Artificial intelligence in healthcare no longer refers to a single class of experimental diagnostic tools. One system may analyze a medical image for a suspicious finding. Another may estimate a hospitalized patient's risk of deterioration. A third can listen to a clinical visit and prepare a draft note for the physician to review.

The technology is also spreading into less visible parts of healthcare. Hospitals are using predictive models for scheduling and billing. Clinicians are using generative AI to summarize medical research and draft patient communications. Researchers are applying natural language processing to clinical-trial recruitment, while pharmaceutical developers are incorporating AI into parts of drug development.

National data show how broad that adoption has become. In 2024, 71% of nonfederal U.S. acute care hospitals reported using predictive AI integrated with their electronic health records, according to the Assistant Secretary for Technology Policy, or ASTP. The figure came from the American Hospital Association Information Technology Supplement, which surveyed 2,253 such hospitals and had a 51% response rate. Results were weighted to account for differences in the likelihood that hospitals responded.

Generative AI was at an earlier stage. A 2025 JAMA Network Open study using the same national hospital survey found that 31.5% of 2,174 responding hospitals reported using generative AI integrated with their EHR in 2024. Another 24.7% said they planned to adopt it within a year. That second figure represented an intention at the time of the survey, not confirmed subsequent deployment.

The two figures are also not directly interchangeable. Predictive AI includes statistical and machine-learning systems that classify patients or generate risk scores. Generative AI generally creates new content, such as clinical notes, summaries, or messages. Hospitals can use both.

Medical imaging remains one of the most established areas for regulated healthcare AI. Algorithms can help identify suspected abnormalities, outline anatomical structures, quantify features, prioritize cases for review, or improve parts of image acquisition and reconstruction.

The U.S. Food and Drug Administration's current AI-Enabled Medical Device List shows authorized products across radiology as well as cardiovascular medicine, neurology, gastroenterology, anesthesiology, pathology, and other specialties. Recent 2026 entries include AI-enabled ultrasound, mammography, CT, sleep-analysis, and image-segmentation systems.

It is important to understand what inclusion on that list means. The FDA says listed devices have met applicable premarket requirements for their intended uses, including review of safety and effectiveness. Depending on the product, the regulatory pathway may involve clearance, authorization, or approval. The list is not a comprehensive catalogue of every healthcare AI product, and FDA authorization does not mean that an algorithm is appropriate for uses outside its specified indications.

AI is also moving beyond conventional radiology. In May 2026, for example, the FDA cleared ArteraAI Breast through the 510(k) pathway. The agency classified it as a pathology software algorithm that analyzes digital images for breast-cancer prognosis. The regulatory decision found the device substantially equivalent to an applicable legally marketed predicate, which is the standard used in the 510(k) pathway. It was a clearance, not a premarket approval.

Research in computational pathology is testing other approaches. A 2025 Nature Medicine study evaluated an AI model designed to help identify lung cancers likely to carry an EGFR mutation from digital pathology slides. In a prospective silent trial, where the model was evaluated without directing patient care, it achieved an area under the receiver operating characteristic curve of 0.890. The researchers estimated that an AI-assisted workflow could reduce rapid molecular tests by as much as 43% while maintaining the clinical-standard performance used in the study. Those findings came from a specific research setting and should not be interpreted as an industry-wide reduction in testing.

Another major use of AI happens inside the electronic health record.

Predictive models can combine information such as diagnoses, laboratory results, medications, vital signs, and previous encounters to estimate the likelihood of a future event. Hospitals use these systems for purposes including early disease detection, readmission risk, inpatient deterioration, fall risk, identifying high-risk outpatients, and predicting missed appointments.

ASTP defines predictive AI in its hospital analysis as the use of statistical analysis and machine learning to classify individuals or produce a risk score. Among hospitals that reported using predictive AI, the most common application in 2024 was predicting health trajectories or risks for inpatients.

This type of system does not necessarily make the clinical decision. A risk model may instead flag a patient for further assessment, helping a clinical team decide where attention may be needed.

That distinction matters because adoption data do not demonstrate that every deployed model improves patient outcomes. A hospital reporting use of predictive AI says that the technology is present in its workflow. Determining whether a particular model is clinically useful requires separate evaluation of measures such as accuracy, false positives, false negatives, bias, and effects on actual care.

Generative AI has opened a different path into healthcare. Rather than primarily predicting a clinical event, it can create text from existing information.

Physicians are using it to summarize research, prepare clinical documentation, generate chart summaries, produce draft patient messages, translate information, and create care instructions.

