Adam Odessky

Excited to finally share our investment in Onboard AI, which just announced a $5M raise and a strategic partnership with Cedars-Sinai. Here's the problem: Healthcare is being flooded with AI solutions—from clinical decision support to operational tools—but most hospital systems are stuck with manual reviews, static checklists, and months-long backlogs. They're forced to choose between moving too slowly or approving AI tools without appropriate controls. What makes Onboard AI different: It's AI-native from the ground up. The platform reads and scores vendor documentation in minutes—not weeks—and continues monitoring AI tools throughout their lifecycle for ongoing safety and compliance. The opportunity to help shepherd responsible AI adoption in healthcare is enormous. We're excited to be supporting Troy Bannister and the Onboard AI team with our friends Scott Law at Meridian Street Capital to build the critical infrastructure layer the industry needs to move forward with confidence. More on why we invested and what's next: https://lnkd.in/d9hnageg #ai #healthcare

69

1 Comment

Three incredible days at #JPM2026. Same conversation with 10+ health system execs: "We're building our AI strategy. Evaluating vendors. Need to pick the right partners." This sounds familiar. 2009-2012: The EHR decision. After the HITECH Act, health systems spent 18-24 months evaluating Epic vs. Cerner. Consultants. Committees. RFPs. Multi-year implementations. The logic made sense: EHRs were monolithic platforms with huge switching costs. You had to get it right. 2025: The AI decision. I'm watching health systems do the same thing. "AI steering committees." "Strategic AI partner" RFPs. Trying to pick winners. But there's a critical difference: EHRs were a single, massive platform decision. AI in healthcare is dozens of different problems that need dozens of different solutions. Here's what health systems are missing: They're looking for "the AI vendor" when they should be building "an AI portfolio." There's no single AI platform that solves:   → Documentation burden (ambient notes) → Radiology throughput (workflow optimization) → Clinical trials matching → Revenue cycle optimization (across specialties) → Clinical decision support These are different problems. Different technologies. Different vendors. Different timelines. Some need FDA clearance and 24-month validations. Some can pilot in 90 days. Some integrate with Epic. Some replace parts of Epic. You can't solve all of these with one "strategic AI partnership”....no matter that Anthropic is saying right now. <ducks> The AI landscape is evolving too fast for monogamous vendor relationships. 6 months ago: GPT-4 for clinical documentation. Today: Ambient listening, agentic workflows, multimodal imaging analysis. 6 months from now: Something else entirely. That's why you need exposure across the whole landscape—not a bet on one vendor. The mistake I see health systems making: Approach 1: "Strategic Partner" Strategy Pick one AI vendor → Enterprise deal → 3-year commitment → Hope they solve everything Result: You're locked in. Technology moves on. You're stuck in vendor negotiations while competitors move faster. Approach 2: "Portfolio Strategy" Deploy capital across multiple companies → Take equity → Co-develop solutions → Maintain optionality Result: You capture upside. You shape products. You're first to new capabilities. You have relationships across the ecosystem. After 20 years building innovation platforms at JPMorgan, Mousse Partners, Blueprint Health, Junto Health, Atlantic Health: In fast-moving markets, portfolio strategies beat single-vendor bets. Healthcare AI is moving fast. The question isn't "Which AI vendor should we pick?" The question is: "How do we build an AI portfolio that gives us the most optionality" If you're thinking about this, let's talk. 📩 douglas.m.hayes@gmail.com ☎️ 215-620-2032 #healthcareAI #digitalhealth #venturecapital #innovation

293

25 Comments

Built a Clinical RAG System: Bridging AI and Healthcare 🏥 I'm excited to share my latest project—a Retrieval-Augmented Generation (RAG) system designed specifically for clinical applications! 🔍 The Challenge: Traditional LLMs can hallucinate medical information, which is dangerous in healthcare. How do we make AI responses accurate, transparent, and grounded in real clinical data? 💡 The Solution: I built a RAG pipeline combining: ✅ FAISS vector search for semantic retrieval ✅ LangChain for orchestration ✅ Google Gemini AI for response generation ✅ Streamlit for an intuitive interface 🎯 Key Features: • Searches through clinical notes using embeddings (all-MiniLM-L6-v2) • Retrieves relevant context before generating answers • Shows users exactly which documents informed the response • Built-in safety mechanisms to prevent hallucination • Responds with "Not enough context" when uncertain 📊 Why This Matters: In healthcare, accuracy is everything. RAG systems provide: → Transparency: Users see the source of information → Accuracy: Responses grounded in actual clinical documents → Scalability: Easy to update knowledge base without retraining → Trust: Medical professionals can verify AI responses 🛠️ Tech Stack: • Python • FAISS (vector database) • LangChain • HuggingFace Transformers • Google Gemini AI • Streamlit 💭 Key Takeaway: RAG isn't just about better chatbots—it's about building AI systems that are transparent, trustworthy, and safe for critical applications like healthcare. 🔗 Full technical walkthrough on Medium: https://lnkd.in/dWWPHMWk 💻 Code on GitHub: https://lnkd.in/dsEkASRZ What do you think about AI in healthcare? Would love to hear your thoughts! 💬 #AI #MachineLearning #Healthcare #RAG #NLP #DataScience #MedicalAI #ArtificialIntelligence #TechForGood #Python #DeepLearning #Innovation

