Legora began as a small Stockholm project: an early product was effectively ChatGPT augmented with Swedish law and a search tool. Jacob Lauritzen, now CTO, joined after helping with the startup’s YC application; he had prior YC experience with a different startup and later became head of engineering before taking the CTO role.
The company has scaled rapidly. Less than 18 months after launching publicly, Legora reported surpassing $100 million in ARR in April, and today serves more than 1,800 customers. The company announced two large financings in 2024: $550 million at a $5.55 billion valuation in March, and another $550 million in October that valued the business at $8.8 billion. Legora has completed multiple acquisitions in the same year; the most recent announced was Wexler, its fifth acquisition that year. In April it acquired the Stockholm legal research startup Qura.
What the agent does today
Legora’s agent drafts and edits Word documents, performs large‑scale contract reviews, and coordinates work between lawyers, clients and other agents. One signature product is Tabular Review: a spreadsheet‑style interface where each row is a document and each column is a question (for example: termination period, IP assignment clause, severance terms). The feature supports very large workloads — the product can handle up to millions of cells — but the UX encourages testing on the first 10 rows so teams can refine prompts before running at scale.
Other core product primitives include projects (containers for files), the agent itself, and lists — a flexible, Notion‑style database with triggers and automations. Intake is a Slack bot that turns inbound requests into tickets on a list, which can in turn trigger agent runs.
Why the billable hour stopped being a blocker
Early skeptics wondered why lawyers would adopt technology that made them faster when they bill by the hour. Lauritzen says the driver was market pressure: in the U.S. routine work has already shifted toward flat project fees or even pro bono work to preserve client relationships. Revenue has migrated to high‑value tasks — M&A, litigation, and time spent on client relationships — so firms have incentives to use tools that reduce cost on low‑value work while allowing human lawyers to focus on high‑value judgement.
Legora’s data show that the role of junior lawyers and paralegals is evolving rather than disappearing: good juniors are more valuable because they accomplish far more with AI‑powered tools.
Evaluations without reading customer data
Most teams iterate on agents by reading traces of real user activity. Legora cannot do that: the documents users upload are confidential and often belong to the firm’s clients. To evaluate models, a dedicated evals team recreates realistic legal tasks in controlled environments. Three engineers and three lawyers build RL‑style environments (for example, a 10,000‑file data room where the agent must run due diligence and produce a report that captures the right red flags). Results are plotted as heat maps by practice area and task: models often improve some capabilities while degrading others.
Legora relies on a pool of roughly 100 legal engineers — lawyers who spent several years in big law, frequently with CS backgrounds — to help shape evaluations and product features. In production, failures are categorized into 10–20 known modes and cheap LLM‑based checks run behind feature flags.
Privacy constraints also shaped technical choices: Legora runs everything on Azure because, at the company’s start, many law firms would not accept AWS or GCP while Microsoft was already an approved vendor via Office 365 reviews.
Where smarter models still matter
Legora resisted a model picker for a long time but eventually added one: the company’s philosophy is that lawyers should not need to choose model families or quality tiers. The platform should select the best trade‑off among quality, cost, latency, governance, and regional availability. Lauritzen estimates current models handle roughly 60% of legal work well today; for those tasks the remaining issues are cost and latency. Complex analysis and editing across million‑token contracts still need more model capability.
Legora is post‑training open‑weight models for narrow tasks such as legal research, but does not plan to train its own frontier general‑purpose model. In drafting tasks, some models (Fable, Opus) tend to be too verbose while Astra produces more concise drafts. To mitigate hallucinations, every output from an agent includes citations that highlight the bounding boxes in source PDFs and allow multi‑layer tracing back to the original document or Tabular Review cell.
Building many products from a few building blocks
Six months into his CTO role, Lauritzen presented a slide of the 300 Spartans and declared the company would stay at 20 engineers; today Legora has about 100 engineers and he still feels understaffed. His current objective is to build a system where the team ships a small set of primitives but customers feel like they get many products.
Primitives (projects, agent, Tabular Review, lists) let legal engineers configure new experiences without code. The same building blocks power acquired products such as Wexler’s fact intelligence. The company aims to enable many specialized products quickly by letting legal engineers compose primitives and, when needed, bringing in forward‑deployed engineers for integrations.
Legal engineers first, forward‑deployed engineers second
Legora’s legal engineers configure workflows and product features as power users without writing code. Forward‑deployed engineers (FDEs) write production code, build integrations into a firm’s CRM or document management system, and feed their learnings back into the core product. The company began hiring FDEs in New York about nine months prior to the interview; FDEs are rotated onto client projects for a couple of months to solve bespoke problems and then move on.
When a major client signs, Legora assigns legal engineers and adds an FDE if the project requires deep integration or custom code. Over time Lauritzen expects a rotating pool of FDEs who will be deployed across customers.
Platform, APIs, and whether the agent can be the API
Legora has been cautious about broad public APIs because the rapid pace of product change makes long‑term API maintenance costly. The company began building APIs more recently but still limits the surface area. Lauritzen is exploring the idea of an agent as an API — i.e., triggering agent behavior rather than providing a fixed endpoint — but notes trade‑offs in cost, latency and determinism.
Engineering priorities: search, retrieval and harnesses
Legora emphasizes search and retrieval: pulling the right precedent documents and facts is a prerequisite for drafting and analysis. Many gains come from improved architecture, pipelines and harnesses rather than model training alone. For certain narrow tasks, post‑training open‑weight models can be inexpensive and effective; Legora has started that work for legal research.
Transcription and speaker attribution are other tool areas: Legora uses various models (early Gemini and Microsoft transcription models were called out) and avoids some vendors (e.g., ElevenLabs) for contractual DPA reasons. The company customizes embeddings rather than relying on defaults, and it performs late‑interaction retrieval experiments to improve relevance.
Hallucinations, citations and long‑form reasoning
To control hallucinations, Legora attaches citations to generated text; users can click through to the precise location in the source PDF or to Tabular Review cells. Lawyers do verify citations selectively — similar to how engineers review code — focusing on the parts that matter. Long‑document reasoning and editing across very large contracts remain challenging: most solutions today use chunking and orchestration (MapReduce‑style or multi‑step edits) rather than a single, one‑shot million‑token pass.
Where this might go next
Lauritzen expects law firms to adopt platforms like Legora directly, while some enterprises will access legal capabilities through broader tools (Claude Cowork, etc.). He does not expect the large model labs to cut off specialized vendors: legal work involves deep vertical features, governance and workflows that general labs are unlikely to match in full for high‑value transactions.
Legora is hiring aggressively — engineering hubs in New York, Stockholm and London are growing — and continues to invest in primitives, evals and targeted model training for bounded tasks rather than attempting to train a general frontier model.
Takeaways
Legora illustrates how a domain‑specific agent lab builds beyond chat: combining search, citation‑aware generation, workflow primitives and privacy‑conscious evals produces a product lawyers will adopt for high‑value work. The company’s funding and acquisitions reflect strong market demand for specialized legal AI that balances capability with governance and client confidentiality.
-- Key names and concepts in this article: Legora; Jacob Lauritzen (CTO); Max Junestrand; Tabular Review; lists; projects; intake (Slack bot); Wexler; Qura; Azure; legal engineers; forward‑deployed engineers; Fable, Opus, Astra.
