How to Build a Legal Agent When You Can’t See the Data — Interview with Legora CTO Jacob Lauritzen

Latent.Space
November 2 · 67 mins
Legora cannot inspect a single trace of what its AI agents do while working for lawyers. That data belongs entirely to the law firm’s clients. How do you build and evaluate an agent when you can’t observe its inner workings?
Jacob Lauritzen is the CTO of Legora, one of the fastest-growing companies in legal tech history. The team started in a Stockholm basement with nothing more than ChatGPT, Swedish legal data, and a search tool. Today, lawyers use Legora to draft and edit Word documents, review 10,000 contracts across a single spreadsheet, and run due diligence. The agent also coordinates workflows across lawyers, clients, and other sub-agents. When Legora made its own tech acquisition, the team conducted the entire due diligence process using their own product.
Key Highlights
• Why Legora resisted adding a model picker for years, and what ultimately convinced them to launch it
• Jacob’s estimate that existing models already handle roughly 60% of legal work well, and why the remaining 40% demands smarter models
• How three engineers and three lawyers built evaluation environments modeled on real-world client workflows—functioning much like the “worlds” used in reinforcement learning (RL)
• How a citation accuracy bottleneck jumped from 70% to 99% overnight following a single model release, without Legora changing a single line of code
• Why Legora deploys all its systems on Azure, and why they don’t use ElevenLabs
• Tabular Review: A million-cell spreadsheet interface whose design—and name—has been widely emulated across the industry
• How legal intake works, how agents construct checklists, and how one checklist spins up downstream agents
• The distinct responsibilities of Legal Engineers vs. Forward-Deployed Engineers, and who actually writes code
• How Jacob once used a “300 Spartans” slide to pitch staying at 20 engineers—and why, at 100 people, he still feels understaffed
• Why Jacob considers models like Fable and Opus too verbose for legal drafting, preferring Astra’s concise style instead
• Why long-form contract drafting is fundamentally a retrieval problem
• Whether foundation model labs will commoditize vertical legal tools the way they disrupted coding tools
• What lies ahead for the law firm partnership model, billable hours, and junior associates
Legora is currently hiring engineers in New York, Stockholm, and London.
Chapters
0:00 Intro
0:40 Meet Jacob Lauritzen, CTO of Legora
1:05 Origins: Underdogs in a Stockholm basement
1:54 How Legora evolved from “ChatGPT + Swedish law” into a legal agent
4:12 The transition from assistant to agent
5:23 Which models Legora runs, and why they added a model picker
7:48 Are today’s models good enough for lawyers?
8:20 What agents couldn’t do nine months ago
9:18 Testing agents in simulated worlds and 10,000-document data rooms
10:07 The evals team and Legora’s 100 legal engineers
11:03 When evals feel futile: jumping from 70% to 99% citation accuracy overnight
12:18 What legal practice looks like in 5 to 10 years
14:18 Can the law firm partnership model survive?
16:40 What happens to paralegals and junior associates?
18:00 The billable hour and why earlier legal tech startups struggled
19:39 Where legal spend is flowing today
22:44 How Legora used its own product for M&A due diligence
23:59 Swedish law firms vs. US law firms
26:12 Product metrics when you can’t inspect execution traces
27:41 Rewriting traces and clustering failure modes into 10 to 20 buckets
28:42 Why data privacy led Legora to host everything on Azure
30:04 Open-weight models and model arbitrage
30:55 Transcription models and why Legora doesn’t use ElevenLabs
33:33 Agents, Tabular Review, project collaboration, and Word add-ins
34:24 How lawyers actually use Tabular Review
35:51 Why spreadsheets work so well, and what to learn from the first 10 rows
37:53 How the legal intake bot operates inside Slack
39:11 How agents build checklists, and how checklists trigger subsequent agents
40:15 Building on shared primitives and the Wexler acquisition
42:37 The role of Forward-Deployed Engineers at Legora
43:52 How Legal Engineers and FDEs onboard enterprise clients
46:43 The Forward-Deployed team’s roadmap for the coming year
47:24 Why Legora initially resisted building an API
49:26 Can the agent itself serve as the API?
50:12 The “300 Spartans” slide and the original plan to stay at 20 engineers
51:32 Big product bets Legora took that no customer asked for
53:05 Post-training open-weight models for legal retrieval
54:05 Search, retrieval, and late interaction
54:51 Anatomy of an effective legal agent test harness, and why drafting is retrieval
56:25 Which model writes the best first draft?
57:03 Hallucinations and Legora’s citation engine
57:41 Do lawyers actually verify citations?
58:19 MapReduce vs. a single massive context window for million-scale workloads