Jonathan Sheffi

There's a new question in every health tech buying cycle that didn't exist 18 months ago: "Is this already on the roadmap of our foundation model vendor?" As we discussed in our 2026 healthcare predictions, we are entering a moment where the end user, clinicians and operators included, can increasingly build software themselves using agentic coding environments. In practice, this shows up when a CIO asks whether a proposed point solution is meaningfully different from what their existing foundation model could deliver with internal prompts, guardrails, and a few weeks of engineering time. This shift is quietly reshaping the competitive calculus. Startups selling into healthcare enterprises no longer compete only with the EHR vendor’s roadmap. They also compete with what a foundation model provider can plausibly ship for that customer, often faster and at lower marginal cost. Importantly, while the moat around building software is eroding, the moat around running software in healthcare is not. Healthcare enterprises are not set up to independently handle deep workflow integration, continuous iteration, reliability at scale, and governance across clinical, legal, and regulatory dimensions. That operational gap is where durable companies are still built. We expect this dynamic to introduce real friction into buying cycles for point solutions with thin moats. At the same time, it expands opportunity for companies with defensible data assets, deep workflow entrenchment, and true platform potential. Artera and Atropos Health are two strong examples from Breyer Capital's portfolio, both highlighted in the article. Special thanks to Brian Gormley for featuring our perspective in today's The Wall Street Journal piece on OpenAI and Anthropic's entry into healthcare. Jim Breyer Daniel Breyer Andre Esteva, PhD Brigham Hyde

272

29 Comments

I just got back from the #HLTH Conference where I participated on a panel with my friend Jim Swanson from Johnson & Johnson and moderated by Jessica J. Federer. Here's what we discussed about delivering on AI's promise in pharma and healthcare: AI Fluency is Our Foundation: We're not just deploying tools—we're building AI literacy across our entire organizations. Jim shared how J&J went from 24% to 80% employee adoption in just a year with 150,000 employees now trained. At BMS, we've taken a three-pronged approach: giving all employees free access to frontier models (Claude, GPT, Gemini, even open-source), empowering them through training to experiment, and concentrating our efforts where we can make the biggest impact. We Think in Ecosystems, Not Point Solutions: Jim introduced J&J's lighthouse framework: Layer 1 improves employee productivity, Layer 2 transforms end-to-end processes, and Layer 3 enhances products and services. Our biggest shared mistake? Letting a thousand flowers bloom and getting stuck in "prototype purgatory." We're now laser-focused on core processes where we're seeing transformational results: At BMS in Drug discovery: 100% of our small molecule programs now use AI before any wet lab work—resulting in 25% faster cycle times In Drug development: We've cut database lock timelines by 50% and believe we can shave 1-2 years off clinical trial timelines In Manufacturing: AI helped us achieve 40% yield improvement on Orencia, one of the world's most complex drugs to make We're Rethinking Processes, Not Just Adding Tools: This was probably the most critical insight: In regulated industries, processes become "safe harbors." where we people end up working for processes rather than than the other way around. But to really innovate, you have to do what others aren't doing. We've shifted from "Can AI help with this task?" to "Should this process even exist?" I call it the "clean sheet approach"—starting from first principles rather than asking where to plug AI into existing workflows. The gap between what AI can do at the frontier and what the average employee is doing with it remains enormous. You need both top-down vision and bottom-up expertise—the people who "know how the sausage is made"—working together. Speed to Learning Beats Perfection: As Jim said perfectly: Today's AI is the worst AI will ever be again. Our traditional 6-12 month implementation cycles are obsolete. We're getting from question to learning in weeks, not months. But you have to be thoughtful—there's a lot of snake oil out there, and if you don't know what's real, you'll waste time and money. The Bottom Line: We're hitting singles and doubles when patients need us to hit home runs. AI isn't our hardest problem—it's process redesign and mindset change. The ability to break out of "thou shalt always follow" thinking is essential to our industry's future. We're not asking where to plug in AI—we're redesigning work from first principles.

554

24 Comments

The result I'm most excited about from Arc's new State model: The ability to generalize on zero-shot out-of-distribution predictions after pre-training on the TAHOE-100M data set. Whereas PLMs have seemingly benefitted less from scaling data and model size, this is an inkling of evidence that we should expect something quite different for perturbation predictions models and virtual cells at-large. Lots more about virtual cells coming to your inbox via CoB very soon... 👀🧬 State manuscript: https://lnkd.in/gt9URsnq

82

🔥🔥 PANEL ALERT 🔥🔥 What gaps may AI fill with the recent policy impacts on Medicaid?  Coyote Ventures 2025 Health Equity Innovators Fall Gathering will feature an expert panel discussing the impact of AI in Healthcare with a particular focus on its implications for Medicaid and Medi-Cal. As we are past the mid-point mark of 2025, AI and OBBBA are reshaping U.S. health policy in profound ways, from expansive new federal and state guidelines to sharp debates around oversight and authority. Our discussion will dive into the themes of bias, equity, and accountability in AI, and what these evolving mandates mean for those developing, deploying, or relying on AI-driven solutions in healthcare. With Medicaid and Medi-Cal covering some of our most vulnerable populations, the stakes couldn’t be higher. I’m looking forward to an engaging conversation with Priyanka Vaidya, Pooja Mittal, and Stella Tran - exploring how we can ensure AI not only advances innovation but also builds trust, fairness, and better outcomes. Curious to hear your thoughts: can AI truly help reduce disparities in Medicaid care, or will it deepen them?

63

6 Comments

ChatGPT Health just made the "Middle" a commodity. With today’s launch, the intelligence layer (analyzing health data) is no longer a startup opportunity. It’s a utility. If your value prop is "we use AI to analyze your health data," you are now competing with a free, ubiquitous giant. You face a data access problem on the input side and a delivery problem on the output side. The "Universal Health OS" is finally here. The future is coalescing all inputs into one engine (ChatGPT Health) to drive real-world outputs. For the consumer, the magic is simple: Talk to the AI, give it context, and get clear "comprehensive" guidance. For investors and builders, the alpha has moved to the Edges: - The Inputs (The Sensors): The AI is blind without unique data. The opportunity is in proprietary capture. Think novel biomarkers and continuous monitoring. - The Outputs (The Solutions & Rails): The AI is paralyzed without real-world interventions. The value is in the independent product itself (the breakthrough medication, the specific supplement, the device) AND the fulfillment (shipping the food, booking the protocol). The barrier to entry for intelligence has vanished. Now, the race is on for execution. Don't build the brain. Build the nervous system. P.S. I wrote about this thesis a while back. The "Gilded Age" of health tech is about infrastructure, not just AI models. Deep dive here: https://lnkd.in/gV8aQycu #HealthTech #ChatGPT #DigitalHealth #Innovation #VentureCapital

194

34 Comments