The Challenge
Roche’s patent organization had built several high-value GenAI tools to support patent research, but as those capabilities expanded, the experience for patent attorneys became increasingly fragmented. Experts were required to switch between multiple interfaces, re-run queries across different systems, and manually rebuild context to reach a complete answer — shifting time and attention away from expert judgment and toward tool navigation.
At the same time, the underlying research methods were not orchestrated together, making it difficult to scale workflows efficiently or extend them without introducing duplication and added cost. As more AI tools were introduced, maintaining consistent governance and oversight for sensitive legal work became harder, increasing operational risk and dependence on external consultants to manage complexity.
Roche needed a way to scale AI that put domain experts first, orchestrated research methods behind a single experience, and embedded governance from the start without slowing teams down or increasing cost.
The Vision
Roche’s goal was to give patent attorneys a single, intuitive interface for every research question — one that could automatically select the right analysis method and preserve context across follow-up questions. The aim was not just to improve usability, but also to reduce time spent navigating tools, improve research completeness, and lower reliance on external consultants by enabling domain experts to extend workflows safely.
At the center of this vision were patent experts themselves, enabling self-service agentic AI without the prerequisite of being an AI or data science specialist.
The Execution
By building Lex as an agentic AI interface in Dataiku, Roche created a scalable foundation for patent research that brings domain expertise, orchestrated intelligence, and embedded governance into a single system — reducing delivery friction, lowering costs, and enabling confident scale.
- Faster delivery to support more people: Lex was shaped by patent attorneys and paralegals who defined real research needs, while citizen developers — Roche team members outside of central IT, enabled by Dataiku’s visual and low-code tools — built and extended the agents in Dataiku to reduce reliance on external consultants and IT. This minimized delivery friction, ensured alignment with real research behavior, and avoided an estimated $375K–$475K in consultancy spend.
- Multiple research methods, one workflow: When an attorney submits a query, Lex orchestrates RAG, full-text search, and deep search behind a single interface in Dataiku, automatically selecting the right research method per query while preserving context across follow-ups. This modular design allows Roche to add new capabilities without rebuilding the interface or fragmenting workflows.
- Controls embedded by design: Access controls, traceability, and C4 scoring and guardrails are embedded directly into Lex workflows, ensuring sensitive legal research can scale safely. Governance is enforced at design time rather than added after deployment, maintaining trust as the platform expands and adoption grows.
Today, the underlying research services support ~100 requests per week across dozens of active attorneys, providing Roche with a production-grade platform to scale patent research with confidence
