Building Legal Agents When You Can't See the Data — Jacob Lauritzen, CTO of Legora
Summary
Jacob Lauritzen, CTO of Legora, discusses the challenges of building AI agents in the highly regulated legal sector, where access to client data is severely restricted. Legora's approach focuses on building a robust platform around core primitives—such as 'Lists' and 'Tabular Review'—rather than relying solely on raw model intelligence. The company emphasizes a sophisticated engineering model involving 'Legal Engineers' and 'Forward Deployed Engineers' (FDEs) to customize and integrate the product, allowing the platform to scale by solving complex, multi-step workflows like due diligence on thousands of contracts. The discussion highlights the shift from traditional billable hours to project-based fees and the necessity of robust governance and citation systems.
Key takeaways
-
Agent Architecture and Core Primitives
18:42
Legora's platform is built on core primitives (e.g., Lists, Tabular Review, Intake) that allow for flexible, scalable automation. Agents can create lists, and lists can trigger new agents, creating a complex, interconnected workflow system that acts as a 'glue' for various legal tasks. This modularity is key to scaling the product.
-
The Role of Specialized Engineering Teams
27:20
The team structure includes 'Legal Engineers' (lawyers with CS degrees) and 'Forward Deployed Engineers' (FDEs). Legal Engineers configure the product as power users, while FDEs write code into production, enabling custom integrations and complex client-specific workflows.
-
Handling Data Privacy and Governance
34:10
Due to the highly confidential nature of client data, Legora's infrastructure is hosted entirely on Azure. The company cannot access raw user activity traces, necessitating an 'emails team' to approximate user behavior and build features based on indirect signals.
-
Model Strategy and Optimization
22:50
Instead of relying on a single frontier model, Legora employs 'model arbitrage,' using different models for specific tasks (e.g., using specialized models for retrieval or transcription). They are also beginning to fine-tune open-weight models for bounded, high-value tasks like legal research, which is considered a 'fan out retrieval problem.'
Technical details
-
Agent Workflow and Data Processing
1450s
The 'Tabular Review' feature allows users to ask the same question across thousands of documents (e.g., 10,000 employment contracts) and collect the answers in a structured, reviewable table. This process is described as a 'fan out' operation, which is highly efficient for due diligence.
-
System Architecture and Integration
1750s
The platform utilizes a 'List' primitive, which functions like a scalable, customizable database (analogous to a Notion database). This list acts as a central hub, allowing agents to create tasks, and those tasks to trigger subsequent agents, enabling complex automation chains.
-
Advanced Retrieval and Analysis
1800s
The system emphasizes robust search and retrieval augmented generation (RAG). The process involves identifying the correct legal precedent (a search problem) and then ensuring all generated output is traceable via a citation system, which links back to the source document and even specific cells in a table.
-
Model Limitations and Improvements
2150s
While large language models (LLMs) are powerful, the most challenging tasks remain: long document reasoning (e.g., millions of tokens) and complex, multi-step edits across a document. The current state-of-the-art is insufficient for these tasks, requiring architectural solutions like chunking and orchestration.
Mentioned resources
- Legora
- Tabular Review
- Intake
- Lists
- Azure
- AG Grid
Channel & topics
Watch on YouTube · Back to latest
This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.