AgentsAutonomy & tool use
Legora CTO Jacob Lauritzen explains how the legal AI company routes tasks across models, tests them by practice area and builds without seeing client data
Legora's CTO says a frontier-model upgrade raised citation reliability from about 70% to 99% without extra work. He says the company's edge comes from practice-area evals, task-level model routing, reusable product building blocks and staff who never handle client data.
From the Latent Space podcast episode of November 2, 2026 and cited sources: Lauritzen says a single model release fixed a citation problem, taking reliability from roughly 70% to 99%. [1] [2]
SOURCES[1] Building Legal Agents When You Can't See the Data - Jacob… lspace.beehiiv.com[2] Transcript: Building Legal Agents When You Can't See the… rss.beehiiv.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Lauritzen says a single model release fixed a citation problem, taking reliability from roughly 70% to 99%. [1] [2]
- Original source
- Latent Space podcast episode, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 2, 2026 podcast episode published
- Covered
- Nov 3 edition
- Evidence
- 9 sources across 8 sites; transcript sourced
- 1Building Legal Agents When You Can't See the Data - Jacob…lspace.beehiiv.com
- 2Transcript: Building Legal Agents When You Can't See the…rss.beehiiv.com
- 3SiliconANGLE: Legora Series Dsiliconangle.com
- 4Artificial Lawyer: Legora buys Wexlerartificiallawyer.com
- 5Legora Series E press releaselegora.com
- 6Artificial Lawyer: Anthropic legal Artifactsartificiallawyer.com
RisingHow it’s measured
YouTube video · 3.6K views · 34 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
Lauritzen says a single model release fixed a citation problem, taking reliability from roughly 70% to 99%. Legora's dedicated eval team tracks every model by legal practice area and task, and finds that new releases often improve some areas while making others worse. [1] [2]
Legora's agent splits work into parts and sends each part to a model chosen for latency, cost, governance and regional availability. Legora has not fine-tuned any model so far. It now plans to post-train open-weight models only for narrow tasks, such as fan-out retrieval for legal research, and does not plan to train its own main agent model. [1] [2]
The product is built on a few reusable building blocks: lists, projects, the agent and Tabular Review, which runs the first 10 rows as a trial. Litigation tools from the Wexler acquisition are built on the same blocks. Forward-deployed engineers build integrations without touching client data, and Legora keeps very few vendors on its data processing agreement. [1] [2] [4]
Lauritzen describes frequent per-area regressions between model releases and a Word add-in architecture rebuilt several times as models changed.
Read the full assessment
Implication: vertical AI builders may need domain-specific evals and swappable scaffolding, especially as Anthropic builds similar legal workflows directly into Claude.
Sources are checked; writing the dossier now.
Executive brief
Legora's CTO, Jacob Lauritzen, says one of the hardest failures in the company's legal agent fixed itself without any work from his team. Citation reliability went from roughly 70% to 99% after a frontier-model upgrade (Latent Space episode, ~11:35). He says Legora's real advantage lies elsewhere: evals by legal practice area, reusable product "primitives", strict data governance and forward-deployed staff. His engineers build all of this without seeing client data. The interview came out days after Legora raised money at an $8.8B valuation and Anthropic moved further into legal with ready-made workflows, including one called "tabular review", a name Legora says it popularised.
What changed and event timeline
Series D at $5.55B
Accel led a $550M round. About 800 customers in 50 markets. Legora said it uses Anthropic's Claude and named Anthropic's February legal plugin as competition ().
Wexler acquisition
This was Legora's fifth acquisition of 2026. Wexler is a London litigation "fact intelligence" startup with 18 staff, who will form a new London engineering hub ().
Series E at $8.8B
Atreides led a $550M round. Legora reports more than $200M in annual recurring revenue (ARR), over 140,000 users and more than 1,000 staff ().
Anthropic ships legal Artifacts
These are ready-made legal workflows inside Claude for Legal, tabular review among them. Artificial Lawyer describes them as repackaged existing Claude capabilities ().
Lauritzen on Latent Space
He covered model routing, evals by legal practice area, plans to post-train open-weight models, and competition with Claude ().
Capabilities and access
- Agent: started as ChatGPT plus a search tool over Swedish law.
- Tabular Review: each row is one document and each column is a question.
- Model picker: Legora resisted adding one, then gave in to customer demand. Models named on air include "Fable", "Astra" and Opus.
Read the full section
- Agent: started as ChatGPT plus a search tool over Swedish law. It now drafts documents, edits Word files, organises files and delegates work to subagents (~02:14–03:00).
- Tabular Review: each row is one document and each column is a question. Every cell cites its source. The agent can launch a review itself and runs the first 10 rows as a trial (~34:40–37:41).
- Model picker: Legora resisted adding one, then gave in to customer demand. Models named on air include "Fable", "Astra" and Opus. Exact version numbers were not given and are not confirmed in the reviewed sources (~06:57–07:48).
- Public APIs: started about six months ago. They do not cover the whole product (~47:38–48:14).
Technical analysis for researchers and developers
- Evals: a dedicated team tracks every model across practice areas and tasks as a heatmap.
- Routing: the agent splits a task into parts and routes each part to a suitable model.
- Post-training: Legora had not fine-tuned any model until now. It is not training its own main agent model (~53:09–53:54).
