What it takes to bring an AI research assistant to thousands of medical experts
Microsoft and Novo Nordisk rebuilt a clinical trials data agent on Microsoft Foundry to streamline secondary analysis for medical experts, improving session isolation and operational efficiency.
A joint Microsoft and Novo Nordisk team rebuilt a clinical trials data agent on Microsoft Foundry to assist medical experts in secondary analysis of trial data. The new hosted agent model simplified operations, strengthened session isolation, and integrated directly with Foundry’s tracing, evaluation, and optimization tools. This shift enabled experts to explore data flexibly, test hypotheses, and visualize trends without rebuilding the underlying infrastructure. The project demonstrated how agent-based tools can adapt to evolving research questions while maintaining control over security and quality.
The original agent ran in Azure Container Apps with PostgreSQL for conversation storage and dynamic sessions for code execution. This architecture required the team to manage container scaling, session lifecycle, and networking, adding operational overhead. Foundry’s hosted agent model shifted responsibility for endpoint management, session compute, and lifecycle to the platform, while retaining control over orchestration logic and tools. Each session operates in a VM-isolated sandbox with on-demand compute, reducing the need for manual capacity planning and improving scalability.
Key infrastructure lessons emerged during the migration, including the importance of private networking early in the design phase. Foundry’s subnet sizing guidance recommends a /24 for production deployments, with /27 as the minimum for smaller setups. The team also emphasized using Azure Verified Modules for repeatable environments and aligning database throughput with expected concurrency. These decisions collectively shaped the system’s reliability and scalability ceiling.
Observability and telemetry were central to the project, with Foundry tracing capturing user inputs, model calls, tool activity, and latency via OpenTelemetry and Azure Monitor Application Insights. The team maintained dual telemetry paths during migration to ensure continuity while connecting to Foundry’s capabilities. Stable identifiers for users, sessions, and traces enabled efficient debugging, while strict access controls and minimization practices were applied to protect sensitive clinical data.