Latent Space

Jacob Lauritzen on building a legal agent that never sees your data

Jacob Lauritzen· CTO at Legora
·~67 min·English·Latent Space
AgentsLLMAI CompanyBusiness Strategy
TL;DR

Legora builds its legal agent blind, never seeing a client file, so it tests in simulated worlds built by lawyer-engineers and anchors every answer to a clickable citation.

01The Core Constraint

The Legal Agent They Can't Watch Work

Legora runs an agent across thousands of legal tasks but can never look at the traces or the data, because it all belongs to the client's clients.

we can't look at any traces basically. We can't look at any data cuz it's all super confidential and that's like extremely frustrating to work with

— Jacob Lauritzen, Latent Space
Key Insight
Unlike a coding-agent company that improves on the traces its users leave behind, Legora's moat and its handicap are the same fact: the data is so valuable precisely because nobody, including Legora, can look at it. Every other design choice in the interview is downstream of this one blindness.
02How They Test

To Test Without Data, Build the World

Since Legora mostly can't observe real matters, its eval team builds simulated worlds, RL-style environments like a seeded 10,000-file data room with known red flags, and runs its evals inside them.

we have different basically like worlds like RL environments essentially

— Jacob Lauritzen, Latent Space
Key Insight
Because they mostly can't observe reality, they have to manufacture it, and the ceiling on product quality is how well their invented worlds match the real matters they rarely get to see. A high score in the sim is a claim about the sim, not proof the customer is being served.
03The Team

The Rarest Hire Is the Lawyer-Engineer

The eval team is three engineers and three lawyers, backed by a pool of legal engineers, big-law lawyers with CS degrees who are the hardest hire and a big force multiplier.

it's a mix. We have uh engineers and then we have lawyers and it's one team. So there's like three engineers and like three lawyers right now.

— Jacob Lauritzen, Latent Space
Key Insight
The scarce resource isn't engineers or lawyers, it's the person who is credibly both. In a domain where you can't see the data, the people who can tell whether an answer is actually good are the ones who have done the job. The legal engineer is how Legora smuggles that judgment back into a blind system.
04Riding the Frontier

A Model Release Fixed It, We Did Nothing

For a long time Legora skipped heavy evals because every few months a new model just fixed things, one citation bug went from 70% to 99% on a release, but the same releases quietly regress other tasks.

next model release solved it went from like 70% to like 99 like we did nothing and like great

— Jacob Lauritzen, Latent Space
Key Insight
Betting on everything improves works right up to the release that doesn't. The eval heat-map across practice areas exists not to make models better but to catch the regressions the free upgrades hide, so a silent drop on one task doesn't ship behind a headline gain on another.
05Trust Layer

Every Answer Has to Point at Its Source

Legora wraps the model in a citation system, every sentence links to a bounding box in the source, and you can drill down layer by layer so the human reviews the links instead of trusting the reasoning.

we have a system of a citation system where we basically check what they're writing has a link

— Jacob Lauritzen, Latent Space
Key Insight
Citations turn an unverifiable claim into a cheap-to-check one. The agent can't be trusted to be right, but it can be forced to show its work, so the human audits the links, not the logic. That is how a model that still produces hallucinated citations becomes usable in a field where getting a cite wrong is a serious problem.
06The Product Primitive

A Spreadsheet That Scales to a Million Cells

Tabular review is fan-out as a product, one row per document, one column per question, scaling to a million cells, and you test the question on the first ten rows before you let it run.

Each row is one employment contract. And then for cells, you ask questions basically.

— Jacob Lauritzen, Latent Space
Key Insight
Fan-out converts one hard judgment call into ten thousand identical easy ones, then lets the lawyer spot-check the pattern. Running the first ten rows first is the whole trick: it makes a million-document question debuggable at the cost of ten.
07Build Philosophy

Build a Few Things That Feel Like Many

Lauritzen's job is to build a handful of deep primitives, lists, agents, projects, that let customers assemble the rest, so it feels like a huge product built by a tiny team.

the customer feels like we've built so much stuff but we've actually only built a few things.

— Jacob Lauritzen, Latent Space
Key Insight
His job isn't to say no to features, it's to make a few deep primitives feel like many. The agent-swarm is the payoff: lists spawn agents and agents spawn lists, so once the right three primitives exist, customers and legal engineers assemble the rest without the core team touching it.
08Where Value Goes

Paralegals Aren't Gone, the Bill Is

Lauritzen says the agents don't delete the junior roles, they move value from menial hours toward judgment, and the billing model is already shifting from hourly to flat project fees.

The billing model itself I think is also going to change. We we're already seeing that particular in the US that a bunch of stuff is actually being priced on like a project fee and not on an hour fee.

— Jacob Lauritzen, Latent Space
Key Insight
The fear (paralegals gone) and the reality (paralegals do more, bills go flat) point the same way: value moves from hours billed to judgment rendered. The work that survives is the work a lawyer became a lawyer to do; much of the lower-end work is already moving to flat fees or no charge.