By Renana Ashkenazi
By Renana Ashkenazi
Earlier this year, a panel of five doctors worked through some of the hardest cases ever published in the New England Journal of Medicine and got 85% of them right. A panel of 21 practicing physicians, given the same cases, got about 20%.
The five doctors were not human. They were AI agents in Microsoft’s diagnostic orchestrator – a hypothesis generator, a test selector, an evidence interpreter, a consensus builder and a final diagnostician – arguing through each case in what Microsoft calls a “chain of debate.” Microsoft’s MAI-DxO isn’t in a clinic yet, and it obviously still needs real-world validation, not to mention regulatory approval. But we can already start fantasizing about the meaning of the result: a council of specialists that reason together better than any one of them could alone, already beat a room full of doctors four to one.
I think this concept of “digital council” is one of the most important shifts in healthcare AI right now. Here is why, and what I’d want to see if you’re building it.
We have cardiologists who know everything about the heart but can miss how a failing heart damages the kidneys. We have oncologists brilliant at targeting a tumor but disconnected from the patient’s metabolic baseline. Medicine has fractured into deep, narrow silos because human cognition is capped by time – we simply don’t have enough hours in a day. But the human body doesn’t care about our org charts; It is a hyper-connected graph where every system affects every other. When the heart fails, blood flow to the brain drops, the brain sends faulty signals to the gut, the gut shifts the microbiome, and the immune response changes with it. To actually solve for the patient you need someone who understands all of it – and understands it at the same time.
If you’re old enough to remember, that only someone was Dr. House. He was never great because of one specialty, but because he held all of them in his head at once, and half of every episode was his colleagues asking how he possibly knew. As great of a show it was, no real physician can log enough hours to be House. The math doesn’t work: a single human can only read so many papers, see so many cases, hold so many interactions in mind. We spent the last centuries optimizing for the part because understanding the whole was computationally impossible for a single human brain.
And the cost of that ceiling is massive. A Johns Hopkins study estimates that roughly 795,000 Americans die or are permanently disabled by diagnostic errors every year, and that nearly 40% of those are related to just five conditions, far from exotic diseases – stroke, sepsis, pneumonia, blood clots and lung cancer. Failures of synthesis, the cases where the signal was there but no one connected it in time.
For the last two years the industry has been obsessed with copilots. You’ve seen the demos: an AI assistant that sits beside the doctor, drafts the note, summarizes the chart. Great efficiency hacks, but really this is an enhancement of the old playbook, squeezing a little more productivity out of a process that is already broken.
The Digital Council is a different bet. Instead of asking one model to act as a doctor, you build a room: a cardiologist agent, a nephrologist agent, a pharmacologist agent, a geneticist agent. The system pulls the patient’s records, genomic data and wearable streams. The cardio agent reads the echo, the nephro agent reads the creatinine trend. Then they argue. The cardio agent wants to push the ACE inhibitor; the nephro agent counters that potassium is trending up, so switches to an ARB and recheck labs in three days.
Microsoft’s orchestrator is the existence proof that this approach can produce better answers (and that the council can do it while ordering fewer tests than the human physicians it beat). What comes out the other side is a synthetic super-generalist: the depth of a sub-specialist with the breadth of a general practitioner, an entity that has never really existed in one place before.
Sitting on the other side of the table, here is what separates a real company from a feature.
Build the general contractor, not the subcontractor. The defensible companies will not be the ones that detect a lung nodule slightly better than a radiologist. That is pretty much a commodity. They will be the ones that take the nodule, correlate it with the patient’s rheumatoid arthritis history, cross-reference their current immunotherapy and propose a plan. The value is in the connections, not the nodes.
Treat the patient as an N of 1. Current medicine optimizes for the average patient, which is a statistical artifact, not a person. The opportunity is a dynamic physiological model – a digital twin of the specific patient in front of you. One that you can simulate thousands of times before the first IV drip.
Start where human synthesis fails most. Sepsis is a multi-system cascade – exactly the kind of cross-domain failure the Council is built for, and one of the five conditions driving those 795,000 harms. Polypharmacy is another. Today about one in three Americans in their 60s and 70s take five or more prescriptions, and the risk of an adverse drug event climbs from 13% on two medications to more than 80% on seven or more. A de-prescribing agent that looks at a senior on fifteen drugs and works out which ones are fighting each other is actually a quantified, urgent product.
If you’ve pitched me before, I’ve probably asked the annoying question: why is this buildable today when it wasn’t a few years ago? The answer is not just better models.
The unlock is that the 2024–2025 wave of reasoning models – the o-series and what followed – shifted from predicting the next word to actually working through a chain of thought before answering. Agents that reason can negotiate, consider alternatives and find points of failure. Microsoft’s result is the ultimate Jewish proof – that the argument yields a better answer than any single model talking to itself.
The second shift is context. We can finally drop the raw imaging, the PDF pathology report and the messy handwritten notes into one window. Data normalization and harmonization which was a killer of so many health-tech startups in the last decade is not a problem anymore.
In the old world, the risk was that AI hallucinated a fact. In this new world, the risk is different: the committee agrees, and you have no idea how.
If five agents debate a diagnosis and converge on a treatment, who audited the argument? Why did the nephrology agent overrule the cardiology agent? If you are building this, observability has to be built in from the start. We will need a trace of the reasoning that a clinician can actually inspect and challenge.
Go back to the 795,000. Most of those patients weren’t failed by a doctor who knew too little, but by systems who didn’t connect what was already in the chart. Connecting it is exactly what a council of agents does well, but an answer a clinician can’t check simply won’t do. No one likes black boxes. The companies that win this will be the ones whose councils can be right, and show how they got there.
