Meet the co-founder of Vega Health: Dr. Mark Sendak

While AI has the potential to drastically improve the outcomes and cost of healthcare, the impact of machine learning (ML) is still just a promise. AI transformation projects inside hospitals rarely survive the pilot stage, because fast-moving startups selling one-size-fits-all ML models often lack the people, technology, and patience to navigate the complex processes and fragmented data sources of a hospital system. 

At Bessemer, we believe that the winning healthcare AI company will look less like a typical software company and more like a full-stack partner. One that sources AI solutions from the innovators closest to care delivery, builds the infrastructure to make it run, and holds itself accountable until it actually works for their customers. To launch such a venture, we ambitiously resolved to find an AI scientist with a proven track record of successfully deploying and scaling AI models in hospitals. This is how we met Dr. Mark Sendak. 

Dr. Sendak spent over a decade at the Duke Institute for Health Innovation building clinical AI, an effort that grew into DIHI—now running more than 50 models. When Bessemer’s Lance Co Ting Keh reconnected with Mark a decade after they studied ML together at Duke, it was clear that what DIHI had built for Duke needed to exist for the rest of healthcare. Lance and Mark co-founded Vega Health, a full-stack AI solutions platform that partners with health organizations. They launched the company alongside Partner David Cowan, who has a long history of incubating companies at Bessemer. We sat down with Mark to talk about the insight behind Vega, how the company has earned trust in a skeptical market, and why 93% of U.S. hospitals are the most overlooked, yet important opportunity in healthcare.

What was the founding insight behind Vega Health?

Our founding story starts with a failure—or at least, what felt like one at the time. I spent over a decade at the Duke Institute for Health Innovation building AI solutions for real clinical problems. One of them was Sepsis Watch, an early sepsis detection model that has been running continuously at Duke-affiliated hospitals for nearly a decade. We tried to commercialize it, but our first external implementation faltered. The same model that had saved lives in one setting stalled at another, not because the technology failed, but because everything around it did.

We eventually tried solving the AI implementation problem in community and rural settings through a non-profit collaborative, Health AI Partnership. While we built a real community across institutions, we couldn't crack the last-mile implementation because the best practices and community weren’t enough.

These experiences clarified something I hadn't been able to fully articulate before: the barriers to scaling effective healthcare AI aren’t technical. What fails is the distribution infrastructure, the implementation support, the workflow integration, and the incentive structures that shape how vendors serve provider organizations.

When I learned what wouldn’t work, I finally had a clear enough picture of what would. Not another point solution, and not another marketplace without real financial upside and accountability for solution developers. The healthcare industry needed a full-stack partner: platform infrastructure, a curated marketplace of validated AI solutions sourced from those closest to care delivery, and the implementation and monitoring muscle to make AI work in a health system's environment. 

How do you build trust in a crowded healthcare space, especially when some of your customers have already suffered from failed pilots or put their trust in the wrong vendor?

So many health systems have been burned by AI vendors, and surprisingly, that’s where we’ve gained early traction. Our first two customers came to us with real scar tissue. One had worked closely with a revenue cycle management (RCM) vendor that sold a grand vision and then largely disappeared, offering no workflow integration and no effort to understand how the existing team operated. With the other, their front-line clinicians and IT staff had spent years trying to get a clinical decision support tool to work. In both cases, technology wasn't the problem; it was poor implementation support, and minimal accountability for results.

When we talk with prospects, we lead with curiosity, not a demo. Recently, our team was on-site with a partner scoping out a solution for medication history and reconciliation. We spent most of the trip interviewing and observing informatics leaders, front-line workers, and clinical and operational champions. One administrator was visibly wary of yet another AI vendor walking through the door. Half an hour later, they were advocates. Not because we showed them something impressive, but because we asked the right questions to be the most helpful to them.

Health systems suffer from failed pilots partly because the incentive structures behind most vendors are missing accountability after the contract is signed. Vega Health’s business model specifically counters that. We commit to solving a customer’s pain point and then install our platform, testing which model addresses the problem most effectively. Our monitoring framework tracks technical accuracy, clinical adoption, impact on outcomes, and ROI. We’re also honest about when AI might not be the right solution for a customer’s problem. That conversation is uncomfortable, but it’s also the reason health systems learn to trust us with their most critical problems.

You mentioned the healthcare industry needing AI solutions sourced from those closest to care delivery. How does that difference show up compared to AI built from a distance?

The real difference shows up before a single line of code is written because it determines who defines the problem.

Most healthcare AI is built by people who understand machine learning deeply and healthcare only at a distance. They identify use cases by analyzing easily accessible datasets, conducting market research, engaging in C-suite conversations, and pattern matching against adjacent industries. That process produces (mostly) technically competent solutions to problems that are sometimes real and sometimes assumed. The gap between AI built by outsiders looking in and insiders solving their own problems is enormous, and most health systems struggle to discern how well a technology will solve their problems.

AI built by those closest to care is different. The problem definition comes from the people living it: the nurse who knows exactly when cognitive load peaks, the physician who can tell you what information is missing at the moment a decision has to be made, the RCM employee who struggles to find relevant information from payer policies. That knowledge doesn't live in any dataset or product document; it lives in the minds of the people doing the work. The innovators whose models we license have contextual knowledge about the problem they solve. We earn access to these innovations because we deeply value and respect that contextual knowledge.

The biggest indication of whether an AI solution will be adopted isn't its accuracy. It's whether the people being asked to change their behavior were part of shaping the solution and whether they trust those who built it. 

How did you build out your GTM strategy? And what would you tell a founder just getting started in healthcare?

We targeted an underserved market of community hospital systems. In healthcare, most AI companies fight over the same 400 hospitals affiliated with elite universities. Our go-to-market focuses on the 93% of American hospitals that no one is taking seriously for innovation. We're building for them, and the early numbers suggest that the market is ready to move faster than conventional thinking would have you believe. 

I began this journey as a clinician and an innovator, not a salesperson. My version of go-to-market for most of my career was: build something that works with people at my organization and help those same people get the most value out of using what I built. If we were successful, we would publish our findings, present our research at conferences, and hope that word got out about how amazing our team at Duke was. It’s a good strategy for academia, but unconventional for starting a company.

Two things I've learned that I didn't expect coming in:

• The community hospital market thesis has been validated faster than I anticipated. We believed that these hospitals, largely ignored by most healthcare AI vendors, represented a significant portion of the market. The prior consensus was that the community hospital system market was too fragmented, small, and slow. Our experience has debunked those beliefs. This thesis is now a foundation for driving our growth, and no longer just a hypothesis.
• The second thing I learned is that clinical credibility is an operational asset. We bring our experience building AI solutions with front-line workers and supporting community hospital systems into every room we enter at every stage of our customer journey. We show up to do the hard work of helping our partners realize as much value as possible from their investments in AI.