AI Agents in Healthcare in 2026: Real Uses, Rules and Limits

In 2026, AI agents and assistants in healthcare mostly draft visit notes, answer practice phones and sort patient requests, and help radiologists read images. None of them makes a diagnosis on its own: a clinician signs off. This page sums up, with dated sources, what is actually deployed, what regulators allow (FDA, HIPAA, EU rules) and the main risks, for clinicians and curious patients alike.

This article covers tools and rules; it is not medical advice. In an emergency, call 911.

What changed since the February 2025 version: OpenAI launched ChatGPT Health (U.S. only), the first randomized trial of AI scribes was published, the FDA’s list of AI-enabled devices passed 1,600 entries, and the EU pushed the AI Act’s medical device obligations back to 2028. All sources were checked Sept. 25, 2026.

Agent or Assistant: What’s the Difference?

An assistant answers or drafts on request: it transcribes a visit, summarizes a chart, explains a lab result. An agent also carries out actions in software: booking an appointment, routing a refill request, preparing a letter. In healthcare the line matters, because a tool that acts without human review can change regulatory status (see below).

In the deployments documented in 2026, a human stays in the loop: the physician reviews the note, staff handle the routed request, the radiologist confirms the flagged image.

Where AI Is Actually Used in 2026

Clinical Documentation: Ambient AI Scribes

An ambient scribe listens to the visit (with the patient’s consent) and drafts a structured note that the clinician edits and signs.

• Microsoft Dragon Copilot (successor to DAX Copilot) is sold in the United States, Canada, the UK, Ireland, France, Germany, Austria, the Netherlands, Belgium and Switzerland (Microsoft partner announcement, April 1, 2026).
• Nabla, a French startup, was one of the two scribes tested in the randomized trial below.
• Other vendors exist; none of these scribes appears on the FDA’s list of AI-enabled medical devices.

What the randomized trial found. Published in NEJM AI on Nov. 26, 2025, the UCLA trial assigned 238 outpatient physicians from 14 specialties to DAX Copilot, Nabla or usual care for two months (Lukac et al., NEJM AI):

• Nabla cut time spent writing notes by 9.5% versus the control group; DAX showed no significant change (-1.7%).
• Both tools improved burnout and task-load scores, a secondary finding the authors say needs confirmation in larger, multicenter trials.
• Physicians reported clinically significant inaccuracies “occasionally,” and one mild adverse event was reported.

The gain is real but modest, and review remains essential.

Regulatory status. In the UK, the MHRA clarified on July 29, 2026, that a tool that only transcribes, summarizes or drafts documentation from a consultation, with clinician review, is not a medical device; it becomes one if it supports diagnosis or treatment, or takes automated action without review (MHRA).

Phone Handling and Request Routing

This is where tools come closest to true agents. In France, Doctolib launched a phone assistant on Dec. 2, 2025, already deployed in Germany, that answers a practice’s calls, books appointments straight into the clinician’s calendar and routes other requests (prescription renewals, certificates, sick notes) to the practice inbox. According to Doctolib, in-house models flag situations that need a human (Doctolib announcement, in French).

Clinical triage is a different matter: the EU AI Act lists emergency patient triage systems as high-risk (see the rules section).

Imaging: The Most Authorized Field

Radiology accounts for most authorized AI devices. The FDA’s list of AI-enabled medical devices (page current as of Sept. 4, 2026, with decisions through June 29, 2026) contains 1,614 devices, including 1,230 in radiology (76%), 154 cardiovascular and 73 in neurology. 335 of them were authorized in 2025, the highest yearly count on the list.

One of the strongest results comes from breast cancer screening. Sweden’s MASAI trial, published in The Lancet on Jan. 31, 2026, covered 105,934 women: AI-supported reading detected more cancers (sensitivity of 80.5% vs. 73.8% for standard double reading) with no more false alarms (specificity of 98.5% in both groups), a non-inferior interval cancer rate and a lower reading workload for radiologists (Gommers et al., The Lancet).

Research and Clinical Questions

Assistants are also used to search the medical literature. Since April 22, 2026, OpenAI has offered ChatGPT for Clinicians, free for verified U.S. clinicians, with clinical search that cites its sources; since Sept. 1, 2026, a Healthcare Public Data plugin searches public sources (research, clinical trials, drug information, Medicare data) read-only, without access to patient charts (ChatGPT release notes). For organizations, ChatGPT for Healthcare is an enterprise version designed to support HIPAA-compliant use, with a Business Associate Agreement available.

ChatGPT Health and Google Health Coach: What Consumers Can Use

Health is one of the top uses of chatbots: according to its own analysis, OpenAI says more than 230 million people a week ask ChatGPT health and wellness questions (OpenAI announcement, Jan. 7, 2026). Two dedicated consumer offers exist.

