Topics Pioneers Insight Method Author
Back

Building Legal Agents When You Can't See the Data — Jacob Lauritzen, CTO of Legora

·Latent Space

Rendering the long image…

Topics Startup Playbook for Vertical AI Applications AI Training Data and Evaluation Benchmarks AI Inference Cost Optimization AI Workforce Displacement and Labor Disruption How AI Reshapes Enterprise Workflows Enterprise AI ROI and Cost Optimization
Pioneers swyx Jacob Lauritzen Basil Chatha

Summary

  • Legora’s defense against frontier labs is deep legal workflow ownership. swyx argues that the labs “are never gonna go as deep as we are.” He says Claude may handle smaller corporate matters, but that for a “$400 million M&A transaction,” clients will want Legora.
  • Confidentiality makes improving a legal agent unusually difficult. Basil says Legora cannot inspect customer traces because documents often belong to law firms’ clients, so the team must reconstruct representative workflows for evaluation. Jacob describes an evaluation team of about three engineers, three lawyers and a pool of 100 legal engineers; Basil’s frustration is “I have no idea what people are actually doing with it.”
  • Model gains are commoditizing routine tasks while complex analysis still needs more intelligence. Jacob estimates roughly 60% of existing work can already be done well, making cost and latency the focus for that portion. One citation problem improved from about 70% to about 99% after a model release—“we did nothing”—but regressions elsewhere make task-specific evaluation important.
  • The pressure on junior labor is substantial, but replacement remains contested. The episode’s “Paralegals, gone” claim meets Jacob’s “That’s actually not what we’re seeing”: strong juniors can become more valuable. Meanwhile, competitive pressure is pushing some routine US legal work toward flat fees or free service, with revenue remaining in strategic advice and client relationships.
  • Product expansion requires more people despite reusable software primitives. Jacob’s plan for 20 engineers has become 100, and the team still feels “way too small.” Lists, projects and agent runs provide the desired software scale; legal engineers configure workflows, while forward-deployed engineers write production code and integrations.
  • Law firms could build durable intellectual capital, but partner incentives work against it. Jacob wants firm-specific judgment and precedent codified into assets that outlast individuals. The challenge is getting partners to “build an asset for the company” when they are rewarded for bringing in clients and receive no dividends after leaving.
  • Governance and retrieval engineering remain priorities alongside better models. Basil says Azure’s existing acceptance helped Legora enter law firms, while data-processing agreements constrain its vendor choices. He says Legora has not fine-tuned models so far; the opportunity is “very specific, bounded tasks,” such as cheaper, faster legal-research retrieval, rather than training a frontier-agent competitor.

Deep dive

1. Legora is moving from legal answers to legal operations

  • Lauritzen describes the starting product as ChatGPT with a Swedish law-search tool: answering research questions by finding precedents. He corrects the chronology—his own YC batch was 2023; Legora’s was 2024—and tentatively dates the assistant-to-agent transition to about a year earlier. The company began as a small Stockholm startup competing as an underdog.

  • The product now drafts, edits Word files, organizes documents and performs complex analysis. Lauritzen’s distinction is between “doing the work” and “orchestrating the work”: Tabular Review handles bulk document analysis, while Lists, Intake and Portal manage tasks and collaboration among clients, lawyers and agents. Word and Outlook add-ins meet lawyers inside their existing tools.

  • The interface can fairly be described as ChatGPT or Cowork for legal work, Lauritzen agrees. The larger question is what replaces “a 200-thread email” with “V5 Word files back and forth.” If capabilities improve and prices fall, he imagines a command center where lawyers react to five things agents have already done, then review and improve the system.

2. The model selector exposes a tension between simplicity and control

  • The ideal, in swyx’s framing as a Cognition adviser, is a domain-specific GPT wrapper that delivers frontier capabilities without making customers model experts. Lauritzen agrees: a lawyer billing $1,000 an hour does not want to judge “Sonnet High or an Opus Mid.” Legora has a primary-model picker; Jacob says the team has gone back and forth on whether it will stay, and that customers want to experiment with new models. Separately, swyx says his team fought a model picker before giving customers the control they wanted.

