The pharmaceutical industry has long grappled with the cognitive dissonance of drug development—the tension between data-driven insights and the deeply ingrained gut feelings of experienced trial designers. As clinical trials grow in complexity, the reliance on fragmented digital tools and legacy benchmarks has often led to operational bottlenecks and unpredictable recruitment cycles. To address this, Novo Nordisk is deploying Study Hub, a comprehensive digital suite designed to harmonize data across the entire clinical development lifecycle.
Guillaume Carbonneau, Vice President of Operational Data Insights at Novo Nordisk, is at the helm of this transformation. By implementing advanced ontologies and agentic frameworks, Carbonneau is shifting the organization from subjective decision-making toward cognitive intelligence. In this interview, he discusses the challenges of building trust in AI-driven predictions, the technical hurdles of entity resolution, and why the future of clinical operations depends on a unified semantic language.
Moe: Looking back at the early development of StudyHub, can you share a specific moment where data insights directly challenged long-held internal gut feelings regarding trial design or site selection?
Guillaume Carbonneau: You mentioned cognitive dissonance, what we’re aiming for with StudyHub is transform towards cognitive intelligence. Our goal is to eliminate information bias as soon as a decision-maker steps into the complex operational process of designing, planning and executing clinical trials. In the past, pharma companies used various digital tools, but they were often isolated and only addressed one or two steps in the process.
You might have excellent data for site selection, but the challenge comes when you try to connect that information to the early stages of a clinical development plan—like study design, endpoints, and I/E criteria. We realized that while we had tools before, they were treated more like optional add-ons rather than the main resources for making final decisions. So, we decided to take a holistic approach, connecting clinical development, operations, and individual countries from the early protocol stages all the way through to trial monitoring.
Moe: Your strategy involves a pragmatic layer in your ontology to define AI models in a functional context. In terms of deep entity regression, what has been the most difficult operational variable to quantify regarding its impact on recruitment cycles?
Guillaume: Proposed: To provide insights that everyone can truly rely on, we need to ensure that the human organization trusts the system. We’ve developed an Ontology model with two key components. The first is a semantic layer, which standardizes how the organization comprehends business decisions, KPIs, and metrics. By virtualizing these decisions, we ensure that every unit communicates in the same language.
Then comes the pragmatic layer, where we build the business logic and AI models—translating data into predictions, recommendations, or benchmarks. To keep that trust intact, our models must be transparent. We utilize AI agent frameworks and large language models (LLMs) tailored to the specific context in which the data is applied. It’s not just about the math; it’s about making the logic behind a prediction clear and understandable to the user.
Moe: When synthesizing disparate sources like investigator histories or naming conventions, the data is often noisy. How do you handle cases where these sources don’t align with the ground truth at the site level?
Guillaume: The ontology really shines here because it pulls together scattered datasets from all over the industry—whether it’s external sources, internal R&D data, or even medical and commercial information. When you combine all this, you often find that different datasets actually contain the same piece of info about an object, be it a healthcare professional, a clinical site, a patient, or a specific disease.
We use entity resolution to break down this data into billions of nodes, figure out their relationships, and then put them back together into clear, defined objects. Over the past year, we’ve been using LLMs and specialized agents to bump up the accuracy of this extraction. It’s not perfect yet, but we’ve reached a level of accuracy that lets us understand a competitor’s protocol and compare it to their operational outcomes or scientific publications. Our Ontology Core is the mother of our insights.
Moe: Clinical teams are traditionally risk-averse. What is the most surprising behavioral barrier you’ve encountered when asking leaders to trust a digital brain over legacy benchmarks?
Guillaume: Technology evolves faster than organizations can keep up. In our personal lives, we’re immersed in new tech constantly, but in the workplace, any suspicion about a tool can undermine the trust needed to rely on it. It’s all about crafting the right user experience and delivering consistently reliable insights. At Novo Nordisk, we’re gradually building this trust together, demonstrating that these models outpace and outperform old methods. We started with small expert groups of 20 to 50 people, and now we’re reaching around 3,000 users. The current challenge is scaling our educational programs so that everyone can fully benefit.
Moe: As your agentic framework becomes more autonomous in suggesting trial pathways, how do you define the boundaries of human-in-the-loop oversight to ensure speed doesn’t come at the expense of clinical empathy?
Guillaume: We’re trying out some new ways to help different groups of people work together more quickly on analytical tasks. Right now, these tasks often use different methods, which means the results can vary. We will be exploring how planning agents can help our organizations stick to a consistent method and get faster to decision – aiming to reduce clinical trial cycle internal time.
We do not aim making the final decision, but by offering a scenario that’s 80% ready for a site selection or allocation plan as early as at Study Outline. We also use agents behind the scenes to clean data, in the middle to generate recommendations, and soon up front to save our clinical operations teams’ time. Our goal is to speed up the process from the initial human dose to commercialization.
Moe: What is one specific lesson you’ve learned from a pilot that failed to meet its predictive accuracy, and how did that reshape your integration of external data?
Guillaume: I believe predicting outcomes perfectly is nearly impossible, but making decisions based on data certainly boosts our confidence in hitting targets like patient recruitment, critical milestones or staying within budget. It cuts out the guesswork. If things don’t go as planned, it’s usually down to having an unrealistic plan or disjointed execution.
With StudyHub, when our plans are backed by solid data, facts, and historical insights, we aim to shift the organization from complaining about visibility issues to focusing on corrective actions. And if our predictions miss the mark, we have the mathematical groundwork to understand why and adjust our strategy to tackle the problem head-on.
Moe Alsumidaie, MBA, MSF, is founder and Chief Editor of Vanguard Publications, which publishes Clinical Trial Vanguard, Pharma Vanguard and BullScope, and Head of Research at CliniBiz. He has two decades in clinical trial operations and data science, with earlier roles at Genentech, Abbott Vascular and Stanford University Medical Center, and is a guest lecturer in clinical trial sciences at Rutgers University.