The American Medical Association's 2026 Physician Survey on Augmented Intelligence found that 81% of physician respondents were using AI professionally, compared with 38% in its 2023 survey. The 2026 survey included 1,692 U.S. physicians across specialties, practice settings, and career stages and was fielded between January 15 and February 2. The results were self-reported.

The comparison also requires some caution. The AMA expanded the evaluated list from 15 AI use cases to 17 in 2026 and changed parts of the questionnaire, meaning the multiyear figures are better viewed as evidence of a strong adoption trend than as a perfectly unchanged statistical series.

The most commonly reported use was summarizing medical research and standards of care, cited by 39% of physicians. Thirty percent reported using AI to create discharge instructions, care plans, or progress notes. Twenty-eight percent used it for billing codes, medical charts, or visit documentation, while another 28% reported AI-generated chart summaries. Nineteen percent used AI to draft responses to patient portal messages.

One of the most visible generative AI applications is the ambient clinical scribe.

These systems can capture a conversation between a clinician and patient, process what was discussed, and generate a draft clinical note. The clinician can then review and edit the documentation before it enters the medical record.

A 2025 JAMA Network Open quality-improvement study examined one ambient AI scribe across six U.S. health systems. The analysis included 263 physicians and advanced practice practitioners providing ambulatory care. Among 184 participants included in the adjusted burnout model, the estimated proportion experiencing burnout declined from 51.9% before using the system to 38.8% after 30 days. Participants also reported less after-hours documentation and lower documentation-related cognitive workload.

Those results show an association rather than proof that the AI system caused the change. The study did not have a randomized control group, participants volunteered, and only 272 of 451 enrolled clinicians completed both the before and after surveys. The findings are therefore encouraging evidence about workflow effects, not a general estimate of how much every ambient scribe will reduce burnout.

That distinction reflects a broader issue in healthcare AI. A technology may work well as a drafting assistant while still requiring clinicians to verify whether the resulting note accurately represents the encounter.

AI can also operate while a procedure is taking place.

During colonoscopy, for example, computer-aided detection systems can analyze the live video feed and visually flag regions that may contain polyps. The physician remains responsible for examining the area and deciding what action to take.

Evidence from randomized trials shows that some systems can increase detection under particular conditions. The COLO-DETECT trial, conducted across 12 National Health Service hospitals in England, randomly assigned 2,032 participants to AI-assisted or standard colonoscopy.

At least one adenoma was detected in 56.6% of participants with AI-assisted colonoscopy, compared with 48.4% with standard colonoscopy. The mean number of adenomas detected per procedure was also higher in the AI-assisted group.

The outcome was adenoma detection, not a direct measurement of subsequent colorectal cancer incidence or mortality. Other trials have produced results that vary by setting, technology, endoscopist experience, and the comparison technique. A 2026 randomized trial in Brazil, for example, found no statistically significant difference in adenoma detection when AI-assisted linked color imaging was compared with linked color imaging alone.

The evidence therefore illustrates both the potential and the limits of generalizing from a single AI system.

Some of healthcare AI's fastest-growing applications are operational rather than diagnostic.

Hospitals use predictive models to forecast appointment no-shows, improve scheduling, automate parts of billing, allocate resources, and identify patients who may require follow-up. Generative systems can help create coding suggestions, summarize records, and prepare administrative communications.

ASTP's hospital analysis found that among nonfederal acute care hospitals using predictive AI, the share reporting its use to simplify or automate billing climbed from 36% in 2023 to 61% in 2024. Use for scheduling increased from 51% to 67%. These were increases of 25 and 16 percentage points respectively, making them the fastest-growing applications measured in the survey.

These percentages describe reported use among hospitals already using predictive AI. They are not shares of all healthcare providers, nor do they measure financial savings generated by the technology.

AI is also entering healthcare payment administration. The Centers for Medicare and Medicaid Services launched its Wasteful and Inappropriate Service Reduction model in 2026 across six states. CMS says participating organizations can use enhanced technologies including AI and machine learning, together with human clinical review, for review of selected Medicare items and services. The model runs from January 2026 through December 2031. It is a defined federal demonstration model, not a description of all Medicare prior authorization nationwide.

Clinical research creates another natural use for AI because trial eligibility criteria can require teams to search large quantities of structured and unstructured medical information.

