Back to all writing

Agentic AI in healthcare, explained plainly: AI that finishes the task

A chatbot tells a patient when their appointment is. An agent moves it. That difference is simple to say, harder to build, and the main thing to check before you believe any vendor using the word.

, November 20, 2026, 9 min read

Agentic AI in healthcare means AI that completes a task rather than answering a question. Ask a chatbot when your next appointment is, and it tells you. Ask an agent to move it to Thursday afternoon, and it checks the schedule, moves the appointment in the EHR and texts you a confirmation. The idea is that simple. Building one that behaves safely in a medical practice is not.

The word is everywhere this year, and a lot of what wears it is a chatbot with a new label. So this post does three things: explains what an agent actually is, shows what agents do in a practice today, and gives you the questions that separate a real one from a rebrand.

What is agentic AI in healthcare?

Agentic AI is software that is given a goal, works out the steps, uses connected systems to carry them out, and checks whether it succeeded, with limited human input along the way. In healthcare the useful version is narrow: an agent with permission to do a small number of specific things, such as booking an appointment in the EHR, and a firm rule for handing everything else to a person.

Every agent, whatever it is sold as, has four parts. It is worth knowing them because each one is a question you can ask a vendor.

  • A goal. "Reschedule this patient to a slot that suits them with the same provider," not "help the patient"
  • Tools and permissions. The specific systems it can read and write: the schedule, the patient record, the provider's task inbox, a text message gateway
  • Steps. The ability to work out the order: find the patient, find the appointment, find open slots, offer them, book one, cancel the old one, confirm
  • Stop conditions. When to hand off to a person, and what it must never do even if asked

The fourth part is where I spend most of my time. An agent that can do many things is easy to demo. An agent that reliably knows when to stop is what you can put on a medical practice's phone line.

How is agentic AI different from a chatbot or generative AI?

The plainest way to see it is to follow one request, a refill, through each kind of system.

Kind of systemWhat it does with "I need my blood pressure medication refilled"What exists afterward
Phone tree or rules automationPlays a menu, routes to a voicemail boxA voicemail someone has to transcribe
Chatbot or conversational AIExplains the refill policy and suggests calling the officeNothing new in any system
Generative AIDrafts a message summarizing the requestA draft someone has to send
Agentic AIIdentifies the patient, confirms the medication and pharmacy, files a task in the provider's inboxA structured refill request waiting for the clinician

Notice what the agent does not do: approve the refill. That is a clinical decision and it belongs with the clinician. A well built healthcare agent is defined as much by where it stops as by what it does.

What does agentic AI actually do in a medical practice today?

Less than the conference slides, and enough to matter. Here is my honest read on each area, from most ready to least.

TaskWhat an agent doesMaturity in 2026Where a person stays in the loop
Scheduling callsBooks, reschedules and cancels in the EHR during the callReady, when it writes to the EHRExceptions, distressed callers, clinical questions
New patient intakeCollects demographics, reason for visit and insurance details, then booksReadyCoverage checks and unusual registrations
Refill requestsCaptures patient, medication and pharmacy as a task for the providerReady for intakeThe clinician approves or declines
Prior authorizationGathers chart documentation and tracks statusEarlyMost of the process, until payers are connected
Claim denialsSorts denials by cause and drafts appealsEarlyBilling staff decide and submit
Clinical decisionsSuggests a diagnosis or urgencyShould not act on its ownAlways the clinician

Prior authorization is the row people most want to be further along. It is held back by the other side of the transaction. The CMS Interoperability and Prior Authorization Final Rule requires impacted payers, including Medicare Advantage and Medicaid plans, to implement a prior authorization API beginning January 1, 2027. Until payers are reliably on the other end, an agent can assemble the packet but cannot finish the job.

Agents are arriving on the payer side too, and physicians are wary. In the AMA's 2025 prior authorization survey of 1,000 physicians, released in May 2026, six in ten said they worry AI could further raise denial rates. That is worth remembering when someone describes agentic AI as purely a gift to practices.

Why are so many agentic AI projects failing?

Gartner predicted in June 2025 that more than 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs and unclear business value. Its analyst Anushree Verma said that most agentic AI projects right now are early stage experiments or proofs of concept. Gartner also described "agent washing," vendors rebranding chatbots and assistants as agents, and estimated that only about 130 of the thousands of vendors claiming agentic AI are real.

The same firm still expects growth: it predicts that at least 15 percent of day to day work decisions will be made autonomously through agentic AI by 2028, up from zero in 2024. Both things can be true. Most projects fail and a minority do real work.