  • Legora routes individual tools and subtasks to different models; Lauritzen finds subdividing the problem more effective than making one global choice. Even when clients ask for Mythos, “a lot of legal work does not require Mythos-level intelligence”; Jacob says choosing the more powerful option can use credits “10 times faster.”

  • Basil’s pushback is whether models are already good enough. Lauritzen’s answer is a distribution: probably about 60% of current work can be done well, leaving cost and latency as the main optimization. Complex analysis still needs more intelligence, however, and the work available to lawyers might itself expand as capabilities improve.

3. Confidentiality forces Legora to reconstruct work it cannot inspect

  • “We can’t look at any traces, basically,” Basil says: confidential material often belongs to the law firm’s clients. That leaves him saying, “I have no idea what people are actually doing with it.” The evaluations team therefore approximates workflows it believes customers are running rather than freely inspecting production use.

  • The core evaluations team has about three engineers and three lawyers, backed by 100 legal engineers with Big Law experience, typically three or four years and sometimes CS backgrounds. Test “worlds” might contain a 10,000-file diligence room and a target report. Evaluations separate red-flag discovery from drafting quality, checking analysis, wording, inclusions and omissions across practice areas.

  • Lauritzen tentatively agrees Sonnet 4.6 and Opus 4.6 were early agent models; subagent delegation for fan-out retrieval works much better now. One citation issue improved from about 70% to about 99% after a release: “we did nothing.” But the team’s heat map also shows releases improving some tasks while regressing on others, more often than people might expect.

  • Basil says product signals include renewals, repeat use, DAU/MAU and hours spent, which he describes as strong without supplying figures. Cheap online LLM judgments support feature-flag comparisons; restrictive classifications sort failures into 10 or 20 known modes. swyx proposes paraphrased traces, but Basil warns that removing names may not protect confidentiality. They favor coarse categories over detailed inspection.

4. Firm-specific AI could create capital that partnerships struggle to reward

  • Lauritzen introduces his outlook with “I should preface this by saying I don’t know.” He expects lawyers to produce higher-quality work at lower cost and spend more time on judgment, manageable risk and clients. Less time would go into searching huge documents for the smoking gun or performing the menial work that currently absorbs junior effort.

  • His partnership explanation centers on human capital: professional-services firms cannot own their people, so they make them owners to encourage retention. Legora wants to help codify “the X law firm way of doing things” into precedents, processes and possibly a self-improving knowledge graph—durable IP capturing expertise and judgment beyond any individual’s continued employment.

  • The incentive problem comes first. Partners are rewarded for bringing in clients under “eat what you kill” or lockstep arrangements, rather than diverting time into codifying knowledge. Once they leave, they receive no dividends on assets they helped build. Lauritzen wants incentives to “build an asset for the company,” but does not claim partnerships must disappear.

5. The junior-replacement thesis meets pushback as billing changes

  • The episode includes the provocation “Paralegals, gone” and “Junior lawyers, gone”; in the fuller discussion Basil voices the claim, while the opening clip attributes a similar line to swyx. Jacob’s opening response is “That’s actually not what we’re seeing.” The host describes document review previously assigned to paralegals working until around 4 a.m. as work that no longer requires the same staffing. Those workers may be able to take on other valuable work.

  • Lauritzen’s counterargument is talent polarization: good junior lawyers, like good engineers, can be worth more because their output multiplies, and firms still need the next generation. swyx suggests “meat proxy” for the AI operator; Lauritzen resists that description when someone knows the tools and delivers useful results: “you’re actually just doing meaningful work.”

  • Lauritzen’s earlier objection was that hourly billing punishes efficiency: “They earn less money.” He now sees competitive pressure forcing change, with some routine US work moving to project fees, flat rates or free service. Asked where money remains, he points to big work, including litigation, not only M&A. Cheaper routine work also supports client relationships.

6. Tabular Review makes large-scale diligence inspectable

  • The concrete workflow is 10,000 employment contracts, one per row, with questions as columns: termination period, IP assignment, severance or compensation. The tool might fill cells with “30 days” or flag unusual terms. Lauritzen says the grid is the easiest form they have tried for overview and review; lawyers can inspect citations, mark reviewed cells and comment with colleagues.

  • The agent can create the review itself: drop in 10,000 files and ask for due diligence. It launches the table, waits, analyzes results and highlights red flags, such as excessive CEO compensation. It can then draft a client memo or leave the flagged results for the lawyer to examine.

