Novo Nordisk accelerates clinical insight with custom agents on Azure

As competition accelerates and pharmaceutical R&D costs continue to rise, Novo Nordisk faced a growing challenge: how to extract more value from decades of clinical research data fast enough to influence critical scientific and business decisions.

The company had already spent years building a cloud-based research foundation through its FounData initiative, harmonizing more than 200,000 patient-years of clinical trial data across studies, disease areas, and research programs. Because the platform is aligned to open industry standards—CDISC, SDTM, and ADaM—analyses can run on data that is already prepared, without bespoke pre-processing. But while the data foundation existed, accessing insights from it still depended heavily on scarce specialist expertise and time-intensive workflows.

"The pharmaceutical industry is under enormous pressure to improve productivity," says Sid Prabhu, who leads the FounData initiative within Novo Nordisk's AI and Digital Innovation organization. "We're spending more and more on research, while competition is moving faster than ever. Every decision matters more."

Inside pharmaceutical R&D, answering a single scientific question can require coordination across medical experts, commercial stakeholders, project managers, data scientists, and highly specialized biostatisticians. Teams may spend months evaluating whether a hypothesis is even worth pursuing—often only discovering late in the process that there is insufficient evidence to justify moving forward.

In pharmaceutical development, delayed or poorly informed decisions can impact trial timelines, investment allocation, and the probability of new therapies reaching patients. Failed trials can also represent significant lost investment, making early confidence in decision-making critical. Novo Nordisk saw an opportunity to change that dynamic by operationalizing AI directly inside scientific decision-making workflows.

Working with Microsoft forward deployed engineers, Novo Nordisk developed a quantitative decision-support agent designed specifically for pharmaceutical R&D. Built on Microsoft Azure, the solution helps researchers and medical experts independently explore hypotheses against Novo Nordisk's proprietary clinical datasets using AI-generated code execution and statistical analysis—with human validation built into the process from the beginning.

"We wanted to move from gut-feel decision-making toward quantitative decision support. If we can validate ideas earlier, fail faster when necessary, and prioritize stronger opportunities sooner, that changes the economics of pharmaceutical development,” adds Prabhu.

Unlike a traditional chatbot, the system does not simply retrieve answers or summarize documents. The agent helps researchers generate and execute code, work with structured clinical datasets, and explore statistical workflows, providing explainable outputs to support exploratory scientific reasoning.

Researchers can ask complex scientific questions in natural language—such as whether a particular patient subpopulation demonstrated a specific treatment response—and receive data-driven exploratory analyses in minutes, rather than waiting weeks for manual workflows to complete.

"What surprised people internally was not just that the system could generate answers," says Prabhu. "It could reason through problems, identify confounding factors, and surface insights in ways that made sense to scientific experts."

Rather than approaching the project as a proof of concept, Novo Nordisk and Microsoft focused from the beginning on enterprise deployment, moving beyond isolated experimentation toward governed, production-ready workflows.

Microsoft helped Novo Nordisk operationalize reasoning agents inside a regulated research environment by providing the infrastructure, governance patterns, and scalable orchestration needed to deploy AI safely against proprietary clinical data.

"You don't just connect an LLM to enterprise data and hope it works," Prabhu says. "There's a tremendous amount of infrastructure, governance, testing, and evaluation required to make something like this production-ready in a regulated environment."

Before production rollout, the teams executed thousands of automated tests and implemented layered evaluation systems combining query validation, system performance metrics, expert review, and continuous user feedback. The solution also incorporates human-in-the-loop workflows that allow biostatisticians and subject matter experts to validate promising outputs before they influence major scientific decisions.

"The agent is there to inform decisions, not to take over control. It drafts the analysis; the scientist decides what to do with it. That distinction is foundational to how we built it," says Rasmus S. Andersen, Associate Director, Products and AI, FounData.

That governance model proved especially important in building organizational trust. Early in the project, some stakeholders questioned whether an AI system could reliably participate in high-stakes scientific workflows. But as researchers and biostatisticians began using the system, skepticism shifted toward adoption.

"People initially said this would be impossible," Prabhu says. "Now the challenge is managing demand because everybody wants access to the tool."

The business impact is already beginning to emerge. Novo Nordisk expects the system to reduce time to insight from weeks to minutes for many exploratory analyses while increasing the number of scientific questions teams can evaluate.

"In the past, we might have had the capacity to pursue 5–10 strong ideas in a quarter," says Prabhu. "Now we can evaluate 50+."

"What's powerful here isn't speed alone—it's that we can now explore far more scientific questions earlier, with greater productivity, efficiency, and scientific quality while maintaining the rigor our work demands. Ultimately, that increases our probability of success at the very front end of drug development," says Mishal Patel, GVP AI & Digital Innovation, Novo Nordisk.

The company is also targeting improvements in trial probability of success—a critical metric in pharmaceutical R&D, where failed studies can cost hundreds of millions of dollars over multiple years. Even modest improvements in decision quality can have significant downstream scientific and financial implications.

Importantly, Novo Nordisk positioned the system as an augmentation tool for experts, not a replacement for them. Rather than eliminating biostatisticians from the process, the goal was to eliminate low-value exploratory work and allow specialists to focus on the highest-priority opportunities.

"The goal was never to replace scientific expertise," Prabhu says. "The goal was to help our experts spend more time on the decisions that matter most."

With that focus, Novo Nordisk is now expanding the solution beyond its initial deployment. With the agent already live internally and broader rollout underway, the company is exploring new applications in trial design optimization, portfolio intelligence, and additional decision-support workflows across the organization.

Looking further out, the roadmap shifts the agent from descriptive analyses—telling teams what the data shows—toward predictive capabilities, including outcome modeling and trial-scenario simulation. It will also reason across a broader set of data modalities, extending beyond today’s clinical and real-world data to incorporate omics, device telemetry, and external scientific literature.

The broader significance may extend beyond pharmaceutical research. As highly regulated industries look for ways to operationalize AI safely, Novo Nordisk’s approach demonstrates how governed reasoning agents can accelerate complex decision-making while preserving human oversight, traceability, and trust.

Discover more about Novo Nordisk on Facebook, Instagram, LinkedIn, X, and YouTube.