Medicine Is Headed Toward Semi-Autonomous Care. Are Doctors Ready? With Counsel Health CMO Rishi Khakhkhar

Key Podcast Moments

• Rishis walk through exactly what an AI is and isn't allowed to decide before a physician has to step in
• The two dig into a JAMA perspective arguing AI already outperforms doctors, and where Rishi thinks the authors get it right and wrong
• Graham pushes on whether residents trained alongside AI will lose the instinct to know sick from not sick in five seconds
• Rishi describes holding medical licenses in all 50 states, and what that actually costs a company trying to recruit good physicians
• The conversation ends on what primary care looks like in 2036, and why Rishi doesn't think it will be built around one person anymore

An emergency physician sees the same patient more than once. Not the same chart, but the same shape of a visit: someone whose blood sugar has been climbing for weeks, who finally shows up in a hallway bed at 2 a.m. because nothing else was open, and who didn't need a trauma bay so much as an earlier conversation.

Dr. Rishi Khakhkhar spent the early years of his career loving the trauma bay part of emergency medicine, the chest tubes and intubations. Now his favorite patient is the asymptomatic hyperglycemic, the one everyone else has learned to discharge quickly once diabetic ketoacidosis is ruled out. Rishi likes to sit with them instead. That instinct, to intervene earlier and more cheaply than the ER allows, is what led him to Counsel Health, an AI-enabled virtual care company where he's chief medical officer, while he continues working ER shifts at Mount Sinai.

The split matters to how he talks about AI. Rishi isn't pitching a product on this episode so much as describing how AI-assisted, physician-supervised care actually functions once you build it: an AI takes a full patient history first, the way a well-trained medical assistant would, then flags when a physician should join the conversation. The decision to escalate belongs to the patient, not the algorithm. Anything touching diagnosis, prescriptions, referrals, or diagnostics stays with a physician.

That structure sets up the episode's sharpest exchange. Graham raises a recent JAMA perspective in which Zeke Emanuel and coauthors argued that AI already performs as well as, or better than, physicians in many clinical scenarios. Rishi doesn't dismiss the claim outright. He agrees that on narrow, well-defined tasks, like reading subtle EKG changes that even experienced physicians miss, AI can already outperform a human reader, and he sees no reason to hold that back. Where he pushes back is on the broader implication that human-in-the-loop care is simply inefficient. Rishi compares it to autonomous vehicles: Waymo spent years running supervised, human-backed driving before removing the driver at all. He argues medicine is at the equivalent stage now, one he calls semi-autonomous care, and one he isn't in a hurry to skip past.

The conversation then turns to what that means for the people coming up through training. Graham raises a worry that won't be unfamiliar to anyone who's watched a resident lean on a tool: if AI is doing the pattern recognition, will residents build the instinct to know, in five seconds, whether a patient in front of them is sick or not? Rishi doesn't fully resolve the tension, but he's clear that some things can't be outsourced yet: reading a patient's affect, the feel of an ultrasound-guided line, the kind of gestalt that takes repetition to build. What can be delegated, he argues, should be, the same way emergency departments already let RNs run chest pain protocols without a physician signing off on every step.

Structural barriers come up too, and Rishi is direct about the ones that frustrate him most. He holds medical licenses in all 50 states, a requirement he describes as a genuine obstacle to recruiting good physicians into AI-enabled care, since asking a new hire to wait months for licenses to clear before they can start seeing patients is a hard sell. It's a small detail that stands in for a larger point: much of what's slowing AI-enabled care down isn't the technology, it's the infrastructure built around a pre-internet version of medicine.

The episode closes on where this is all heading. Rishi doesn't think primary care in 2036 looks like a relationship with one person. He imagines something closer to an AI that knows a patient's full history, preferences, and health literacy, supervised by physicians managing much larger panels than they do today, with that kind of AI-enabled panel management eventually carrying the same prestige, in his telling, as a surgical subspecialty. Whether that vision arrives on his timeline is an open question. That medicine is already moving toward some version of semi-autonomous care is, by the end of the conversation, harder to argue with.

Thank you to our wonderful sponsors for supporting the podcast:

Evidently - Leading AI-powered clinical data intelligence https://evidently.com/

Top 4 Takeaways

AI beating doctors on a narrow task is not the same argument as AI replacing doctors, and this episode is where that distinction actually gets tested. Here's what held up.