Skip to content
Main Site News Console

Building AI for Lawyers — Legora CTO Jacob Lauritzen

· Latent Space Translated
播客深度访谈

Evals Without Traces, Million-Cell Spreadsheet Review, and Cracks in the Billable Hour

Basil Chatha

Nov 2 · 75 min read

In this episode, we sat down with Legora’s CTO, Jacob Lauritzen, just weeks after they announced a $550M Series E at an $8.8B valuation. We discussed what it takes to build AI agents for lawyers and what other agent builders can learn from their playbooks.

Legora started in a basement in Stockholm with a very simple product: ChatGPT, a Swedish legal database, and a search tool. Jacob, who had previously gone through YC with his own startup, helped Max Junestrand write the application for Leya (Legora’s predecessor). He subsequently joined as Head of Engineering and now serves as CTO.

Today, the agent can draft and edit Word documents, review thousands of contracts in bulk, and collaborate across lawyers, their clients, and other agents. In April of this year—less than 18 months after launch—the company crossed $100M in ARR, and currently serves over 1,800 customers. The company raised $550M at a $5.55B valuation in March, followed by an additional $550M at an $8.8B valuation in October. Its recent acquisition of Wexler marks the company’s fifth acquisition this year alone.

We asked Jacob: How do you run evals on agents when you have zero access to customer data? Which tasks still require more capable frontier models? How do legal engineers collaborate with forward deployed engineers? And why is the billable hour—the bedrock business model of law firms—being turned on its head?

Watch the full episode on YouTube →


What’s happening to the billable hour?

When Jacob first read Leya’s YC application, his feedback was blunt: people have tried LegalTech countless times, and it never works. If lawyers bill by the hour, efficiency just means fewer billable hours and less revenue. Why would they ever buy a tool like this?

He pointed out that what actually changed was the external pressure on law firms. In the US, routine work like low-end contract review is already shifting to flat project fees. Some firms will even do this work for free just to retain clients. Margins are shifting toward M&A, litigation, and advisory time spent directly with clients.

We asked whether this means law firms will no longer need paralegals and junior associates. Jacob noted that’s not what Legora is seeing. A great junior associate is more valuable today than they were two years ago because of the sheer volume of work they can accomplish with these tools—mirroring what is happening in software engineering.

Jump to the discussion on billing models →

Evals without customer data access

Most teams iterate on their agents by analyzing traces. But Legora can’t look at traces at all. The vast majority of materials uploaded by users are confidential, and largely belong to the law firms’ clients rather than the firms themselves.

As a result, the eval team has to meticulously recreate the real-world customer environment. A team of three engineers and three lawyers built RL simulation environments to mimic actual customer workflows—for example, creating a 10,000-document data room where an agent conducts due diligence and generates a report catching every single red flag. The team visualizes test results on a heatmap broken down by practice area and task type. Often, a new model might improve in one specific legal domain while degrading in others.

The team draws on a pool of roughly 100 legal engineers—most of whom spent three to four years at major law firms and often have a background in computer science. In production, Legora categorizes failures into 10 to 20 known failure modes, running dirt-cheap, specialized LLM checks in the background behind feature flags.

These strict security and compliance constraints also dictated Legora’s infrastructure choices. All their services run on Azure, because when Legora was founded, law firm clients wouldn’t accept AWS or GCP, whereas Microsoft had already secured vendor compliance approval through Microsoft 365.

Jump to the discussion on evals without traces →

Where smarter models are still indispensable

For a long time, Legora resisted offering a model picker, but eventually relented. Jacob’s take is that lawyers shouldn’t be forced to think about or choose between underlying model architectures