Read the full section
- Evals: a dedicated team tracks every model across practice areas and tasks as a heatmap. Lauritzen says new releases often improve some areas and regress others (~08:55–09:16).
- Routing: the agent splits a task into parts and routes each part to a suitable model. Model choice weighs latency, cost, governance and regional availability (~05:49; ~30:11–30:36).
- Post-training: Legora had not fine-tuned any model until now. It plans to post-train open-weight models only for narrow tasks, such as fan-out retrieval for legal research. It is not training its own main agent model (~53:09–53:54).
- Long documents: edits across very long contracts remain hard for models. Legora breaks them into chunks and coordinates the edits, and it has rebuilt its Word add-in architecture several times as models changed (~58:30–59:18).
- Citation checking: every claim links to a highlighted region of the source document, and the links can be followed through several layers of derived work (~57:07–57:37).
Claims and evidence
- Funding and scale: the $8.8B valuation, more than $200M ARR and over 140,000 users come from Legora's own release.
- 70%→99% citation fix, eval regressions, "about 60% of legal work done well today": these come only from the interview.
- Wexler metrics: the 8x revenue growth and over 400% net revenue retention are company-reported (Artificial Lawyer).
Read the full section
- Funding and scale: the $8.8B valuation, more than $200M ARR and over 140,000 users come from Legora's own release. Press reports repeat these figures (Tech Funding News) but no audit is cited. Some outlets date the announcement October 16 instead of October 21 (Nordic9).
- 70%→99% citation fix, eval regressions, "about 60% of legal work done well today": these come only from the interview. No eval data has been published, and no independent confirmation was found.
- Wexler metrics: the 8x revenue growth and over 400% net revenue retention are company-reported (Artificial Lawyer).
Context and prior work
- Harvey, the main rival, raised $550M at a $15.5B valuation (Tech Funding News).
- Claude for Legal: sources disagree on when Anthropic entered legal. SiliconANGLE describes a February 2026 plugin (SiliconANGLE).
- Agent labs: host swyx frames the business model as a "domain-specific wrapper" that keeps customers on the best cost/quality trade-off among frontier models, and compares it to Cognition, where he is an advisor (~06:07–06:34).
Read the full section
- Harvey, the main rival, raised $550M at a $15.5B valuation (Tech Funding News).
- Claude for Legal: sources disagree on when Anthropic entered legal. SiliconANGLE describes a February 2026 plugin (SiliconANGLE). Other coverage dates Claude for Legal to May (Unite.AI).
- Agent labs: host swyx frames the business model as a "domain-specific wrapper" that keeps customers on the best cost/quality trade-off among frontier models, and compares it to Cognition, where he is an advisor (~06:07–06:34).
Limitations, safety and contested findings
- Verification: Lauritzen says many lawyers no longer check citations. They spot-check the important parts, the way they would review a junior lawyer's work (~57:37–58:23).
- Writing style: some models are too chatty and verbose, and prompting has a hard time fixing it (~56:36–56:54).
- Data access: forward-deployed engineers never touch client data. Some are set up at the client like employees when the work requires it (~46:09–46:31).
Read the full section
- Verification: Lauritzen says many lawyers no longer check citations. They spot-check the important parts, the way they would review a junior lawyer's work (~57:37–58:23).
- Writing style: some models are too chatty and verbose, and prompting has a hard time fixing it (~56:36–56:54).
- Data access: forward-deployed engineers never touch client data. Some are set up at the client like employees when the work requires it (~46:09–46:31).
- Data processing: Legora keeps very few vendors on its data processing agreement. It does not use ElevenLabs for that reason (~31:23–31:42).
- Transcription errors: the automatic transcript renders "Legora" as "Lagora" and "Lagoon".
Business and practitioner implications
- Pricing: Lauritzen says US firms are moving low-end work to flat fees, and some of it to free.
- Platform risk: Anthropic offering tabular review and Harvey-built workflows directly inside Claude (Artificial Lawyer) puts pressure on simple wrapper products.
- For builders: keep evals by domain, expect to rebuild your scaffolding with each model generation, and post-train only narrow, high-volume tasks.
Read the full section
- Pricing: Lauritzen says US firms are moving low-end work to flat fees, and some of it to free. Value is shifting toward strategic M&A and litigation work (~19:21–19:51).
- Platform risk: Anthropic offering tabular review and Harvey-built workflows directly inside Claude (Artificial Lawyer) puts pressure on simple wrapper products. Lauritzen expects Claude to win small clients' work, but not $400M deals (~64:10–64:18).
- For builders: keep evals by domain, expect to rebuild your scaffolding with each model generation, and post-train only narrow, high-volume tasks.
The source trail.
Sources (9)
Building Legal Agents When You Can't See the Data - Jacob Lauritzen, CTO of Legora
Transcript retrieved via published_transcript; language en. Timestamped text, not direct audiovisual review. Source: https://rss.beehiiv.com/podcasts/transcriptions/01a19df6-a61b-7375-825b-0ce2635bee20.vtt. Automatic captions/transcription may contain errors.
lspace.beehiiv.comTranscript: Building Legal Agents When You Can't See the Data - Jacob Lauritzen, CTO of Legora
published_transcript
rss.beehiiv.com