In healthcare, the AI agents actually running at scale do paperwork, not diagnosis: ambient scribes that draft a clinical note from the visit conversation, tools that draft patient-message replies, and back-office agents for prior authorization. In every working deployment, a clinician reviews and signs — the medical decision stays human.
Healthcare is a regulated, high-stakes industry, so it's the clearest test of this site's one rule: automate the work, keep a hand on the decision. The agents that are real here cluster on documentation and administration — the tedious work around care — while the clinical call stays with a licensed professional. Below: what's genuinely deployed, the review gate every real build keeps, and where the human line is drawn by policy, not preference.
What an "AI agent for healthcare" actually does today
The real deployment is ambient clinical documentation — an agent that listens to the visit and drafts the note. Microsoft's Dragon Copilot, which merges its Dragon Medical One dictation and DAX Copilot ambient-listening products, drafts clinical notes, referral letters, and after-visit summaries from the conversation in the room. Per the December 2025 athenahealth collaboration announcement, the underlying Dragon ambient technology is already used by more than 650 healthcare organizations, and athenahealth's Ambient Notes is available to over 170,000 clinicians on its network. The notes are drafted "ready for review and editing" — the clinician, not the agent, owns what goes in the chart.
The same shape shows up inside the EHR itself. Abridge built "Abridge Inside" directly into Epic's clinician workflow, and its Linked Evidence feature lets a clinician trace every line of the AI-drafted summary back to the source transcript — so the draft is auditable before anyone signs it. That traceability is the healthcare version of the discipline this whole site keeps: the agent does the work where you can read its sources, and a person commits.
A second real job is the patient inbox. Epic's generative-AI feature for drafting replies to MyChart messages — reported by eMarketer, relaying the New York Times, as in use by nearly 15,000 doctors and assistants across more than 150 health systems — generates a draft that clinicians "review and approve, or change, before replying." Epic's own R&D leader told Becker's Hospital Review the tool has guardrails to stop it giving clinical advice and "is not designed to improve clinical outcomes." In UNC Health's pilot, clinicians found the drafts most useful for routine replies — work notes, medication refills — and said the tool cannot substitute for clinical judgment. This is the same drafting pattern that holds across every job on this site: the agent gets you to 80%, a person owns the last mile.
The number everyone quotes, and the one that matters
The headline is the scale — 650-plus organizations on Dragon, 150-plus health systems drafting patient replies. It's real, and it's the part every vendor deck leads with. But the number that runs a safe clinic isn't how much the agent drafts; it's whether a clinician actually read and corrected the draft before it entered the record or reached the patient.
That's not a nicety — it's the design. Abridge's Linked Evidence exists so the review can be fast and grounded; Epic's reply drafts are explicitly "review and approve, or change" (eMarketer), not auto-send. An ambient scribe that drafts a flawless-looking note still records a hallucinated detail if nobody checks it, and in a medical chart that error compounds. The gauge is review quality, not draft volume — the same lesson the customer-service deployments relearned the hard way, with a patient's safety instead of a refund on the line.
Where the human line has to stay — and who draws it
In healthcare the line isn't a preference; a professional body draws it. In June 2026 the American Medical Association adopted policy stating that AI should assist rather than replace physician judgment, and that transparency, accountability, and physician oversight are required whenever AI touches patient care. The AMA specifically opposes autonomous or semiautonomous AI as a substitute for physician review in insurance coverage determinations. That is the clinical-decision gate, written into policy: the agent drafts, researches, and summarizes; a physician decides.
It maps exactly onto the rule this site states in general terms — gate the irreversible, let the reversible run. In a clinic, drafting a note, retrieving a record, and summarizing a message are reversible work an agent can own. A diagnosis, a prescription, a coverage denial, a care-plan change — those are the irreversible, high-stakes calls that keep a human's name on them. The healthcare agents that get in trouble are the ones that blur that line to make a throughput chart go up.
The other non-negotiable is the data. Clinical agents handle protected health information, which makes the security risks of AI agents concrete: an agent with broad access to patient records is a breach surface, so least-privilege access and audit logging aren't optional in a care setting. Automate the work on the record; govern the access to it.
The administrative back office
Away from the bedside, the sturdiest healthcare use case is the paperwork payers and providers fight over: prior authorization and claims. This is a draft-and-submit job with a human approval, structurally identical to the invoice-processing split — an agent assembles the authorization request from the chart, a person confirms it. The regulatory clock is pushing it: CMS's Interoperability and Prior Authorization Final Rule (CMS-0057-F) shortens required decision timelines and, generally beginning January 1, 2026, adds operational requirements for impacted payers, with a standardized Prior Authorization API due by January 1, 2027. Agents that prepare and track these requests fit the new rails — but, per the AMA, the coverage decision itself is exactly where autonomous AI shouldn't sit.
If you're new to the whole idea of what an agent is and isn't, the agent basics primer starts from scratch.
FAQ
What AI agents can help healthcare teams? The ones deployed at scale do documentation and admin, not diagnosis: ambient scribes that draft the clinical note from the visit (Microsoft Dragon Copilot, Abridge inside Epic), tools that draft patient-message replies (Epic's MyChart drafts), and back-office agents for prior authorization and claims. In each, a clinician or staffer reviews before anything is final.
Can an AI agent diagnose patients or make clinical decisions? No — not on its own, by design and by policy. The AMA's June 2026 policy says AI should assist rather than replace physician judgment, and Epic says its patient-reply tool has guardrails against giving clinical advice and "is not designed to improve clinical outcomes" (Becker's). The diagnosis and the prescription stay with a licensed clinician.
Is an AI scribe safe to use in a medical record? It's safe as a drafting tool that a clinician reviews and signs — which is how Abridge (with traceable Linked Evidence) and Dragon Copilot (notes "ready for review and editing") ship it. The risk is an unreviewed note: the fix is the same human-in-the-loop gate this site applies everywhere, plus least-privilege access to protected data.
Can AI agents handle prior authorization? Yes, as a prepare-and-submit job with a human approval. CMS's prior-authorization final rule adds operational requirements from January 1, 2026 and a standardized API by 2027, which agents that assemble and track requests can plug into — the same draft-vs-pay discipline as invoice processing. The coverage decision, per the AMA, stays with a physician.
The healthcare agents that work follow the same discipline as every build on this site: they do the paperwork and stop before the step you can't take back — the diagnosis, the prescription, the note nobody read. Want the real builds, the exact jobs each professional automates and the ones they still sign by hand? Subscribe free and get each week's build in your inbox.