Manish Balakrishnan

Hot take: betting or investing in any niche EHR right now (e.g. AI-native EHR for X clinical specialist or X vertical like Longevity) when AI code development exists and only continues to get better with each new model release has to be dumbest pre-AI SaaS era thing an investor can do. Here's why ⤵️ The EHR market is essentially split into 5 categories: #1. Outpatient EHR-enabled private practice (doctor or clinical owned) #2. Outpatient tech-enabled practice (PE, VC, family office owned) #3. Outpatient tech-enabled practice (MBA or bootstrapped owned) #4. Startup enterprise (PE or VC backed) #5. Inpatient hospital VC backed EHRs primarily make their money selling to #2-4, with the Epic's of the world taking most if not all of 5. For #1 there really isn't any true TAM since it's difficult to acquire this customer even when giving the EHR product away for free. But here's what these EHRs are not telling you and why they're all racing to update their API docs. They hope that #2-4 will continue to use their solutions to build with AI vs exiting and vibe coding their own in-house solutions. So when you hear them say their EHR product is safe from AI, that's not true. Right now the Epics of the world are safe but it's customers in categories 1-4 who will 100% vibe code their own solutions. #VibeWithCline #LetDoctorsVibe

24

21 Comments

As LLMs enter the healthcare niche more directly, they introduce a new temptation: anchoring an entire AI strategy to a single smartest model or provider. It feels simpler - one provider, one interface, one roadmap. But how well does that hold in practice? Healthcare workloads are not uniform. Some require speed, others demand traceability. Some require multilingual support, others require extreme caution. No single model is best across all of these dimensions. And even if one appears to be today, that advantage will not remain static. This is why portfolio architectures consistently outperform single-model strategies. In a portfolio approach, foundation models are selected per task rather than by brand. An orchestration layer routes work based on risk, cost, language, and latency. Connectors tie into systems of record and institutional knowledge. Channels reflect where staff and patients already operate. Governance and evaluation span the entire system rather than living inside a single tool. This is a resilience-focused design. Direct vendor tools absolutely make sense for narrow, well-contained use cases, especially when speed matters and the blast radius is small. But as soon as AI touches multiple workflows, departments, or risk levels, flexibility becomes a requirement rather than a preference. Single bets feel efficient until pricing shifts, policies change, or new regulatory or operational constraints appear. Portfolio architectures survive those moments without forcing organizations to rebuild from scratch.

16

3 Comments

Stack Overflow’s decline is a great example of the crazy evolutionary pressure exerted by LLMs. Coding answers live inside tools now and their forum traffic fell off a cliff (and with it advertising revenue). They pivoted to selling years of forum posts as training data input to LLM vendors and a custom edition as a B2B coding assistant but their proprietary data is going to get stale pretty soon. Interesting times!

10

4 Comments