25

4 Comments

As AI is increasingly used to analyze clinical decisions after-the-fact, AI may be used more by providers and payers to protect their own interests rather than to help patients. Dr. Eric Topol, MD and colleagues recently warned how this could play out: → Clinicians are constantly under pressure to make fast decisions with imperfect information. → When things go wrong (which inevitably happens on occasion in a non-deterministic world like medicine and life!), AI will be used retrospectively by patients, lawyers and insurers to argue the clinician should’ve done things differently. → Clinicians will face pressure to concede to as many AI model outputs as possible, regardless of their clinical judgment: “hey I followed the AI model like you said and it told me to do X against my clinical judgment, and the patient died - so I can’t be liable!” → To further self-protect themselves from scrutiny, clinicians will face pressure to over-test and over-treat findings that AI will increasingly detect Ultimately they argue this could lead to a Prisoner’s Dilemma:  each individual stakeholder in a rationally self-interested manner would use AI to protect themselves instead of cooperating for the best patient outcome. My thoughts about this: → The AI train has left the station - we can’t stop patients from using AI to self-diagnose illnesses, self-read medical imaging or get a second opinion. But we have a moral imperative to educate patients on what AI can and can’t do well when it comes to patient care. Yes this takes time, work and patience. → Regulators and lawmakers need to ensure that clinicians still lead on evaluating good/bad clinical judgment. That decisions on whether a clinician deviated from a reasonable standard of care are still left up to clinicians and not purely an AI algorithm. Of course increasingly over time the clinician body will find use cases where applying AI will become evidence-based practice - but the point is we need to let our clinician societies remain independent on advancing what evidence-based standard of care is in a world of AI. If they lose that trust and autonomy, patients will suffer as clinicians will become sycophants to AI - instead of doing what’s best for patients. → There does exist some very small number of subpar clinicians who need to level up their care, and AI applied appropriately could be a forcing function to shine a light on this. But given the complexity of medicine and mistakes AI still often makes, we still need to have clinician-led, evidence-based approaches to using AI to weed out clearly and repeatedly poor clinical decision making to flag clinicians for review - instead of second guessing every single decision an otherwise good clinician makes. If we let AI become the judge instead of a tool that empowers clinicians to improve patient outcomes, we’ll have a system where everyone protects themselves from blame - and no one protects the patient.

22

8 Comments

Before you commit to building any component of a telehealth platform, talk to me. I’ve been down that road. A custom build can be a great investment, but only under the right conditions. Too many variables, too little time, and you end up owning tech instead of traction. I’ve rarely seen regret from going to market with a partner first. You gain speed, save cost, and validate your vision. You can always own the stack later once you’ve earned the scale to justify it.

16

Agentic AI Isn’t Coming — It’s Already a Pressure Point in Healthcare M&A Healthcare consolidation is accelerating in 2026 — providers, payers, and marketplace platforms are combining to gain scale, margin control, and operational leverage. Take for instance Spring Health announcing their acquisition of Alma. What’s changing quietly in these deals is how AI is being evaluated. Agentic AI is no longer a roadmap item or a demo feature. It’s becoming a diligence issue. Here’s why. 🔹 Consolidation turns AI from a feature into an operating risk. In a standalone company, an AI agent can live inside a narrow workflow. But post-merger, that same agent spans: - multiple care models - different policies - incompatible data standards - conflicting incentives At scale, agentic AI doesn’t just automate work; it arbitrates between systems. That’s power. 🔹 Buyers aren’t asking “does it work?” anymore. They’re asking: - Who controls the agent when workflows collide? - How are overrides handled across merged organizations? - What happens when an agent optimizes locally but harms the system globally? - How do we shut it down without breaking operations? If those answers aren’t clear, the AI isn’t an asset; it’s integration debt. 🔹 Agentic AI exposes governance gaps faster than any other technology. In traditional systems, unclear decision rights slow down execution. In agentic systems, they decide outcomes automatically. Once agents act across scheduling, billing, care navigation, or utilization workflows, ambiguity stops being theoretical. It becomes encoded behavior. 🔹 This is why consolidation favors platforms with restraint, not just intelligence. The winning healthcare marketplaces won’t be the ones with the most autonomous agents. They’ll be the ones that can: - scope authority precisely - enforce policy consistently - surface decisions transparently - and absorb accountability when automation moves faster than certainty That’s not an AI capability problem. It’s an operating model problem. In the next wave of healthcare consolidation, agentic AI won’t differentiate products. It will differentiate platforms that can be trusted to operate at scale. And that trust won’t come from model performance. It will come from governance, ownership, and control — designed before automation hardens them. #HealthTech #AIinHealthcare #ProductLeadership #ChiefProductOfficer #HealthcareInnovation #OperatingModel #BoardroomReality #DigitalHealth

7

1 Comment