Two numbers tell you where healthcare AI actually stands.
$995 million went into healthcare AI in 2025. Roughly 70% of it never reached scale.
That is not a technology failure. The models are good. The demos are impressive. The pilots start fast.
Then they stall, and 18 months later nobody can explain what changed for the patient or the physician.
I have spent the last three years asking why. The answer says a lot about where generative AI in healthcare goes next.
Where we are now: smarter suggestions, same broken system
Generative AI entered healthcare through documentation. Ambient scribes that listen to a visit and draft the note. It was the right entry point. Documentation is the pain every physician feels every day, and the technology works.
But documentation is maybe 20% of the problem.
The average hospital runs on hundreds of software tools. For every physician seeing patients, there are around ten administrative staff behind them. Four out of five hospitals still cannot share a patient record across departments without printing a PDF. In 2026.
Into that environment, the industry shipped point solutions. A phone bot from one vendor. A coding tool from another. A scribe from a third.
Each one built in isolation. None of them talk to each other.
So the physician becomes the connector. They use one tool, set it down, pick up the next, switch context, and manage handoffs nobody designed. A tool saves minutes on one task. The workflow stays broken.
That is the “now” of healthcare AI. Smarter suggestions bolted onto the same broken system.
The question that separates the present from the future
Here is the fastest way to evaluate any healthcare AI: ask what it does once it is done thinking.
A copilot hands you a suggestion and waits. The next generation of systems opens the tab and finishes the task.
Calls the patient. Books the appointment. Writes the note.
Assigns the codes. Sends it back to the EHR. Follows up three days later.
One gets measured in licenses and daily active users. The other gets measured in minutes worked and tasks closed out.
A “daily active user” in healthcare is a physician clicking buttons at 9pm. Calling that adoption is how the metric hides the problem. The market was never short on smarter suggestions. It is short on the work actually getting done.
Generative AI’s real contribution to healthcare will not be better text. It will be labor capacity. The ability to execute administrative work, not just describe it.
The future is a system, not a collection of tools
The bet we made at Sully.ai from day one was simple: healthcare does not need another tool. It needs a single operating system for AI agents, one that can run hospital operations autonomously.
Why an operating system? Because healthcare is a journey, not a series of disconnected tasks.
It starts the moment a patient calls the hospital, or the hospital calls the patient for a checkup. From that first call you learn things: call volume, patient sentiment, why people are reaching out. An agent schedules the appointment, confirms it, chases the no-shows. Before the visit, another agent collects symptoms, history, and medications, so the physician opens a complete chart before entering the room.
During the visit, the conversation becomes a clinical note. A checklist verifies itself. The full patient summary sits at the physician’s fingertips instead of buried across EHR screens.
After the visit, the note flows back to the EHR with the codes attached. Follow-up agents call the patient, confirm information, and update statuses in the record.
The critical word in all of that is context. Every agent along the journey works from the exact same picture of the patient. That is what point solutions can never do, because each one only sees its own slice.
Gartner recently called multi-agent AI “an operating model shift.” That is analyst language for: if you are still buying disconnected tools, you are already behind.
Autonomy will arrive in levels, and the order matters
Nobody should promise a fully autonomous hospital tomorrow. The honest way to think about this is levels, the same way the auto industry thinks about self-driving.
The first level is hospital operations: the calls, the scheduling, the intake, the documentation, the coding, the follow-ups. This is where autonomy belongs right now, because the stakes are administrative and the evidence has been in for years.
From there, the stack goes up. The machines and devices inside the hospital. Then patient-generated data from home: wearables, home labs, remote monitoring. Each level feeds the next, until the system sees the whole patient, not just the ten minutes they spend in an exam room.
This sequencing matters more than most people realize. The “AI in medicine” conversation collapses two different problems into one.
Should AI touch clinical decisions? Maybe, with real evidence and heavy oversight.
Should AI handle scheduling, intake, documentation, and follow-up calls? Yes. Right now.
These are not a choice. They are an order of operations.
Every front-desk call, every intake form, every visit note is patient context. Today most of it dies the moment a human hangs up the phone. Automate the administrative layer first and you capture proven ROI, and you quietly build something else: a continuously updated record of who the patient actually is.
A clinical AI sitting on top of a broken admin layer is being asked to reason about a patient it does not know. A clinical AI sitting on top of months of clean, structured context is doing a different job entirely. Same model. Very different output.
What this means by 2030
Play the levels forward and healthcare stops being reactive.
The physician shortage is projected to reach 86,000 in the US by 2036. Replacing one burned-out physician costs roughly $500,000. We cannot hire our way out of this. The people do not exist, and the training pipelines do not produce fast enough.
Generative AI is the first technology that adds labor capacity to healthcare instead of adding software to it. When the administrative layer runs itself, the 7-minute appointment becomes 20. The physician looks at the patient instead of the screen. And the system starts catching things early, because for the first time it holds continuous context on every patient, not fragments scattered across departments.
The goal was never to replace physicians. AI is not coming for doctors. It is coming for the mountain of work between doctors and their patients.
There are 4.5 billion people in the world who cannot reliably see a doctor. The ones who can wait weeks and get seven minutes.
The doctors were never the bottleneck. The system around them is. Generative AI is how we finally rebuild it.
One Human, One Doctor. That is the future worth building toward.
