Build or Buy Healthcare AI? Five Lessons From Merck

Every AI conversation in Life Sciences reaches the same moment: “We can build this ourselves.” 

They’re right, to a certain point. With today’s foundational models, a capable team can stand up a compelling prototype in a few weeks. The bigger challenge is making it trustworthy enough for the demands of a regulated enterprise.

Watch the Webinar On Demand

In Komodo HealthⓇ’s recent webinar, Ritu Saxena, PhD, Merck’s Senior Director of AI Agents, and John Wollman, General Manager, Marmot™, at Komodo, shared what that journey actually requires. Below are five key takeaways from their discussion:

1. The demo takes weeks. Trust takes months. Komodo had a working prototype, running on real data, in about three weeks. Getting Marmot to full enterprise readiness took far longer, with much of that time spent testing alongside clinicians. As John put it, the real question for any team deciding between in-house development or going with a vendor is “whether you want to spend 16 months earning trust, or start with it.”
2. Enterprise trust has five requirements. Healthcare analytics demands answers that executives, clinicians, regulators, and compliance teams can all stand behind. Komodo frames this as the five pillars of AI trust: grounded, explainable, reproducible, verified, and governed. John explains why the typical MVP meets one or two, and which pillar is hardest to earn.
3. Build vs buy turned out to be the wrong question for Merck. The answer was both. After a year of building in-house, Merck chose to accelerate and own its AI integration by partnering with Marmot’s healthcare context and proven technology, while retaining control of its roadmap, data strategy, and governance.
4. At Merck, the real shift is from searching to conversing. Its first commercial agent is currently in production with about 200 users. Ritu shares how it’s tangibly changing the way marketers and analysts work. “The biggest change is moving from searching for information to having a conversation with it,” she said.
5. Start with the problem.  Ritu’s advice for peers weighing the same decision was simple: “Start with the problem. What are you actually trying to solve?” She walks through the six-part framework her team used to make the call.

The live Q&A went further, covering deployment models, compliance guardrails, and Marmot’s handling of customer data.