What will it take for artificial intelligence to move from impressive demonstrations to meaningful impact in real-world healthcare?
Dr. David Rhew, Global Chief Medical Officer and Vice President of Healthcare at Microsoft, explored this question through a practical lens. Drawing on his background as both a physician and computer scientist, Rhew described AI not as a distant promise, but as a rapidly advancing set of tools that must now be carefully integrated into clinical workflows, population health strategies, and human decision-making.
His message was clear:
AI is already real. The challenge now is to apply it responsibly, redesign workflows around it, and ensure it expands human capability rather than simply replacing human roles.
Translating Between Technology and Care
Rhew began by describing his role at Microsoft as that of a translator: someone who connects what is possible on the technology side with what is needed on the clinical side. The goal is not simply to showcase innovation, but to understand how it fits into workflows, how it can be adapted, and where it creates real value.
He reflected on studying computer science as an undergraduate in 1988, when his thesis focused on AI in healthcare. At the time, few people were interested. Decades later, AI has exceeded many of the benchmarks set for it, from reading and writing to clinical examinations and diagnostic tasks.
But Rhew cautioned that benchmarks are not enough. The real test is whether AI can improve care in practical, real-world settings.
Moving Upstream in Population Health
One of the first opportunities Rhew highlighted was using AI to help healthcare systems become more proactive.
Too often, patients arrive at the hospital or clinic only after disease has already progressed. A heart attack, stroke, or heart failure admission may represent the visible endpoint of a process that began months or years earlier. By that point, intervention is more difficult, more costly, and often less effective.
AI can help healthcare systems move upstream by identifying people who are at risk, people who may already have disease but do not know it, and people whose known disease is silently progressing. This matters because a small percentage of the population accounts for a large share of healthcare costs, and earlier detection can help prevent complications, reduce avoidable utilization, and allocate limited resources more effectively.
Finding Risk in Everyday Data
Rhew shared examples of how familiar clinical tools can become more powerful when paired with AI.
The electrocardiogram, long used to detect arrhythmias, can now also support detection of heart failure, valvular disease, hospitalization risk, readmission risk, and even longer-term mortality risk. The value is not simply prediction for its own sake. When clinicians know a patient is at high risk, they may act more aggressively, monitor more closely, and prioritize limited resources toward those most likely to benefit.
He also described retinal imaging as a powerful window into broader health. A simple, non-invasive eye exam can help detect diabetic retinopathy, hypertensive retinopathy, cardiovascular risk, chronic kidney disease, and even early biomarkers associated with neurological conditions such as Alzheimer’s, Parkinson’s, and multiple sclerosis.
Importantly, these tools are moving beyond specialty clinics. Rhew described examples of retinal screening being deployed in primary care, health fairs, and rural clinics, where it can identify high-risk patients who might otherwise go undetected until much later.
The Data Side of the AI Coin
Rhew emphasized that an AI strategy cannot focus only on models. AI is like a coin: one side is the model, and the other side is the data.
For healthcare AI to be truly useful, data must become more multimodal. Images, EKGs, medical records, lifestyle factors, genetics, social determinants of health, and environmental data all provide important context. A model trained on one type of data may be helpful, but a model that can combine multiple signals can better risk-stratify individuals, identify patient cohorts, and predict likely outcomes.
This opens the door to more personalized and proactive care. By comparing a patient with similar patients across longitudinal datasets, AI may help clinicians understand not only what diagnosis is likely, but what disease course, treatment response, or risk trajectory may be expected.
Rhew described work using large pathology datasets to identify disease and predict genetic mutations, cancer subtypes, and outcomes that traditionally require additional testing. For him, this is a glimpse of where healthcare AI is heading: not just recognition, but prediction and guidance.
From Generative AI to Agentic AI
Rhew then turned to one of the most important emerging developments: agentic AI.
Traditional AI and machine learning have largely focused on prediction. Generative AI has introduced conversational interfaces and new ways to synthesize information. Agentic AI goes further. An agent can be understood as a system built around a large language model, with instructions, protocols, access to tools, memory, and the ability to take actions.
Rhew compared agents to “minions”: each can perform a specific task, but the real power comes when multiple agents work together under human supervision.
He illustrated this through diagnostic reasoning. In a Microsoft model tested on complex New England Journal of Medicine case studies, multiple agents each performed a distinct role: one generated a differential diagnosis, another identified possible tests, another reviewed evidence and guidelines, another considered cost, and another challenged whether something had been missed. Working together, the agents performed significantly better than a single model alone.
The lesson was not that AI should make clinical decisions independently. It was that structured collaboration between specialized agents can support more rigorous, evidence-informed, and cost-conscious decision-making.
Humans Must Stay in the Loop
Rhew also used the example of tumor boards to show why human judgment remains essential.
Tumor boards bring clinicians, patients, and experts together to weigh complex treatment options. They are highly valuable, but resource-intensive, and only a small fraction of cancer patients have access to them. AI agents could help by extracting information from the medical record, summarizing visits, creating timelines, identifying prior treatment failures, and comparing treatment pathways based on survival, quality of life, and cost.
But those outputs do not determine the “right” answer on their own.
When decisions involve tradeoffs between life expectancy, quality of life, cost, and patient preferences, humans must remain central. AI can organize the evidence, surface options, and accelerate insight. People still have to weigh values, context, and what matters most to the patient.
AI Plus Clinicians, Not AI Instead of Clinicians
Rhew closed with a cautionary message about over-reliance on AI and the danger of simply replacing human roles.
He shared the story of a patient who came to the emergency department with severe abdominal pain and died from a ruptured aneurysm. The CT scan that showed the aneurysm before it burst was waiting in the queue to be read by a radiologist. AI could likely have detected the finding earlier. But the lesson, Rhew emphasized, is not to replace the radiologist with AI. The lesson is to combine AI with the radiologist in a redesigned workflow that allows urgent findings to be flagged and acted on sooner.
That is where AI can create real impact: not as a substitute for humans, but as a way to expand what humans and healthcare systems can do.
The Takeaway
AI in healthcare is no longer just hype. It is already being used to detect risk, interpret data, support diagnosis, guide population health, and assist clinical decision-making. But its success will depend less on the technology alone and more on how healthcare leaders implement it.
The future will require stronger data strategies, multimodal models, workflow redesign, governance, human oversight, and a clear focus on outcomes that matter.
Rhew’s core message was ultimately practical: AI becomes meaningful only when it is translated into better care. The opportunity is not simply to automate what healthcare already does, but to redesign systems so clinicians can act earlier, patients can be reached sooner, and organizations can achieve more than they could before.