Sources, checked Sept. 25, 2026: OpenAI Health help article, Google Health help and Google Store.

ChatGPT Health. It started as a waitlisted pilot on Jan. 7, 2026, and opened more widely to U.S. users on July 23, 2026, on web and iOS. OpenAI says connected medical records and Apple Health data are not used to train its foundation models or to target ads. Worth knowing: since Sept. 14, 2026, new Health connections default to your global plugin setting, “Allow low-risk actions,” so ChatGPT can use connected data without asking each time; you can change this under Settings → Plugins → Health.

Google Health Coach. Available since May 19, 2026, it is built with Gemini but runs separately from the Gemini app. Google warns that its answers “may be inaccurate or incomplete” and that the app is not intended to diagnose, treat or monitor any disease (Google blog).

What the research shows. A randomized study published in Nature Medicine on Feb. 9, 2026, gave ten medical scenarios to 1,298 people in the UK. On their own, the models (GPT-4o, Llama 3, Command R+) identified the right condition in 94.9% of cases. Used by participants, they led to the relevant condition in fewer than 34.5% of cases and to the right course of action (GP, ambulance, etc.) in fewer than 44.2%, no better than a control group using its usual resources (Bean et al., Nature Medicine). The weak link is the conversation: what people describe, and what they take away from the answer.

What’s Allowed: FDA, HIPAA and EU Rules

FDA

The FDA keeps a public list of authorized AI-enabled devices (1,614 as of Sept. 4, 2026). It does not yet flag any device based on a large language model; the agency says it “will explore methods” to tag devices built on foundation models in a future update. Its final guidance on clinical decision support software, reissued Jan. 29, 2026, explains which decision-support tools for clinicians fall outside the device definition (FDA).

HIPAA and Consumer Health Apps

HIPAA applies to covered entities (health plans, clearinghouses, and providers that conduct standard electronic transactions such as billing) and their business associates; an organization that is neither does not have to comply (HHS). In practice, the health data you type into a consumer chatbot is generally outside HIPAA. Most health apps not covered by HIPAA must instead follow the FTC’s Health Breach Notification Rule, updated in 2024, which requires them to notify users of a breach (FTC).

EU: CE Marking and the AI Act

In Europe, software with a medical purpose is a medical device and needs a CE mark, usually after review by a notified body; under the Medical Device Regulation, decision-support software is at least class IIa (rule 11).

An AI-enabled medical device that requires a notified body is also a “high-risk” system under the AI Act. The Digital Omnibus on AI, Regulation (EU) 2026/1744 adopted July 8, 2026, moved the dates (EUR-Lex):

These obligations cover risk management, relevant and sufficiently representative training data, documentation and human oversight. For the broader picture, see our overview of AI regulation. France adds its own layer: its health authority (HAS) published generative AI guidance for professionals in October 2025, and health data hosted for providers must sit with a certified host inside the European Economic Area.

Limits and Risks

• Errors: scribes produce “occasional” inaccuracies, and a chatbot can give a correct answer that the user misreads. Clinicians remain responsible for checking AI output, as the MHRA points out.
• Bias: a model trained on one population may perform worse on another, which is why the AI Act requires representative training data for high-risk systems.
• Health data: a consumer chatbot is not your doctor’s HIPAA-covered system. Avoid uploading documents that identify you or someone else unless you accept the service’s terms.
• Autonomy: the more an agent acts alone (ordering a test, sending a message), the higher the risk and the requirements. In the UK, a scribe that acts without review is regulated as a medical device.
• Clinician workload: the UCLA trial suggests less burnout with scribes, but over two months only. Long-term effects, including on how junior clinicians learn, are still unmeasured.

FAQ

Can AI diagnose patients?

Not on its own. Authorized tools, mostly in imaging, help clinicians spot findings; the decision stays with them. ChatGPT Health and Google Health Coach both say they are not intended for diagnosis.

Is ChatGPT Health available?

Yes, in the United States: Health in ChatGPT is open to logged-in Free, Go, Plus and Pro users 18 and older, on web and iOS, as of Sept. 25, 2026. It is not available in the EU or the UK.

Is my data protected by HIPAA when I use ChatGPT or Gemini?

Generally not, because consumer chatbots are not HIPAA covered entities. OpenAI says connected Health data is not used for training or ads, and Google says Fitbit health data is not used for Google Ads, but those are company commitments, not HIPAA protections.

What AI tools do doctors use for notes?

Mostly ambient scribes such as Microsoft Dragon Copilot or Nabla. They listen to the visit, draft a note, and the clinician reviews it before it goes into the chart.

Are AI scribes FDA-approved?

None appears on the FDA’s list of AI-enabled medical devices as of Sept. 25, 2026. They are sold as documentation tools, and the clinician remains responsible for the final note.