Healthcare AI Isn't One Thing

When people say "AI in healthcare," they mean ten different things at once. Ambient scribing. Prior auth automation. Surgical robotics. Drug discovery. Radiology triage. Predictive risk scoring. All under the same umbrella.

These are not all the same kind of work. They demand different teams, different timelines, different unit economics, and different definitions of what success looks like. The investor, the operator, the clinician on the receiving end all deserve a more useful map than "AI is coming to healthcare." Lumping the layers together is how you end up disappointed in all of them at once.

So here's how I'd draw the map. At least the part of it I have any business commenting on.

The first layer is burden-easing AI. Ambient scribing. Prior authorization. Revenue cycle. Patient communications. The work focused on the operational and administrative overhead that's been piling up on clinicians and practices for the last fifteen years. This work is real. It frees time. It restores some part of what made medicine feel like medicine to the people who chose to practice it. A lot of really smart entrepreneurs are working in this layer because they see a problem they know how to fix, and they're right. There are still plenty of problems left at this layer worth solving (anyone who's spent five minutes inside a prior auth workflow knows this).

The second layer is established clinical AI. Radiology models that flag a single finding on a scan. Risk calculators built into the EHR. ML-driven sepsis alerts that have been quietly running in ICUs for years. This layer already has many active solutions doing important work in clinics and hospitals.

The third layer is predictive modeling and precision medicine. This is the layer Voythos is working in, along with several other companies. The goal is to predict how an individual patient's disease is going to behave over time. Not what the average looks like for someone with that diagnosis. The specific person, with their specific anatomy, their specific history, their specific likelihood of a complication at month six. The aim is to give a clinician a real view of the future, not just a snapshot of the present.

This third layer is hard. Really hard. Prediction is fundamentally harder than detection or classification. Patient outcomes are variable in ways that imaging findings are not. The models often have to fuse multiple data types together (imaging, EHR data, structured measurements, sometimes genomics). The data has to be longitudinal, because you can't predict a trajectory from a single point in time. The volume of data needed is larger.

That's why this layer is less crowded. Not because it's not worth doing. Because it's hard, and the timelines are long.

Let me be clear about something. We didn't start Voythos because we wanted to "solve this third layer." That's not how it happened. We started because we saw a problem, and the third layer was where it had to be solved. We followed the problem here. We didn't pick the layer first.

The reason this third layer matters, and matters now, is that it's where the biggest delta lives between what medicine can do today and what we believe it can do in the future. The first two layers ease the burden of practicing medicine and improve the precision of specific clinical tasks. The third layer changes what medicine is able to do. All three are worth working on. But if the future we want is one where a vascular surgeon can sit down with a patient and know, with real certainty, what's going to happen if she intervenes now versus in two years, that work has to happen at the third layer.

That future is going to need more capital, more clinical conviction, and more patient teams working in this layer than are working in it today. We need more companies, not fewer. And we need the field, the investors, the clinicians, and the operators evaluating those companies to actually understand what they're looking at.

Healthcare AI isn't one thing. It's at least three (probably more). The work happening at the third layer, with all its difficulty and all its mess, is where some of the biggest gains in the next decade of medicine are going to come from.