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

Host interviews Jacob Lauritsen, CTO of Legora, about the company’s European roots and rapid expansion into New York and San Francisco. Lauritsen says the company remains culturally humble because it began as a small Stockholm basement startup competing against a much larger, more advanced American rival. As Legora has entered the United States, the team has had to become more willing to tell the market how good it is. He explains that his Danish surname means “son of Lauritsen,” while the host frames Legora as one of legal tech’s fastest-growing companies.

Legora began as ChatGPT with a Swedish legal-search tool, answering questions about prior cases, precedent, and how lawyers had handled similar situations. The product now drafts documents, edits Word files, organizes files, runs complex analyses, and performs “Tabular Review,” which bulk-extracts information from large numbers of agreements for due diligence and fundraising. Lauritsen divides the product into doing legal work and orchestrating it: Lists acts like Linear for legal tasks, while Intake and Portal coordinate lawyers, clients, and agents. He describes the overall experience as ChatGPT or Cowork, but specialized for legal work.

Legora uses multiple models by task, subdividing problems and routing different steps to the most suitable model. The agent includes a primary-model selector, although Lauritsen is unsure it will remain permanently. The host notes that users increasingly want to experiment with models such as Fable, Opus, Astra, and “Mythos,” but Lauritsen says perhaps 60% of legal work is already handled well; the remaining challenges are intelligence, cost, latency, and expanding the range of work. Users can choose expensive frontier models, but doing so may consume credits roughly ten times faster.

Legora’s evaluation team combines three engineers, three lawyers, and a larger pool of “legal engineers”—former big-law lawyers, often with computer-science backgrounds, who use the product and help shape it. Evals recreate legal environments such as a 10,000-file due-diligence data room, then separately score red-flag detection, report drafting, wording, completeness, and analytical soundness. Lauritsen says model releases can improve one practice area while regressing in another. Earlier, citation accuracy rose from about 70% to 99% simply because a new model fixed the issue, requiring no work from Legora.

Lauritsen expects lawyers to become more productive, produce higher-quality work at lower cost, and spend more time on risk profiles, judgment, strategy, and client relationships rather than searching huge document sets. The host provocatively says “paralegals, gone” and “junior lawyers, gone,” but Lauritsen pushes back: strong junior lawyers will become much more valuable because they can produce far more with AI, even if the profession becomes more polarized. A paralegal handling a million-document review may no longer be needed for that task, but an AI-skilled paralegal could take on more meaningful work—or simply have time to sleep.

The speakers discuss whether faster work will reduce legal fees. Lauritsen says US firms are already pricing more lower-end work by project or flat fee, with some small reviews even done for free, while strategic M&A, litigation, and client relationships remain valuable. He also argues that partnerships exist because professional-services firms mainly own human capital, not durable assets. Legora wants firms to codify their methods, precedent, and judgment into systems that become firm IP. That could require new incentives, since partners currently receive little benefit from building assets that remain with the firm after they leave.

Legora used its own product during its acquisition of Kira and other transactions in Canada and Australia. Its team performs much of the document review with Tabular Review, while external counsel still evaluates strategic risks and jurisdiction-specific issues such as Australian payroll. Intake emerged from Legora’s internal Slack workflow, where employees were already sending legal requests and complaining that the process was inefficient. The company says its tools have surfaced red flags early enough to stop deals, although Lauritsen declines to disclose examples. American customers generally move faster and request more custom integrations than European customers, who more often prefer an off-the-shelf product.

Because law-firm data belongs partly to clients and is highly confidential, Legora generally cannot inspect customer traces or real work. It hosts primarily on Microsoft Azure because law firms already trust Microsoft through their Microsoft 365 security reviews; earlier resistance to AWS, Google, and on-premise alternatives has diminished as cloud adoption became more accepted. Legora measures renewals, repeat usage, time in product, and online evaluations, using inexpensive language-model checks and feature flags to detect improvement or catastrophic decline. When real data cannot be used, the team relies on safe synthetic or anonymized cases and classifies likely failure modes rather than exposing client material.

For a typical diligence project, each row in Tabular Review represents an employment or other contract, while columns ask repeated questions such as the termination period, whether an IP-assignment clause exists, or how much an employee receives upon termination. Legora can process up to roughly a million cells, identify red flags, draft a client memo, and provide citations back to source documents. The agent can also initiate the review automatically after a user uploads thousands of files. Legora runs the first ten rows before scaling up, letting lawyers refine the questions and check that results make sense. Users can comment, collaborate, and mark cells as reviewed.

Lists functions as a configurable database and glue layer: agents can create lists, read them, trigger actions, and populate fields. Legora uses this primitive for custom workflows, including litigation fact intelligence from its Wexler acquisition, where facts, entities, dates, and contradictions are extracted from huge case files and compared with hearing transcripts. Legal engineers configure workflows without coding; forward-deployed engineers write production integrations and may temporarily work as embedded customer employees. Lauritsen initially wanted to keep Legora to 20 engineers, but the company reached 100 because demand is intense. He is now considering whether the agent itself could serve as an API, despite cost, latency, and occasional unpredictable actions.

Legora has not historically fine-tuned its own models, but it is beginning to post-train open-weight models for bounded tasks such as legal research, where a specialized retrieval model can be cheaper and faster than a frontier model. Lauritsen identifies search, analysis, and drafting as the most important areas for custom harnesses. Complex contracts require finding the right precedents and combining clauses from multiple documents; long-context editing remains difficult because models must remember definitions and changes across a million-token document. Legora uses citation chains that can lead from an answer to a table cell and then to the source document. Lauritsen says concise models such as Astra may draft better than verbose models such as Fable or Opus.