AI is already widely used in healthcare, and its footprint is growing fast. The U.S. Food and Drug Administration has authorized over 1,350 AI-enabled medical devices for marketing, with the vast majority designed for radiology and cardiovascular care. Beyond devices, AI now plays a role in diagnosing cancer, predicting life-threatening infections, guiding surgery, accelerating drug development, and even delivering mental health therapy.
Medical Imaging and Cancer Detection
Radiology is where AI has gained the most traction. Of the 1,357 AI-enabled devices listed by the FDA, roughly three-quarters fall under radiology. These tools analyze X-rays, CT scans, and mammograms to flag abnormalities that a human eye might miss, or might catch a few seconds later.
A systematic review published by the British Institute of Radiology pooled data from 23 studies and found that clinicians working with AI assistance detected cancers with 79% sensitivity, compared to 66% without it. Specificity (the ability to correctly rule out cancer) also improved, from 82% to 87%. For lung cancer on CT scans specifically, AI-assisted reading pushed sensitivity from 78% to 89%. These aren’t dramatic leaps in any single case, but across millions of scans per year, the difference translates to a meaningful number of earlier diagnoses.
Predicting and Catching Sepsis Earlier
Sepsis, the body’s deadly overreaction to infection, kills more than 250,000 Americans each year. Speed matters enormously: every hour of delayed treatment raises the risk of death. AI early warning systems monitor patient vitals in real time and alert clinicians when patterns suggest sepsis is developing, sometimes hours before traditional screening would catch it.
Results vary depending on how well these systems are integrated into clinical workflows. One implementation led to a 39.5% reduction in in-hospital mortality, a 32.3% reduction in length of stay, and 22.7% fewer 30-day readmissions. Another hospital system, called TREWS, showed improved mortality rates for high-risk patients and shorter hospital stays after adjusting for how sick patients were on arrival. But not every rollout succeeds. An eight-month trial at two hospitals found no significant improvements in mortality or ICU transfers, largely because clinicians didn’t change their behavior in response to the alerts. The technology works, but only when the humans around it are trained and motivated to act on its recommendations.
Drug Discovery Gets Faster
Developing a new drug traditionally takes over a decade and costs billions. AI is compressing the early stages of that process. Current AI-enabled workflows are cutting early discovery timelines by 30 to 40 percent and shrinking the preclinical candidate development phase from a typical three to four years down to 13 to 18 months. In antibody design, AI-driven approaches are achieving hit rates of 16 to 20 percent, compared to a computational benchmark of just 0.1 percent.
The important caveat: AI hasn’t yet shortened the parts of drug development that take the longest. Clinical trials, regulatory review, and manufacturing scale-up still proceed at their traditional pace. So while AI is delivering real savings in the lab, the total time from concept to pharmacy shelf hasn’t changed as dramatically as some headlines suggest.
Robotic Surgery With AI Guidance
Surgeons have used robotic systems for years, but AI is adding a layer of real-time decision support. AI-assisted robotic procedures have shown a 25% reduction in operative time and a 30% decrease in complications during surgery compared to manual methods. Recovery times shortened by an average of 15%, and patients reported lower pain scores afterward.
In spinal surgery, one study found complication rates dropped from 12.2% with traditional techniques to 6.1% with AI-assisted robotic systems, along with shorter hospital stays. These systems don’t replace the surgeon. They provide enhanced visualization, steadier instrument control, and alerts when the surgical path approaches sensitive structures.
Wearables That Watch Your Heart
AI-powered wearable devices are turning consumer electronics into screening tools, particularly for atrial fibrillation (AFib), a heart rhythm disorder that raises stroke risk fivefold. The Apple Heart Study enrolled more than 400,000 participants with no prior AFib diagnosis. Among those flagged for an irregular pulse, the positive predictive value was 84%, meaning the watch was right about four out of five times.
Performance varies by device and population. A study in elderly patients with known AFib risk factors found sensitivity and specificity both above 98%. A smartphone-based app called FibriCheck showed 96% sensitivity and 97% specificity in primary care settings. But in a more real-world scenario with broader patient groups, one study found both sensitivity and specificity dropped to about 68%. The takeaway: these tools are quite good at catching AFib in people who are already at risk, but less reliable as a general screening tool for everyone.
Mental Health Chatbots
AI chatbots that deliver cognitive behavioral therapy are now being tested in randomized clinical trials, not just marketed as wellness apps. A trial published in NEJM AI compared a generative AI chatbot called Therabot to a control group over eight weeks. Therabot users showed significantly greater reductions in depression symptoms at both four and eight weeks, with effect sizes in the moderate-to-large range. Participants used the chatbot for an average of more than six hours over the study period and rated their connection with it as comparable to working with a human therapist.
These tools aren’t positioned as replacements for human therapists. They’re filling a gap: the majority of people with depression and anxiety never receive treatment, often because of cost, stigma, or long wait times. A chatbot available at 2 a.m. on a Tuesday addresses a different need than a weekly therapy session.
Personalized Cancer Treatment
AI is helping oncologists match patients to the treatments most likely to work for them. By analyzing clinical records, genomic data, and medical imaging together, AI models can predict how individual patients will respond to specific therapies. One study focused on kidney cancer patients undergoing targeted drug therapy developed an AI model that predicted whether patients would survive beyond three years with over 93% accuracy. That kind of precision helps doctors avoid treatments that are unlikely to benefit a particular patient, sparing them unnecessary side effects.
The broader vision is a system where AI integrates a patient’s tumor genetics, medical history, and imaging results to recommend a treatment plan tailored specifically to them. This is already happening in pockets, particularly at large cancer centers, but it’s far from standard practice everywhere.
Administrative Tasks and Paperwork
Some of AI’s most immediate impact in healthcare has nothing to do with diagnosis or treatment. It’s reducing the paperwork burden that consumes a significant chunk of every clinician’s day. AI systems now automate appointment scheduling, patient communications, billing, coding, and claims processing. By handling these tasks, they reduce errors, speed up reimbursement, and free up staff to focus on patient care.
Virtual health assistants powered by AI manage patient triage, answer routine questions, and handle pre-visit intake, reducing the administrative load on nurses and front desk staff. Revenue cycle management, the complex process of billing insurers and collecting payment, is one area where AI adoption is accelerating because the return on investment is immediate and measurable. Fewer coding errors mean fewer denied claims, which means faster payments.
What’s Holding AI Back
Despite these advances, AI in healthcare faces real limitations. Algorithms trained on data from one hospital or one demographic group don’t always perform well elsewhere. Concerns about racial and socioeconomic bias in healthcare algorithms have prompted calls for more rigorous testing across diverse patient populations before deployment. The sepsis example illustrates another challenge: even accurate AI predictions are useless if clinicians ignore the alerts or don’t have workflows designed to respond to them.
Regulation is also playing catch-up. The FDA’s authorization process for AI devices is evolving, and questions remain about how to handle algorithms that continuously learn and change after they’re deployed. Privacy concerns around the massive datasets needed to train these systems add another layer of complexity. AI is firmly embedded in healthcare today, but the gap between what it can do in controlled studies and what it reliably delivers in everyday clinical settings remains significant.