Legora builds legal agents without seeing customer data, so a six-person eval team approximates user behavior and tracks known failure modes.
Legora's citation accuracy jumps from 70% to 99% with a model release, with no work from the company.
Lauritzen says only about 60% of legal work can be done well today, and Legora now offers a model picker after long resisting.
Legora Builds Legal Agents It Is Not Allowed to Watch
Legora is building agents for lawyers under a hard constraint. It cannot read customer data, because the documents belong to law firms and their clients, so quality is measured by recreating user work in evals instead of watching real traces. Everything else follows from routing across models and chasing reliability, and Lauritzen argues the payoff is more legal work done better and cheaper, not fewer junior lawyers.
Takeaways
Legora cannot see customer data at all. Traces are off limits because the data often belongs to the firms' clients. Three engineers and three lawyers recreate user work, sort failures into 10 to 20 known modes, and build environments like a 10,000-file due diligence data room.
Model routing beats one model for legal work. Legora routes between models on capability, latency, cost, governance and regional availability, and ships a picker it resisted for a long time. Lauritzen puts well-served legal work at about 60%, with complex analysis still needing frontier intelligence.
Citation accuracy jumped without Legora doing anything. A citation problem went from 70% to about 90% on a model release, showing the scaffolding gets rebuilt as models move. Citations link output to a highlighted box in the source PDF.
Junior lawyers are not being replaced. Lauritzen says good juniors, like good engineers, are worth more than before, and Legora is not seeing the replacement the host describes. US billing is shifting to project fees and flat rates for lower-end work.
Law firm distrust of AWS shaped the stack. Everything runs on Azure because firms did not trust AWS or Google and Microsoft was already approved through Office 365. Legora keeps very few vendors on its data processing agreement and uses ElevenLabs for speech partly for that reason.
A hundred legal engineers shape the product. They are lawyers from big law, often with computer science degrees, who use the product daily as power users and write no code. Forward deployed engineers, first hired about nine months ago, write code into production.
Tabular Review turned rows into a product. Each row is a document and each column a question, up to a million cells, built on AG Grid. The agent drafts and edits Word files, and intake runs as a Slack bot that starts agents and adds tasks to lists.
What Lauritzen predicts next
Enterprises will reach Legora through Claude, the way he uses Linear through Claude while still opening Linear, while law firms use Legora directly because it is simpler and better.
Lawyers become far more productive, doing more work at higher quality for lower cost, and legal services become more widely available.
Tap a name to go there · tap a line to see why they are linked
On this episode 3
Show 1
Legora will post-train open-weights models only for bounded tasks where the cost or latency gap against a frontier model is large, naming legal research fan-out retrieval.
The company is toying with whether the agent itself can be the API, having started opening its APIs only about six months ago.