This article is drawn from the Season 2 premiere of The Signal Room, my conversation with David Niewolny of NVIDIA. Listen to the full episode: AI in Hospitals: Building the AI-Native Hospital.
Most health systems I have worked with would describe themselves as digital. The records are electronic, the devices stream data, and the dashboards refresh on schedule. When I asked David Niewolny, director of business development for healthcare and medical at NVIDIA, what an AI-native hospital is, his answer showed how far that description sits from the hospital of the future. In his framing, a hospital today holds data much the way a database does. People still have to read it, interpret it and act on it, one image and one schedule at a time. An AI-native hospital puts the intelligence it already creates from patient and operational data to work for the nurses, doctors, surgeons and administrators who carry the load.
Digitized and Intelligent Are Two Different States
David described his career as a twenty year journey that started with wireless technology and the early work of pulling data off medical devices, on the premise that intelligence would come out of it someday. The digitizing is largely done. What he pointed out is that it added a whole set of workloads with nothing to do with caring for a patient. Clinical documentation is the obvious one. Scheduling is another. Administrators still plan the next three days with a spreadsheet and partial information, and a surprising amount of the hospital day still runs on whiteboards and clipboards. I have watched this pattern for most of my career. Each new system promised relief and handed the front line another screen to feed. The hospital of the future has to reverse that direction, so the data does work for clinicians instead of creating work for them.
Capacity Is the Promise Worth Measuring
I asked David whether he could name a significant role that AI had actually eliminated. He could not think of one. Radiology was his example. Years ago the prediction was that radiologists would become extinct. Today there are more of them than ever, because the efficiency AI created raised demand for imaging, one of the most effective diagnostic tools medicine has, and that demand called for more radiologists. The goal he sets for an AI-native hospital is capacity, meaning a facility that can see double, triple or even quadruple the patients it sees today. That matters when staffing is short, margins are thin and patients in many systems wait based on need and priority. Framed that way, the leadership question becomes how many more patients the same team can care for well.
One X-Ray Room Shows the Whole Idea
The example from this conversation I keep coming back to is the X-ray room. Today a technologist sets up what David described as essentially a giant robot. In the version he sees coming, a patient walks in and talks with a digital agent that helps with positioning. The device reads that position, sets itself up, takes the image and checks its quality before the patient ever leaves the room. That last step is bigger than it sounds. David noted that poor image quality is one of the main reasons patients have to come back. Add the reading that can now take a couple of days, and the entire workflow can fit inside a single visit. Nobody in that room was replaced. The patient got an answer sooner, the technologist and radiologist spent their time where judgment matters, and a return trip never happened.
Driver Assistance for the Operating Room
David borrowed a frame from autonomous vehicles, an area where NVIDIA does a great deal of work. Those systems run from level one, very basic autonomy, to level five, full self-driving. He does not expect surgery to reach level five, because care is too personal. What he does expect is the surgical equivalent of lane change warnings and automatic braking inside a procedure. By his estimate roughly twenty percent of surgeries are now done robotically, and he cited a study by Andrew Hung, presented at the Surgical Robotics Conference, that linked how a surgeon moves during a procedure to patient outcomes. That link is what turns surgical skill into a trainable data set.
From there the operating room connects to the rest of the building. Cameras and microphones capture what is happening, and if a tool is missing, a mobile robot can retrieve it using visual context alone. David called this an orchestration layer, where agents and software reason together much as the professionals in the room would. He pictures the entire hospital almost acting as a robot, with every AI tool working from the same information.
Explainability Is a Chain of Decisions
I pressed David on explainability, because brilliant engineering does not earn trust by itself. His answer started with open models. NVIDIA backs open models and open model architectures so the questions people ask about a black box have answers: how the model was trained, on what data, and with what weights. He then described explainability as a series of choices a team should be able to defend. Be clear about the problem you are solving. Explain why you chose the data you chose and how it was labeled, ideally by leaders in that specialty. Show how the model was trained and how you augmented the data set.
That last piece is where healthcare parts ways with the frontier labs. Large language models train on enormous pools of data, while healthcare pools are much smaller. A surgical corner case may appear once in ten years, yet clinicians still need to know how to handle it. World foundation models can generate synthetic versions of those rare cases to add to training, and simulated environments let teams test the finished application before it reaches clinical trials or a patient. David was direct that the burden of proof sits with the companies building the technology, and he gave the agencies that review it credit for insisting it be safe and effective. I agree with him. Speed in this market will belong to whoever can show their work.
