Teamday
← Back

Legora CTO Jacob Lauritzen: Legal Agents Without the Data

Latent Space
agentsenterpriseknowledge-workinterviewbusiness

Jacob Lauritzen is CTO of Legora, the Stockholm-founded legal AI company (YC 2024, originally called Leya) that now sells to large and small law firms and in-house legal teams in Europe and the US. On Latent Space he walks through how the product grew from a ChatGPT-style assistant with a Swedish legal research tool into an agent that drafts, edits Word files, organizes documents and runs large reviews, and what it takes to build that for an industry where engineers can't look at the work.

The constraint that shapes everything: no traces. Law firm data mostly belongs to their clients, so Legora's engineers can't inspect what users do with the agent. "We can't look at any traces basically. We can't look at any data because it's all super confidential." Instead, an evals team of roughly three engineers and three lawyers builds environments that approximate real projects, for example a 10,000-file data room with the red flags a good due diligence report should catch, and scores drafting and analysis separately. In production they run cheap online LLM checks on whether a task was solved, classify failures into a known set of 10 to 20 modes, and compare feature-flagged rollouts. When a client cares about a failure, Legora asks for anonymized documents so it can reproduce it.

Why they mostly didn't do evals at first: For a long time every model release simply got better, so building the harness beat measuring it. Citations were one example. "It went from like 70% to like 99%. It's just like we did nothing and like, great." Today they track every model across legal practice areas and tasks, and Lauritzen says releases that improve some areas while regressing others happen "a lot more than you'd think."

Tabular Review, the spreadsheet form factor: Lawyers drop thousands of contracts into a grid, one document per row and one question per column (termination period, IP assignment, severance), and the agent fills in each cell with a citation back to the source. Lauritzen says it scales to about a million cells, runs the first 10 rows first so users can iterate on their questions, and is built on AG Grid. The agent can also start a review itself, wait for it to finish, flag red flags and draft the client memo. Citations chain down through layers, so a sentence in a memo can link to a table cell and from there to a bounding box in the PDF.

Legal engineers versus forward deployed engineers: Legora's hardest hire is a lawyer with three or four years in big law and often a CS degree. About a hundred of them configure the product for clients, shape the roadmap and write evals. Its forward deployed engineers, the first hired roughly nine months ago in New York, handle what configuration can't. "Legal engineers configure the product as a power user would. They don't write code." Both build on a small set of primitives: Lists (a Linear or Notion-style database with triggers that can start agent runs), projects, the agent and Tabular Review. Intake, Legora's Slack bot, files incoming legal requests onto a list where lawyers and agents work them. Fact intelligence from Wexler, the UK litigation startup Legora acquired, is built on the same primitive.

From 20 engineers to 100: "We're just going to be 20 engineers for the rest of the company's history. And like, we're going to beat everyone who's much bigger." That was his pitch six months in. Legora now has about 100 engineers and he says it feels too small, so his job is to build primitives that let the team ship many products while building only a few things. He is also wondering how much of the product, including its API, could simply be the agent.

6 Lessons from Legora on Building Vertical AI Agents

  • Build evals from expert approximations when you can't see production - Lawyers on the evals team turn known workflows into synthetic environments; failure modes are classified, never read in full.
  • Post-train only bounded tasks - Legora hasn't fine-tuned a main agent. It is now post-training open-weight models for narrow jobs like fan-out legal research, where a small model beats a frontier one on cost and latency.
  • Keep the vendor list short - All infrastructure runs on Azure because law firms had already approved Microsoft. ElevenLabs is excluded for data processing agreement reasons; transcription runs on Gemini and Microsoft's newer model.
  • Model pickers are a concession - Legora resisted letting lawyers pick models, then gave in. Lauritzen thinks about 60% of legal work is already handled well and is now a cost and latency problem.
  • Long documents remain the frontier - Editing a very long contract while keeping definitions and ordering consistent is still hard, and Legora keeps rebuilding its Word add-in scaffolding, from map-reduce to fuller context, as models change.
  • Billing is shifting - US firms already price some lower-end work as flat project fees or include it free, while lawyers spend more time on strategic M&A and litigation work.

What Legora Shows About AI Agents in Regulated Knowledge Work

Legora's advantage isn't a better model. It is domain staff who can stand in for data it will never see, plus a few flexible primitives those people can configure without code. Lauritzen expects junior lawyers who master these tools to become more valuable, not obsolete, and law firms to codify their own way of doing things into systems they own. Any company deploying agents into confidential work faces the same blind spot and can borrow the same answer: expert-built evals, cited outputs and narrow post-training.