How Legora Built a Legal-AI Agent Without Seeing Client Data

Legora’s rise from a Stockholm basement prototype—ChatGPT, Swedish law, and a search tool—to a $550 million Series E raise at an $8.8 billion valuation is anchored in a single tension: how to improve AI agents when the data they must process is legally confidential. Jacob Lauritzen, CTO, describes the company’s solution in a conversation hosted on Latent Space.

Simulated evals replace real-world traces

Because law-firm uploads are client-confidential, Legora cannot read the documents that power its agents. Instead, a six-person eval team—three engineers and three lawyers—builds reinforcement-learning environments that mimic typical client workflows. A flagship scenario is a 10,000-file data room where the agent conducts due-diligence and produces a red-flag report. Results are plotted on a heat map by practice area and task, revealing that a new model may improve on some legal functions while regressing on others.

In production, failures are bucketed into 10 to 20 known modes, and cheap LLM-based checks run behind feature flags. The company draws on a pool of roughly 100 “legal engineers,” lawyers with three to four years in big firms and often a CS background, to design these simulations and interpret outcomes.

Model arbitrage and the limits of current LLMs

Legora resisted a model-picker for months, arguing that lawyers should not decide whether a contract is a “Sonnet-high” or “Opus-mid” task. The platform now automatically selects the best model based on quality, cost, latency, governance and regional availability. Jacob estimates that today’s models handle about 60 % of legal work well; the remaining tasks—complex analysis and editing of million-token contracts—still need more intelligence.

For narrow tasks like legal research, Legora uses open-weight models trained post-hoc, but it does not plan to develop its own frontier model. Drafting models such as Fable and Opus are deemed too verbose, while Astra is preferred for concision. To mitigate hallucinations, every agent-generated passage includes a citation that highlights a bounding box in the source PDF, adding a layer of traceability.

Modular primitives drive product velocity

Legora’s engineering philosophy centers on a small set of primitives: projects (containers for files), the agent, Tabular Review, and lists. Lists function as Notion-style databases with triggers; agents can read and write to them, and list changes can launch new agent actions. An intake Slack bot converts user requests into tickets on a list, feeding the workflow.

Tabular Review, now copied across the industry, lets users drop a large contract set—up to a million cells—and query specific clauses. The interface starts with ten rows to let users refine prompts before scaling. Legal engineers configure new products on these primitives without writing code, while forward-deployed engineers (FDEs) write production code, build CRM or document-management integrations, and feed insights back into the core platform.

The first FDE, hired nine months ago in New York, exemplifies this split: legal engineers shape the product like power users, and FDEs handle the integration work without ever touching client data. Jacob expects a rotating pool of FDEs to embed with large clients for a few months, then move on to the next engagement.

Infrastructure choices dictated by privacy

Law firms’ reluctance to trust AWS or GCP led Legora to run all workloads on Azure, leveraging Microsoft’s existing approval through Office 365. This decision also aligns with the need to keep data within compliant environments while supporting parallel evals—Legora runs about 2,000 agent evals on Modal sandboxes, according to the source.

Business model shift and scaling

The traditional billable-hour model discouraged speed-enhancing tools, as faster work meant less revenue. Jacob notes that U.S. firms are already moving routine tasks to flat fees or even offering them for free to retain clients, shifting revenue toward M&A, litigation and client-focused time.

Outlook for legal AI

Jacob projects that in two years large enterprises will access Legora through tools like Claude Cowork, while law firms will continue to work directly inside Legora.

Our read: Legora demonstrates that a focused engineering stack, privacy-first infrastructure, and simulation-based evaluation can overcome the data-access barrier that hampers most industry-specific AI efforts. Its blend of legal engineers, forward-deployed engineers, and modular primitives offers a template for other niche verticals seeking to scale AI without compromising confidentiality.