AI in Healthcare Use Cases: What's Actually Working Now

This is a grounded look at AI in healthcare use cases: what’s actually deployed and delivering value today, and what’s still mostly promise.

AI in healthcare attracts both enormous optimism and understandable skepticism. Depending on what you read, artificial intelligence is either about to transform medicine or is another technology that performs impressively in research papers but struggles in the real world.

The reality is somewhere in between.

Some AI systems are already being used in clinical practice and delivering measurable value. Others are promising but still need stronger evidence. And some of the most ambitious claims — particularly around replacing doctors — remain well ahead of what the evidence supports.

Here is a grounded look at where AI in healthcare actually stands.

Where AI is already working

Medical imaging

Medical imaging is one of the clearest examples of AI moving from research into regulated clinical use.

AI-enabled systems are being used across radiology and other imaging specialties for tasks such as detecting abnormalities, assisting image interpretation, improving image quality, segmenting anatomical structures and helping clinicians prioritise cases.

This is no longer a small experimental category.

The U.S. Food and Drug Administration maintains a regularly updated list of AI-enabled medical devices authorised for marketing in the United States. As of September 2026, the FDA reports that it has authorised more than 1,600 AI-enabled medical devices across a range of clinical areas.

That does not mean an AI system replaces the radiologist or clinician. In many deployments, AI acts as an assistive layer: identifying potentially important findings, improving workflow or providing additional information that a clinician evaluates alongside the rest of the clinical picture.

That distinction matters.

The strongest case for AI in healthcare today is often not autonomous diagnosis. It is AI helping clinicians perform specific, well-defined tasks more efficiently or consistently.

Ambient clinical documentation

Another area showing meaningful real-world adoption is ambient clinical documentation.

These systems listen to a clinician-patient conversation and generate a draft clinical note, which the clinician reviews and approves. The objective is straightforward: reduce the amount of time clinicians spend typing and completing documentation after patient visits.

There is now growing evidence that this can make a practical difference.

A 2025 multicentre study published in JAMA Network Open evaluated ambient AI scribes among clinicians across six U.S. healthcare systems. After 30 days of use, the technology was associated with reductions in reported burnout, cognitive documentation burden and time spent documenting after hours, along with increased attention to patients during consultations.

These findings are encouraging, although they should be interpreted appropriately. Studies of this kind show associations in real clinical settings; they do not mean every ambient documentation product will produce the same outcomes in every hospital.

Still, this is a useful example of AI solving a relatively unglamorous but very real healthcare problem.

Where AI looks promising but remains early

Drug discovery and protein structure prediction

Machine learning is also changing parts of biomedical research and drug discovery.

AlphaFold is perhaps the best-known example. DeepMind’s AlphaFold demonstrated highly accurate protein-structure prediction at a scale that would have been extremely difficult using experimental methods alone.

That matters because understanding protein structure can help researchers investigate biological mechanisms and potentially support parts of the drug-discovery process.

Machine-learning techniques are also being explored for molecule generation, virtual screening, target identification and candidate prioritisation.

But there is an important distinction between accelerating a stage of research and accelerating the entire drug-development lifecycle.

A model may help researchers identify a promising molecule more quickly. That molecule must still pass laboratory testing, preclinical evaluation, clinical trials, regulatory review and eventually real-world safety monitoring.

AI can therefore accelerate parts of discovery without turning a decade-long pharmaceutical process into a six-month one.

Clinical risk prediction

Another promising area is predicting clinical deterioration — for example, identifying patients who may be at increased risk of sepsis, readmission or other adverse outcomes.

These models can perform well in retrospective datasets.

The difficult part begins when they leave the dataset and enter the hospital.

A well-known example is the external validation of a widely deployed proprietary sepsis prediction model published in JAMA Internal Medicine. Researchers found that its performance at an external health system was substantially weaker than might have been expected from earlier reported performance.

This illustrates a broader problem in healthcare AI: a model that works well on the data on which it was developed may not perform equally well across different hospitals, patient populations and clinical workflows.

And predictive accuracy is only part of the challenge.

A hospital also has to decide what happens when the model generates an alert. Who receives it? How quickly must someone respond? What happens when there are too many alerts? Does the intervention improve patient outcomes?

These are workflow and implementation questions, not simply machine-learning questions.

Where the hype still exceeds the evidence

Autonomous AI doctors

The idea of an AI system independently diagnosing patients and deciding treatment without meaningful clinician oversight remains far ahead of routine clinical practice.

Large language models can produce impressive responses to medical questions and perform strongly on some medical examinations and benchmarks.

But benchmark performance is not equivalent to clinical competence.

