AI Agents Could Save Lives—If We Let Them

The warnings about artificial intelligence keep getting louder.

The companies building the world’s most powerful AI systems are reporting troubling behavior from their own creations. And some of the researchers closest to frontier AI development warn that powerful new systems could “cause human extinction.”

Scientists fear losing control. Government leaders in the United States and abroad are debating how to keep increasingly capable models from causing harm.

These concerns must be taken seriously. But as a physician, I worry that fear of AI will cause our nation to overlook an extraordinary opportunity: harnessing narrower applications to improve patient care and save lives.

Consider a question I often pose to medical professionals. If you had time to check on patients every day, could you detect complications earlier and improve outcomes?

Almost always, the answer is yes.

For postoperative patients, that could mean identifying a surgical wound infection early enough to begin antibiotics without delay. For patients with chronic heart failure, it could mean treating at the earliest signs of acute decompensation. And for people with diabetes or hypertension, it could mean monitoring closely enough to maintain effective blood glucose and blood pressure control.

I saw the value of more frequent outreach and evaluation as CEO of The Permanente Medical Group. From 2000 to 2015, heart-disease and stroke mortality among Kaiser Permanente members declined roughly twice as fast as it did nationally. TPMG also controlled hypertension in 85% of patients, compared with 54% nationally.

However, that kind of care requires substantial clinician time and organizational resources. Most medical groups cannot afford to provide that level of attention consistently. There are not enough doctors, hours or dollars to do so nationwide.

Generative AI agents could change that calculus. These are software tools designed for specific tasks. They could monitor data, respond to health changes, and take predefined steps without direct clinician oversight.

Consider, for example, how AI might help one of the 6.7 million Americans living with chronic heart failure. A cardiologist monitoring a patient at home would look for increases in weight, ankle swelling, and fluid in the lungs. Combining those signs with reduced endurance and difficulty lying flat comfortably would indicate worsening heart failure.

These assessments are neither especially complex nor beyond what AI could do today. Bluetooth can connect agents to electronic scales and stethoscopes. Video monitoring would allow other agents to evaluate ankle swelling, measure stair-climbing ability, and assess whether the patient can lie flat comfortably.

In each case, these agents would augment what the doctor already does, not replace it. The potential benefit is significant. Across dozens of randomized trials, remote monitoring of patients with heart failure has reduced hospitalizations by roughly 20%. AI agents could make that kind of continuous monitoring easier and more affordable nationwide.

The same logic applies to chronic diseases such as hypertension and diabetes. Nearly 120 million American adults have high blood pressure, and 38 million live with diabetes. Both illnesses can damage the heart, kidneys, and blood vessels when poorly controlled.

Yet treatment still depends primarily on information collected during office visits, several months apart. Millions of patients already own home blood pressure cuffs or glucose monitors, but doctors rarely have the time to evaluate those readings. AI agents could track them, recognize when control is failing and notify physicians based on the same clinical parameters they would use if they could review the information frequently.

Consider the upside for patients with hypertension. The CDC estimates that better “team-based control” of high blood pressure could prevent roughly 92,000 heart attacks, 139,000 strokes and 115,000 cardiovascular deaths over five years. AI agents could support that care by making monitoring and evaluation easier.

A third opportunity comes after surgery. Once home, patients are instructed to watch for signs of infection at the incision site: redness, drainage, and pain. But because some of those changes are normal after surgery, patients may struggle to recognize when they signal a complication.

If surgeons could see patients daily, they would compare the wound’s appearance with how it looked the day before. AI agents could do the same by collecting photos or video from the patient’s phone, analyzing progression for signs of improvement or deterioration, and asking about pain or drainage.

If the application identified a likely problem, it could send the images and findings to a clinician, who could quickly determine whether the patient needed to be seen. Patients with concerns could still contact the surgeon directly.

Surgical-site infections add about $3.3 billion to U.S. health care costs each year. Earlier recognition would bring patients to medical attention sooner, when treatment — including antibiotics — would be more effective.

These opportunities to improve quality and save lives raise an important question. If Bluetooth, video, and conversational AI already exist, why haven’t they been assembled into tools that augment what physicians do today?

The biggest barrier is the complexity and expense of obtaining regulatory approval for medical tools used by patients. But Medicare offers a starting point. Its new ACCESS model supports technology-enabled chronic care and rewards providers for improving outcomes rather than simply delivering more services. Working with the FDA and NIH, ACCESS could provide a structured setting to test these agents, determine whether they improve medical outcomes, and create a path toward approval.

A second barrier is collaboration. Safely developing these agents for patient use at minimal cost will require clinicians, medical societies, and health systems to work with AI companies such as OpenAI, Anthropic, and Google. In parallel, independent researchers will need to test reliability and ensure that false alarms do not overwhelm clinicians.

The existential risks posed by increasingly powerful AI systems demand serious safeguards. But our nation should not treat every use of AI as equally dangerous. Yes, the risk of moving too fast is real. So is the danger of standing still.

Robert Pearl, the author of “ChatGPT, MD,” teaches at both the Stanford University School of Medicine and the Stanford Graduate School of Business. He is a former CEO of The Permanente Medical Group.