Novo Nordisk: a governed reasoning agent on clinical trial data

Getting insight from clinical research data depended on scarce specialists and slow workflows. Teams could spend months deciding whether a hypothesis was worth pursuing, and sometimes learned late that the evidence was not strong enough.¹

Working with Microsoft forward deployed engineers from the AI Acceleration Studio, Novo Nordisk built what it calls a governed reasoning agent, or quantitative decision-support agent, on Microsoft Azure. The agent writes and runs code and performs statistical analysis on structured clinical datasets. Biostatisticians and subject matter experts review its output before it informs decisions.

The agent draws on FounData, Novo Nordisk's harmonized store of more than 200,000 patient-years of clinical trial data aligned to the CDISC, SDTM and ADaM standards. Before production, the teams ran thousands of automated tests and layered evaluations covering query validation, performance, expert review and user feedback. Databricks separately describes FounData as using a Databricks data mesh architecture on Azure, and a different multi-agent Co-Scientist system built with Databricks tools; the two vendor stories describe different systems.¹²³

In our view, the lesson for regulated R&D teams is the discipline before launch: thousands of automated tests, layered evaluation and human sign-off before an agent touches decision support. The gains depend on years of data harmonization to industry standards, which most organizations would need to do first. The headline capacity figure is a statement by one manager without a published method, and the time-to-insight gain is an expectation. Teams without harmonized data will get more value from data standardization than from an agent.