Real patients arrive with incomplete information, ambiguous symptoms, multiple conditions, medication interactions, social circumstances and changing clinical states. Medical decisions can also have serious consequences when they are wrong.

For these reasons, the near-term role of AI is much more likely to be decision support and augmentation than wholesale replacement of clinicians.

General-purpose AI across hospitals

Another attractive idea is a single AI platform that works reliably across every specialty, hospital and patient population.

Healthcare does not behave that neatly.

Data collection differs between hospitals. Patient populations differ. Clinical workflows differ. Equipment differs. Documentation practices differ.

A model trained in one environment can therefore lose performance when deployed somewhere else.

This is why external validation matters so much.

Regulatory approval is useful — but it is not the end of the story

Regulatory authorisation is an important signal because it means a medical device has gone through an applicable regulatory review process.

But regulatory clearance or authorisation should not be interpreted as proof that a product will improve outcomes in every clinical setting.

AI-enabled systems still need appropriate local validation, monitoring and governance after deployment.

Performance can change when patient populations, workflows or underlying data change. AI systems can also introduce new operational risks — including automation bias, inappropriate reliance on alerts and failure to recognise when a model is operating outside the conditions under which it was validated.

Healthcare AI therefore needs to be treated as a lifecycle rather than a one-time model deployment.

What good clinical evidence looks like

When evaluating an AI healthcare claim, I find three questions particularly useful.

Was the system tested prospectively?

Retrospective studies can provide valuable early evidence, but they evaluate historical data. Prospective studies tell us more about how a system behaves when it encounters patients and workflows in real time.

Was it tested outside the institution where it was developed?

External validation helps determine whether performance generalises beyond the original dataset and clinical environment.

Was it compared with existing clinical practice?

A model achieving impressive accuracy is not enough by itself. The more important question is whether using it improves something that matters: patient outcomes, diagnostic accuracy, clinician workload, treatment time, cost or another meaningful clinical measure.

Retrospective, single-site studies can still provide useful early evidence, but they can be more susceptible to optimistic performance estimates and may not generalise to different patient populations, institutions or clinical workflows.

These questions are not a perfect test of whether an AI system is ready for deployment, but they are useful filters for separating promising research results from stronger clinical evidence.

The pattern emerging

The most successful healthcare AI systems today tend to have something in common.

They do not try to replace medicine.

They solve a narrower problem.

They help interpret an image. They reduce documentation burden. They identify a patient who may require closer attention. They help researchers analyse biological structures or narrow a search space.

That may sound less dramatic than an autonomous AI doctor.

But it is also how important technologies often enter complex industries: not by replacing the entire system at once, but by making individual parts of it measurably better.

The interesting question for healthcare AI is therefore no longer simply:

“Can AI do this?”

A better set of questions is:

Does it work outside the lab? Does it improve something that matters? And does it continue to work safely when real clinicians and real patients depend on it?

That is where the evidence becomes much more interesting.

References and further reading

U.S. Food and Drug Administration. Artificial Intelligence-Enabled Medical Devices. FDA.
https://www.fda.gov/medical-devices/digital-health-center-excellence/artificial-intelligence-enabled-medical-devices
U.S. Food and Drug Administration. List of Artificial Intelligence-Enabled Medical Devices. FDA.
https://www.fda.gov/medical-devices/artificial-intelligence-enabled-medical-devices/list-artificial-intelligence-enabled-medical-devices
Olson KD, Meeker D, Troup M, et al. Use of Ambient AI Scribes to Reduce Administrative Burden and Professional Burnout. JAMA Network Open. 2025;8(10).
DOI: 10.1001/jamanetworkopen.2025.34976
https://jamanetwork.com/journals/jamanetworkopen/fullarticle/10.1001/jamanetworkopen.2025.34976
Shah SJ, Devon-Sand A, Ma SP, et al. Ambient artificial intelligence scribes: physician burnout and perspectives on usability and documentation burden. Journal of the American Medical Informatics Association. 2025;32(2):375–380.
DOI: 10.1093/jamia/ocae295
https://pubmed.ncbi.nlm.nih.gov/39657021/
Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596:583–589.
DOI: 10.1038/s41586-021-03819-2
https://www.nature.com/articles/s41586-021-03819-2
Wong A, Otles E, Donnelly JP, et al. External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients. JAMA Internal Medicine. 2021;181(8):1065–1070.
DOI: 10.1001/jamainternmed.2021.2626
https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2781307

This article discusses AI systems and research in healthcare for educational purposes. It is not medical advice and should not be used to make individual clinical decisions.

AI in Healthcare: What’s Actually Working