  • Capacity is “up to a million cells, I think,” Basil says. The interface uses AG Grid, and users are happy working there rather than exporting to Excel. Before running the whole job, Tabular Review runs the first 10 rows twice; users can iterate their questions on that small set until results look right, then expand the fan-out.

  • Legora uses its tools for acquisitions, including Kira in Sweden, and diligence in Canada and Australia. Basil says the company does much of the diligence itself but still brings in counsel for strategic questions, including Australian payroll and deal structure. The tools have found red flags that made Legora “pull the brakes”; he declines to discuss those cases.

7. Lists turns legal requests into coordinated agent work

  • Intake is a “storefront thing” rather than simply a chat interface. A request arrives in Slack, kicks off an agent and creates a task in an intake list. Lawyers and agents work inside Legora, mark the task done and return the result to Slack. The person requesting the review does not need to change their workflow.

  • Lists is Linear-like for tasks and Notion-database-like in flexibility, with fields, triggers and automations. Projects hold files; configured agents read and populate those structures. “Agents can create lists, and lists can create agents,” Lauritzen says, describing how one piece of work can trigger another, update fields or add documents without someone initiating every step.

  • swyx asks whether this becomes “end user programming.” Lauritzen’s answer is that flexibility helps even if customers never see it: legal engineers could build fund-management dashboards with generative UI, or a potential marketplace of specialized products. His analogy is “a Lovable inside Legora”—also a way for the company to build its own features faster.

  • The UK acquisition Wexler provides litigation fact intelligence: extract entities, dates and facts from large document sets, then compare corroboration and contradiction. Lauritzen’s example checks testimony against an email: “he said this thing, but actually the email over here says that he wasn’t there.” He sees this as another product buildable on the same primitives.

8. Deployment needs legal power users and engineers who ship

  • Lauritzen distinguishes two roles: legal engineers configure the product as power users and “don’t write code”; forward-deployed engineers write “code into production,” build integrations and feed missing capabilities back into the product. A large, engaged customer might receive both, with the team sitting beside lawyers to teach the product, map workflows and implement them.

  • The data restriction remains: FDEs build the system without touching underlying client data. Exceptionally, they may be onboarded like employees, including receiving a law-firm laptop. There is no fixed Palantir-style rotation between client sites and product work. Lauritzen says it is early, but imagines a circulating group that might spend two months on one project before moving on.

  • Basil says US law firms and corporate customers move faster and want more custom systems; European customers more often start off the shelf. US requests range from Salesforce integrations to entire new products Legora cannot build for them. Jacob asks whether this reflects greater existential fear; Basil says he does not know, and stresses that both regions see AI as strategically important.

9. The 20-engineer plan gave way to reusable primitives

  • Lauritzen once used the 300 Spartans from the film to pitch staying at 20 engineers and beating larger competitors. “We’re 100 engineers now, and it feels like we’re way too small.” Competition and customer demand made that position untenable. His job is now to make customers feel Legora has built many things while actually building a few reusable primitives.

  • Basil says most product work responds to a “mountain of demand,” rather than low-conviction experiments. He identifies two strategic bets customers were not requesting: greater agent agency and collaboration inside Legora instead of email. The team judged these to be the industry’s direction; legal engineers then help firms restructure workflows around those capabilities.

  • Legora started offering APIs around six months earlier after resisting them for a long time; it does not promise the web app’s entire surface through APIs. Lauritzen’s reason is maintenance: changing data models become harder to evolve when external API versions must keep working. That cost felt unaffordable at 20 engineers and more manageable as the company grew.

  • “Can the agent be the API?” is an idea Lauritzen is considering: customers could trigger an agent to achieve an outcome instead of calling a dedicated endpoint. He emphasizes that he is not saying Legora is doing this. swyx raises unbounded execution, cost and latency; Lauritzen adds nondeterminism—“one or two times out of a hundred” it might do something unwanted.

10. Governance changes which models are usable

  • Basil says Legora’s infrastructure is on Azure because early law-firm customers rejected AWS and distrusted Amazon and Google, but already accepted Microsoft through 365 and prior security reviews. He says that has changed: customers understand cloud adoption, and on-premises discussions have “disappeared completely.” The hosts note that this substantial shift took only about three years.

