Legora CTO Jacob Lauritzen says his engineers generally cannot read the confidential client work their legal agents perform. On Latent Space, he explains how the company combines lawyer-built tests, coarse production signals, and customer-supplied reproductions to improve the product.

Test the project, then separate its failure modes

At 08:44, Lauritzen describes a heat map covering models, legal practice areas, and tasks. New releases improve some cells while making others worse. An overall model ranking hides the decisions Legora actually needs to make.

The evaluation team includes roughly three engineers and three lawyers. It also draws on about 100 legal engineers, former practicing lawyers who use the software and work with customers.

One test environment resembles a 10,000-file data room for acquisition diligence. The desired report must find particular risks, avoid irrelevant material, and remain concise. Finding the red flags and drafting the report are evaluated separately. A fluent memo can conceal a failed search.

Lauritzen estimates that Legora can already handle roughly 60% of the work well. He frames the remaining improvements for that portion largely around speed and cost. That is his estimate, not a measured result across the legal profession. He still wants more intelligence for complex analysis.

Keep production feedback narrow

At 26:25, Lauritzen describes ordinary adoption signals: repeat use, time in the product, activity, and renewals. A low-cost model also produces an online assessment of whether a task was solved. Comparing feature-flagged rollouts helps identify regressions.

His team cannot simply inspect the underlying conversations and documents. He says much of the material belongs to the law firm’s clients. The evaluation team must approximate real projects without that debugging access.

A host suggests generating paraphrased versions of production traces. Lauritzen cautions that removing names may leave sensitive information intact. He prefers restricted classifications: the practice area, task category, or one of roughly ten to twenty known failure modes.

These categories preserve some feedback while sacrificing detail. They also leave a gap around unexpected failures. Later, he describes asking willing customers for anonymized example documents and problems the team can reproduce. The episode offers an engineering approach to limited visibility, not proof that its tests cover every real matter.

Preview bulk work before paying for it

Tabular Review gives each document a row and each question a column. For employment contracts, questions might cover termination periods, intellectual-property assignment, or severance.

The agent can create the review, analyze its results, and draft a memo. Lawyers can comment, flag cells, mark them reviewed, and follow citations back to the documents. Lauritzen mentions reviews reaching a million cells, which differs from a million-document benchmark.

The product first runs ten rows. Users check the answers and refine their questions before launching the full job. A small preview contains the cost of a badly framed instruction.

Citations also work across layers. A memo can link to a table cell that links to a highlighted passage in the original document. Lauritzen acknowledges that lawyers differ in how carefully they inspect those links. Evidence being available does not mean someone checked it.

Governance shapes the product as much as intelligence

Lauritzen says Azure became Legora’s infrastructure choice because early law-firm customers had already approved Microsoft. A data processing agreement with few vendors limits where material can travel. Model selection therefore includes regional availability and governance alongside quality, speed, and price.

The same trade-off appears in customization. Legal engineers configure workflows and train users. Forward-deployed engineers write integrations and production code. Shared lists, projects, and agents let the company reuse a few building blocks across many customer requests.

There are firm limits to what the interview describes as shipped. At 53:09, Lauritzen says Legora has not fine-tuned models up to this point. Narrow post-training for search is a direction. Having an agent serve as the product’s API is another idea, with cost and unpredictability still unresolved.

Long contracts also remain difficult. Definitions and edits interact across distant passages, while independent chunk summaries can lose the connection between them. Better retrieval, document editing, and human review remain part of the job even as models improve.

Caveat: Privacy and legal-work claims are Lauritzen’s account, not independent audit findings or a guarantee of correctness.

Listen to the full Latent Space interview, released November 2, 2026.