By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Latent SpaceBuilding Legal Agents When You Can't See the Data — Jacob Lauritzen, CTO of Legora
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
At a glance
WHAT IT’S REALLY ABOUT
How Legora builds privacy-first legal agents with evals and workflow primitives
- Jacob Lauritzen explains Legora’s progression from a Swedish-law research assistant into an agentic legal workspace that drafts, edits, analyzes, and organizes legal work.
- He describes core product primitives—Agent, Projects, Lists, Intake, and Tabular Review—that aim to replace email-heavy collaboration with in-product orchestration.
- He outlines how Legora evaluates models using simulated legal “worlds” like a 10,000-file due diligence data room, while acknowledging frequent model regressions across tasks.
- He emphasizes the operational reality of legal confidentiality: Legora can’t inspect user traces, so it relies on privacy-preserving signals, failure-mode classification, and anonymized customer examples.
- He argues AI will raise lawyer productivity and shift monetization toward strategic judgment and client relationships, while changing (not eliminating) junior roles and billing norms.
IDEAS WORTH REMEMBERING
5 ideasLegora’s product evolved from simple legal research into a broader “legal agent” that can both do and orchestrate work.
He describes the earliest version as essentially ChatGPT paired with a Swedish law search tool, and says it expanded from question-answering into drafting, editing Word files, organizing files, and running more complex analyses.
Legora is building workflow primitives (not just AI outputs) to reshape how legal teams collaborate.
He says the product now emphasizes not only producing work (research, drafting, analysis) but also “orchestrating” it via collaboration features like Lists, Intake, and Portal, aiming to replace painful email-and-versioned-Word workflows.
Model choice is treated as an internal routing problem, with limited user-facing control.
He says they route different models by task, offer a primary-model picker in the agent (though he’s unsure it will remain), and use specialized models inside certain tools; the discussion frames this as keeping users on the “Pareto frontier” without forcing lawyers to understand model tradeoffs.
Evaluation is structured around realistic project simulations, not isolated benchmarks—and regressions are common.
He says they built evals as simulated “worlds” (e.g., a 10,000-file data room due diligence) and track performance across practice areas/tasks, noting model releases can improve some tasks while regressing others.
Privacy constraints force Legora to learn from approximations and structured feedback rather than raw traces.
He says Legora can’t look at user traces because the documents are highly confidential (often belonging to the law firm’s clients), so they rely on safer paraphrasing, coarse failure-mode classification (e.g., 10–20 types), and customer-provided anonymized examples when available.
WORDS WORTH SAVING
5 quotesNo one believed the smallest startup in a basement in Stockholm was gonna, you know, out-compete this huge behemoth that was, like, way ahead of us in the US.
— Jacob Lauritzen
The way everything started was ChatGPT with a Swedish loss- search tool in it. That was literally the entire everything.
— Jacob Lauritzen
You can't own people, but you can rent them.
— Jacob Lauritzen
We can't look at any traces, basically. We can't look at any data 'cause it's all super confidential, and that's, like, extremely frustrating to work with 'cause you, you basically don't know what people are doing.
— Jacob Lauritzen
I put up this slide of, um, you know, the 300 Spartans and, and, and- for the movie, and I was like, "We're just gonna be 20 engineers for the rest of the company's history, and like we're gonna beat everyone who's much bigger." And we're 100 engineers now, and it feels like we're way too small.
— Jacob Lauritzen
High quality AI-generated summary created from speaker-labeled transcript.
