About This Episode
In Season 2 Episode 4 of Stratola Spectrum, Dinesh Chandrasekhar sits down with Dr. Ron Razmi—ex-cardiologist, former McKinsey consultant, health AI investor at ZOI Capital, and author of AI Doctor: The Rise of AI in Healthcare—to examine what is actually happening beneath the surface of healthcare AI and why the gap between promise and reality is wider than most people admit. The conversation does not begin with technology, but with a simple and striking observation: despite years of investment in AI for radiology, the shortage of radiologists is worse than ever. AI is assisting, not replacing, and the industry continues to blur that distinction.
The discussion moves through several critical themes, including why large language model (LLM) accuracy drops sharply—from around 95% in controlled environments to as low as 20% in real clinical settings—once exposed to messy, incomplete, and inconsistent real-world data. It challenges the common assumption that doctors need help with diagnosis, arguing instead that the real opportunity for AI lies in addressing the overwhelming administrative burden surrounding care, such as documentation, coding, claims processing, prior authorizations, and denial management. The conversation also explores how the B2C2B model is quietly solving healthcare’s notoriously slow enterprise sales cycle, and why the biggest bottleneck in deploying AI is no longer technology but organizational resistance, unclear ownership, and legal complexities.
A significant portion of the episode focuses on the data fragmentation problem at the core of healthcare AI. Dr. Razmi shares a personal story about his father nearly receiving an unnecessary pacemaker due to a missing medication record from a clinic in Las Vegas. This example underscores a critical reality: in healthcare, partial data is not merely incomplete—it can be dangerous. Unlike other industries where missing data leads to suboptimal recommendations, in healthcare it can lead to incorrect clinical decisions with serious consequences. Yet, many AI systems today operate on incomplete datasets without awareness of these gaps.
The episode also sheds light on the commercial challenges that many health AI founders underestimate. Established players like Epic can stall a startup’s growth without even launching a competing product—simply by signaling that a similar feature is in development. Combined with enterprise sales cycles involving numerous stakeholders and veto points, even strong products can remain stuck in evaluation for years. Ultimately, this episode is not a speculative look at the future of healthcare AI, but a grounded examination of its current reality—one that builders, buyers, and investors must understand before making their next move.
Key Takeaways
AI accuracy in healthcare drops dramatically in real-world conditions. While benchmark results on structured datasets may reach 94–95%, studies from Nature Medicine and Microsoft show accuracy falling to 20–30% in real clinical environments due to fragmented and inconsistent data.
Doctors do not need help with diagnosis—they need help with everything around it. Most diagnoses are already known before patient interaction; the true bottleneck lies in administrative workflows.
Partial data in healthcare is dangerous. Even with 99% of a patient’s information, a single missing detail can lead to catastrophic decisions, making data completeness critical for AI systems.
Incumbents can block innovation without building. Companies like Epic can delay startup adoption simply by suggesting future competing solutions, effectively freezing customer decisions.
The B2C2B model is transforming enterprise adoption. By first gaining traction with individual clinicians, startups can create internal demand that accelerates institutional sales.
Most digital health companies do not reach an exit. With only about 2% achieving successful exits, challenges such as fragmented data, long sales cycles, reimbursement issues, and shrinking margins make early strategic awareness essential.