrow legal tasks.
Legora’s agent automates drafting and editing of Word documents, reviews thousands of contracts in parallel, and coordinates work between lawyers, clients, and other agents, according to an interview with Jacob Lauritzen, CTO of Legora, published by Latent Space.
The company evaluates its agents without access to customer data, which is often confidential and belongs to clients rather than the firms using the tool. Instead of reading traces, Legora’s evals team—comprising three engineers and three lawyers—recreates the kinds of work customers do using RL environments. One example is a 10,000-file data room where the agent performs due diligence and produces a report identifying red flags, with results plotted on a heat map by practice area and task.
Legora classifies agent failures into 10 to 20 known failure modes and runs cheap LLM-based checks behind feature flags in production. All infrastructure runs on Azure because law firms initially required an approved vendor through Microsoft 365, and Legora retained the platform even as enterprise acceptance broadened.
Legora’s agent stack is built around four primitives: projects that hold files, the agent itself, Tabular Review, and lists. Tabular Review allows users to drop in up to a million cells of data—starting with 10 rows to refine prompts before scaling—and ask questions like identifying termination periods or IP assignment clauses across employment contracts. Legal engineers configure new products on these primitives without writing code, and even Wexler’s fact intelligence product runs on the same building blocks.
The company has scaled from 20 to 100 engineers since Lauritzen joined, aiming to ship a few core components that customers perceive as a comprehensive suite. Legal engineers—lawyers with three to four years in big law and often a CS degree—configure products like power users, while forward deployed engineers write production code, integrate with CRMs or document management systems, and feed learnings back into the core product without touching client data.
Legora estimates that models already handle about 60% of legal work effectively, shifting the bottleneck to cost and latency for those tasks. For complex analysis and editing across million-token contracts, more intelligence is still needed. The company has started post-training open-weight models for narrow tasks like legal research and has no plans to train its own frontier model. For drafting, Lauritzen notes that Fable and Opus are too verbose compared to Astra, and the agent mitigates hallucinations by attaching citations that highlight bounding boxes in source PDFs, multiple layers deep.
- Nov 11, 2026 · Latent Space — swyx
OpenAI demonstrates voice steering, cloud handoffs, and personal agents for Codex at AI Engineer Code Summit
Trust79 - Nov 11, 2026 · The Verge — AI
OpenAI to launch verified AI Business Agents in ChatGPT for US e-commerce and service businesses
Trust79 - Nov 10, 2026 · Latent Space — swyx
NaiveAI releases Naive-N0.5-Flash coding model with ultrafast token throughput
Trust74