The 9-Person Insurance Company and the Real Line in AI-First Healthcare

Y Combinator has a name for it: burn tokens, not headcount. A health insurance company called Decent runs with nine people total. Twofold does revenue cycle management with three. Deep Cura Health handles patient scheduling, prior authorization, and insurance verification with two humans and seven AI agents.

The AI-first model is real. It's working. And in healthcare, every one of these companies has made the same quiet choice: they're attacking admin, not clinical.

This episode unpacks why — and why the conventional wisdom ("admin is safe to automate, clinical isn't") is the wrong frame. The real line isn't admin versus clinical. It's decision support versus decision making. IBM burned $4 billion learning the difference with Watson Health. The next generation of healthcare AI companies will either learn from that, or rebuild the same trap with better UX.

What's covered:

• How AI-first companies are quietly rewriting healthcare staffing
• What IBM Watson Health actually got wrong — and why "the AI was wrong" misses the lesson
• Why most "decision support" products today are decision-making in a trench coat
• The payment-model problem nobody is pricing: a 5,000-patient panel breaks fee-for-service
• The companies positioning themselves on the right side of the line

🎥 Watch on YouTube: https://youtu.be/9fHKQqm15qo
📝 Companion essay (with the receipts I had to cut for length): https://danmccoymd.substack.com/p/the-part-of-ai-first-healthcare-that

Transcript

SPEAKER_00 0:00

There's a company called Deep Cura Health that runs on two humans and seven AI agents. Two people, seven agents. They handle patient scheduling, insurance verification, prior authorizations, the administrative chaos that buries every medical practice in America. And they're not alone. Twofold does revenue cycle management with three people. Three claims, denials, payment posting, work that used to require teams of 15 or 20 people. Decent, a health insurance company, operates with nine people total. We're talking underwriting, claims adjudication, member services, all automated. This is what Silicon Valley calls the AI first model. YC or Y Combinator just released a video about it, the idea that you can build a $100 million company with a skeleton crew by replacing middle management with AI agents. So burn tokens, not headcount. That's the philosophy. And in tech, it's working. Small teams are doing what used to take hundreds of people to do. But healthcare isn't tech. IBM learned this the hard way. They spent $4 billion, billion with the B, acquiring companies to build Watson Health. They partnered with MD Anderson, Memorial Sloan Kettering, the best cancer centers in the world. The pitch was simple. Watson would analyze patient data and recommend cancer treatments, AI-powered oncology, the future of medicine. And it failed catastrophically. Here's what actually happened. Watson was trained primarily on synthetic cases and treatment preferences of a small group of doctors at Memorial Sloan Kettering. It learned their biases, their preferences, their specific approach to oncology. Then IBM sold it globally as if it were a universal medical truth. Now, internal documents later revealed that Watson made unsafe and incorrect treatment recommendations. In fact, in one case, it recommended a cancer treatment for a patient with severe bleeding, a recommendation that could have been fatal if it was actually followed. The trust gap in healthcare isn't about whether the technology works. It's about the gap between this performs well in a demo and I'm willing to stake my medical license on it. When you burn tokens in tech and the AI gets it wrong, you get a bad product recommendation. Maybe you lose a customer. But when you burn tokens in clinical care and the AI gets it wrong, you can burn bodies. You can burn tokens, but you can't create bodies. So what's actually happening with these AI-first healthcare companies, the twofolds and deep curas and decents? Well, right now, they're all doing the same thing. They're attacking administrative healthcare, not clinical healthcare. Admin AI is working. Prior authorization automation saves 50 to 20 minutes per request. Revenue cycle AI cuts denial rates. Scheduling agents reduce no-shows by up to 30%. These are tasks where errors mean paperwork delays and not patient harm. Clinical AI telling a doctor what to prescribe, what treatment to recommend, that's where Watson died. The liability chain is very personal. When something goes wrong, the doctor signed the order. The AI doesn't have a medical license to revoke. But here's where I think the conventional wisdom actually gets it wrong. The line isn't admin versus clinical. The line is decision support versus decision making. Think about what actually limits a primary care physician today. A typical doctor manages a panel of maybe 2,000 patients, but they're not providing continuous care to 2,000 people. They're reacting, seeing whoever shows up, reviewing charts one at a time, hoping they catch the patient whose blood pressure has been creeping up for six months. Now imagine AI that surfaces the 47 patients in your panel who haven't refilled their diabetic medication, the 12 lab values that need review, the eight who are overdue for cancer screenings. The AI isn't deciding what to do. The physician still makes every clinical decision, but instead of hunting through charts, the doctor's attention goes exactly where it's needed. That's not administrative, that's clinical, but it's decision support, not decision making. And it changes the math entirely. If AI handles the cognitive load of tracking 2,000 patients, surfacing who needs attention, synthesizing their records before visits, closing care gaps through automated outreach, well, suddenly one physician can meaningfully manage 5,000 patients, maybe even more. Population health at a scale that's never been possible before. Medication adherence programs that actually work because the follow-up is persistent and it's very personalized. The shift from reactive medicine waiting for patients to get sick enough to show up to proactive monitoring that catches problems early. This is where the AI first model in healthcare actually goes. Not three people replacing 30 doctors. That's the wrong frame. It's one doctor doing what used to require five, with better outcomes. Because the AI isn't making decisions, it's eliminating the cognitive overhead that limits how many patients one human can meaningfully care for. The admin side of healthcare is a trillion dollars of waste. That's the immediate opportunity, and I agree. But the bigger play is clinical decision support, expanding a physician's reach without replacing their judgment. The companies that figure out where that line is, decision support, not decision making, human oversight at every step, they're building something much larger than a building automation tool. If you're building in this space or investing in it, that's the question. Not can AI do medicine, it's can AI make one doctor effective as five? The technology isn't there yet for everything, but it's getting there. And the companies positioning themselves on the right side of that line, they're the ones to watch. And thanks for watching here. I'll see you in the next one.