Build the Hospital Twice
Digital twins were the part of this conversation I had been waiting for. David pointed to the work NVIDIA did with BMW to model an entire manufacturing facility down to the last component, running simulations on staffing and on how robots would work together before anyone broke ground. Health systems planning new hospitals are starting to do the same. Because a hospital is built once and stands for decades, the design questions now include ones nobody asked twenty years ago. Should robots have their own hallways for moving supplies between storage areas and patient rooms? How should the back of house work when humanoids and autonomous mobile robots take on tasks that are not patient facing? Anyone who has walked through hospital supply closets knows they look nothing like an Amazon warehouse. David sees a huge opportunity in treating the facility like a manufacturing plant in simulation, and I share his view that the payoff starts long before the first patient arrives.
Where Hospital Leaders Should Start
When I asked what an executive should do now, with the worry that whatever they buy today will be leapfrogged tomorrow, David gave two pieces of advice. Pilots will happen, so aim the first serious one at the single biggest problem the organization faces, whether that is patient care, cost or something else. Then remember that hospitals buy technology far more than they build it, which makes the choice of partner as important as the use case.
I added a point from my own experience. Healthcare has a habit of biting off more than it can chew. We overestimate our ability to take on complex technology problems and underestimate the time it takes to become proficient. Scale is the other blind spot. Most hospitals operate in a confined geography, and even the largest systems rarely deal with overlapping markets, so few of us have had to solve the problems technology companies solve every day. Our first obligation is still caring for the people who come through the door and keeping the lights on. A partner that has already worked through scale lets a health system stay focused on that obligation.
Humans Stay in the Loop
My last question was what would shock David ten years from now. His answer was a world where humans are removed from the loop. Research on fully autonomous robotic surgery exists, and some public figures predict that all surgery will eventually be done by robots. David does not see it that way. In every use case his team supports today, humans remain in the loop and the technology makes them more efficient. That is the version of the hospital of the future worth building, one that can see, listen and coordinate, with clinicians still in control and better informed than they have ever been.
I came away from this conversation convinced that the hard work ahead is organizational. The technology is arriving on schedule. Whether our hospitals are ready to put it to work for the people delivering care is the question every leader should be asking now.
Related reading from Hutchins Data Strategy
Hutchins Data Strategy Consultants helps health systems put this into practice. See Clinical & Operational Integration.
Authoritative sources
- NIST AI Risk Management Framework
- FDA: Artificial Intelligence-Enabled Medical Devices
- World Health Organization: Ethics and governance of artificial intelligence for health
Frequently asked questions
What is an AI-native hospital?
An AI-native hospital puts the intelligence it already creates from patient and operational data to work for the people delivering care. Most hospitals today are digitized, yet their data still behaves like a database that people have to read and act on by hand. An AI-native hospital adopts agentic and physical AI, from digital agents to robotic medical devices, so that data actively supports nurses, doctors, surgeons and administrators and frees their time for patients.
Will AI replace clinicians in the hospital of the future?
David Niewolny of NVIDIA could not name a single significant role that AI has eliminated. Radiology is his example. Early predictions said radiologists would disappear, yet there are more today than ever because efficiency raised demand for imaging. The goal he describes is capacity, helping the same teams see two, three or four times as many patients while humans stay in the loop.
How can healthcare AI be made explainable?
Explainability rests on a chain of decisions a team should be able to defend: the problem being solved, why a data set was chosen, how it was labeled, how the model was trained and how the data was augmented. Open models and open model architectures help answer how a model works. Synthetic data from world foundation models can cover rare corner cases, and testing in simulated environments shows the likely impact before clinical trials or any contact with a patient.
How are digital twins used in hospital design?
Health systems planning new facilities can build the entire hospital in simulation first, much as BMW modeled a full manufacturing plant with NVIDIA before breaking ground. Teams can test staffing and workflow and plan for robots that move supplies, including whether those robots need their own hallways. Since a hospital stands for decades, getting that design right before construction has an outsized payoff.