In healthcare, the failures I see come from two mistakes. The goal is vague ("handle patient communication") so nobody can tell whether it worked. Or the permissions are broad, so the agent can touch things it should not and the practice, sensibly, refuses to switch it on. Narrow goals with checkable outcomes are what survive.

What should an AI agent be allowed to do on its own?

These are my rules for any agent that touches a medical practice's systems.

  • Reversible actions can be autonomous. A booked appointment can be moved or cancelled. A wrong one is a nuisance, and it shows up on the schedule for staff to see
  • Clinical decisions are never autonomous. Refill approval, triage and medical advice stay with clinicians, however capable the model seems
  • Emergencies end the task. A caller describing emergency symptoms is told to call 911, and the agent does not try to finish the booking
  • Rules may raise urgency and never lower it. A rule that downgrades a call because of what the caller said has turned into clinical judgment
  • Every action is logged where staff can see it, with a transcript and a summary
  • Permissions are the minimum the task needs. An agent that books appointments has no reason to edit medication lists

Physicians want the same guardrails. In the AMA's 2026 survey of nearly 1,700 physicians, 88 percent said robust safety and efficacy validation is critical to adopting AI, and 85 percent wanted to be consulted on or directly involved in adoption decisions. If a vendor's plan does not include your clinicians deciding the stop conditions, it is the wrong plan.

How do you tell real agentic AI in healthcare from agent washing?

Ask these, and ask for the answers on a live system rather than a slide.

  1. What exactly can it write, in which system, today? A list of actions, not a list of capabilities
  2. Show me the record it changed. After a demo booking, open the EHR and find the appointment
  3. What does it do when the system says no? A slot taken a second earlier, a provider template that blocks the visit type
  4. What will it refuse to do, even if the patient insists?
  5. Where is the log of every action, and who reviews it?
  6. Is it live in production at a practice on my EHR, and can I speak to them?

A product that answers the first question with "it understands patient intent" is a conversational system. Fine for some jobs, but not an agent.

What MedPhone does as an agent, and where it stops

MedPhone is an agent in the narrow sense this post argues for. On a phone call it identifies the patient against the EHR, books, reschedules and cancels appointments against real availability, takes new patient intake and insurance details, captures refill requests as tasks in the provider's clinical inbox, sends text confirmations, and transfers to staff by rules the practice sets. athenahealth is the full integration, verified Marketplace Partner. eClinicalWorks and NextGen each cover core scheduling workflows today, with further capabilities in active development.

At Synergy Medical, a family medicine practice in Michigan on athenahealth, MedPhone handled 2,848 inbound calls between June and mid August 2026. About 53 percent were finished without staff on day one, and that rose as call flows were tuned. It is one practice, and the only production deployment I can name.

Where it stops: it does not approve refills, give clinical advice or triage symptoms. It does not verify insurance eligibility or handle prior authorization. It does not make outbound calls, and it works in English only. If the task you want an agent for is one of those, MedPhone is not it.

See what MedPhone does on a call

Questions people ask

What is agentic AI in healthcare?

AI that is given a goal and completes it by taking actions in connected systems, such as booking an appointment in the EHR or filing a refill request as a task for a clinician. It differs from a chatbot, which answers questions but leaves the work for someone else.

What is the difference between agentic AI and generative AI?

Generative AI produces content, such as a drafted message or a note. Agentic AI uses tools to change something in a real system, such as the schedule. Many agents use generative models to understand and respond, but the defining feature is the action, not the text.

What are examples of agentic AI in healthcare?

An AI phone agent that books, reschedules and cancels appointments in the EHR during the call, takes new patient intake and routes refill requests to the provider's inbox. Earlier stage examples include tools that assemble prior authorization documentation and sort claim denials for billing staff.

Is agentic AI safe in healthcare?

It can be, when its permissions are narrow, its actions are reversible and logged, and clinical decisions stay with clinicians. It is not safe when an agent is allowed to judge symptoms, approve medications or act in systems it does not need.

Can agentic AI do prior authorization?

Partly. Agents can gather documentation and track status, but completing a request depends on payers. CMS requires impacted payers to implement a prior authorization API beginning January 1, 2027, and much of the work stays manual until they do.

Will agentic AI replace front desk staff?

No. It takes over routine, rule bound tasks like rescheduling, which gives staff time back for patients in the room and for calls that need judgment. In every deployment I have been involved in, the practice kept its staff and moved them off the phone.

Written by

Agni Patel is the founder and CEO of MedPhone, an AI phone agent for medical practices. He spends much of his time deciding what the agent is not allowed to do.

Want this run on your practice?

Bring fifteen minutes and your call reports. We will use your call volume and your payer mix rather than the benchmarks in this post.

Book a 15 minute demo