A 2026 study published in JCO Clinical Cancer Informatics described a Yale system using artificial intelligence and natural language processing to prescreen cancer patients for clinical trials. Since September 2022, the system had screened 98,348 patients across 29 trials, identified 825 eligible candidates, and facilitated 117 enrollments, according to the researchers.

For patients who still required manual review, the system reduced chart-review workload tenfold and shortened average screening time from 3.1 minutes to 1.8 minutes per chart. These are results from one implementation rather than a benchmark for clinical-trial recruitment generally.

The practical role of the system was not to decide who should enter a trial. It narrowed the pool so research teams could concentrate their review on patients more likely to meet complex eligibility criteria.

AI use in healthcare begins well before a medicine reaches a physician or patient.

Pharmaceutical and biotechnology researchers are applying AI in areas that include analysis of biological and clinical data, biomarker assessment, trial design, dose optimization, real-world data analysis, manufacturing, and regulatory work.

The scale can easily be overstated if regulatory figures are described incorrectly. The FDA's Center for Drug Evaluation and Research says its experience included more than 500 regulatory submissions containing AI components between 2016 and 2023. That figure refers to submissions in which AI appeared somewhere in the development or regulatory evidence. It does not mean that 500 AI-designed medicines were approved.

The FDA has been developing a risk-based framework for evaluating AI models used to generate information supporting regulatory decisions about drug and biological products. This reflects a wider shift in healthcare regulation as agencies increasingly have to assess not only the medical product, but also how an AI model was developed, validated, monitored, and used within a particular context.

Generative AI can also sit between complex medical information and the patient.

Systems can draft responses to portal messages, translate medical language, convert technical information into plainer explanations, and prepare discharge or care instructions. The AMA's 2026 data show these activities are already part of physician workflows, although their prevalence varies considerably by task.

The central design choice is often whether AI communicates directly with a patient or prepares material for a clinician to review.

The second model preserves a human checkpoint. That can be especially important when source information is incomplete, a question requires clinical judgment, or generated language sounds convincing despite containing an error. Generative systems can produce fluent text without guaranteeing that every statement is medically correct.

For that reason, the most consequential question is not simply whether AI can write a response. It is how the response is checked, who remains accountable, and whether the system is being used for a task appropriate to its tested capabilities.

As healthcare AI moves into routine workflows, evaluation is becoming part of deployment rather than something that happens only before purchase.

In ASTP's 2024 hospital data, 82% of hospitals using predictive AI reported evaluating models for accuracy, 74% evaluated for bias, and 79% performed post-implementation evaluation or monitoring. ASTP noted, however, that fewer hospitals performed those checks across most or all of their models. There was also uncertainty inside some organizations, with 15% responding that they did not know about accuracy evaluation, 21% about bias evaluation, and 18% about post-implementation monitoring.

Physicians are asking for similar safeguards. In the AMA's 2026 survey, 88% said validation of safety and efficacy was important for wider adoption, while 86% cited assurances around data privacy. Eighty-five percent wanted to be consulted or directly involved in AI adoption decisions within their practices.

International health organizations are emphasizing the same principle. A 2026 World Health Organization discussion paper identified bias, opacity, equity, data governance, and regulatory gaps as important risks and called for human oversight and risk-based governance. WHO's framing is that AI should augment human judgment rather than replace it.

The evidence now shows a healthcare AI landscape that is much broader than the idea of an autonomous machine diagnosing patients.

Some of the most established tools perform narrow, well-defined jobs such as analyzing images or calculating risk scores. Generative AI is taking on language-heavy tasks such as documentation, summaries, and draft communications. Other systems operate almost entirely behind the scenes, helping hospitals schedule appointments, process administrative work, screen potential research participants, or analyze information used in drug development.

The evidence supporting these applications is not uniform. Some medical uses have been tested in randomized trials or reviewed through FDA regulatory pathways. Others are supported primarily by observational studies, quality-improvement projects, institutional deployments, or self-reported adoption surveys. A regulatory authorization, a hospital reporting that it uses AI, and a randomized trial showing an improvement in a clinical measure are three different kinds of evidence.

That difference is central to understanding artificial intelligence in healthcare. Adoption is already widespread, but adoption alone does not establish clinical benefit. The more meaningful question is increasingly whether each AI system performs a clearly defined job accurately, improves the workflow or outcome it is intended to affect, and remains subject to appropriate human review.

In 2026, that is what healthcare AI increasingly looks like in practice. It is less a replacement for doctors, nurses, researchers, or administrators than a growing layer of software working alongside them across almost every stage of care.