  • Model arbitrage means optimizing capability, cost, latency, governance and regional availability together. Basil says an early Gemini was strong at transcription and speaker attribution; he finds Microsoft’s transcription model better than that older Gemini, “from what I’ve seen.” ElevenLabs is excluded because Legora’s data-processing agreement keeps a short vendor list. Integrated providers can help even without leading every individual task.

  • Basil emphatically rejects the idea that occasional transcription errors are acceptable. swyx recalls a court role, possibly in Australia, consistently becoming “condom”—specialist jargon still breaks these systems. His proposed repair runs an LLM over the transcript to extract keywords, then supplies those keywords to another transcription pass.

11. Search and drafting offer bounded opportunities beyond frontier scaling

  • Basil says Legora has not fine-tuned models “up until now,” but capable open weights make post-training attractive for “very specific, bounded tasks.” Legal research is the example: fan-out retrieval queries could become fast and cheap. He is not proposing that Legora train its own foundation model to be the main agent; he thinks it would most likely lose that contest against Anthropic.

  • Basil sees “low-hanging fruit” in pipelines and the harness before training models. He says the search-focused engineer Song talked to TurboPuffer about multi-vector support; Basil clarifies that Legora is not currently using late interaction. Jacob recommends customized embeddings rather than choosing a general benchmark leader. swyx agrees, adding that Legora’s many languages complicate that work.

  • Drafting is also a search problem. Basil says lawyers do not begin by typing a huge contract from scratch: they find the firm’s previous clauses from similar transactions. “Similar” might mean industry, company type or something else, and the right clauses may come from several precedents. Selecting and combining those pieces is itself non-trivial.

  • Fable and Opus are “extremely chatty” and “very verbose,” usually undesirable in legal writing; Astra is more concise. Lauritzen says prompting has not made verbosity easy to fix. He thinks the next Anthropic model will probably improve substantially on this, while acknowledging uncertainty. The team already gives providers direct feedback on these shortcomings.

12. Citations support review while long-document reasoning stays difficult

  • Legora links written claims to highlighted bounding boxes in source PDFs; an agent’s memo can cite a Tabular Review cell or row, letting users drill through several layers. Asked whether lawyers verify citations, Basil says “No.” He then describes a code-review-like distribution: some check every comma, others trust the agent, and experienced reviewers focus on the points that matter.

  • Long-contract editing remains difficult because changed definitions, clauses or ordering can require changes elsewhere. Basil says Legora chunks the work and orchestrates edits. The discussion notes that the Word add-in has moved through MapReduce and other scaffolding as models change; swyx says his own news workflow recently moved the other way, from MapReduce to putting everything into context, because results had become better.

  • Retrieval has a cross-chunk failure mode: a two-page discussion can be highly relevant as a whole while each individual chunk looks irrelevant. Summarizing chunks separately can lose critical nuance. swyx says customers sometimes anonymize problem documents and supply a reproducible example: “if we can recreate it, we can fix it.”

  • Jacob raises buying a law firm for its data. swyx is uncertain whether that resolves client-data restrictions, including after clients leave. Acquiring inputs and finished outputs also does not reveal how work happened or whether the result was good: someone must “reverse-engineer it.” Real task traces are much more valuable.

13. Lauritzen expects generalist AI and specialist platforms to coexist

  • Asked where lawyers will work in two years, swyx answers “both.” Enterprises might invoke Legora through a broader assistant while still opening its interface, just as he uses Linear through Claude but also visits Linear. For law firms, swyx’s prediction is more direct: “I think law firms will just use Legora” because it is “simpler and better.”

  • Jacob asks whether coding labs cutting off competitors will happen in law; swyx answers, “No, I don’t think so.” Jacob notes that sales representatives do not make that call. swyx’s argument is that frontier labs have broad enterprise knowledge work to cover, while they will not go as deep in legal as Legora.

  • swyx says Claude taking a small corporate client’s legal work is fine. His dividing example is the large transaction: “But if you’re doing a, you know, $400 million M&A transaction, you’re gonna want Legora.” His prediction about the labs is that “they’re never gonna go as deep as we are.”