AI for medical practices in 2026 is useful in a few specific places and oversold in most others. The parts that work today are narrow and administrative: answering and booking phone calls, drafting visit notes for a clinician to sign, and checking claims before they go out. Diagnosis, symptom triage and anything that stands in for clinical judgment are where caution still belongs.
AI for medical practices means software that uses machine learning or large language models to take over or speed up a specific task in an outpatient practice, such as answering calls, drafting notes or sorting claim denials. The useful question is never whether a practice uses AI. It is which task, who checks the output, and what happens when it is wrong.
Adoption is no longer the question. An AMA survey of nearly 1,700 physicians released in March 2026 found 81 percent use AI professionally, more than double the 2023 figure. The most common uses were summarizing research, drafting care plans and progress notes, and documenting charts and billing codes. Seven in ten respondents saw AI as a way to automate tasks that contribute to burnout.
Where AI for medical practices actually works in 2026
Here is my read on each area, from the most mature to the least. Maturity here means a typical independent practice can buy it, run it with ordinary staff and get the promised result, not that a large health system has a pilot.
| Area | What the AI does | Maturity in 2026 | Who checks it |
|---|---|---|---|
| Phones and front desk | Answers, books, reschedules, routes refill requests | Ready for routine admin calls | Staff, through transfers and call review |
| Clinical documentation | Ambient scribe drafts the visit note | Ready, widely adopted | Clinician, before signing |
| Revenue cycle | Claim checks, coding suggestions, denial sorting | Rules mature, generative tools newer | Billing staff |
| Patient portal inbox | Drafts replies to patient messages | Early, drafts only | Clinician edits and sends |
| Prior authorization | Gathers documentation, tracks status | Early, waiting on payers | Staff and clinician |
| Clinical decision support | Imaging reads, risk flags | Specialty specific, often FDA regulated | Clinician, always |
| Symptom triage | Suggests urgency from symptoms | Not ready to run unsupervised | Nurse or physician |
AI receptionists and the front desk
This is the most mature use for most practices because the work is high volume, repetitive and well defined. A large share of calls are booking, rescheduling, cancelling, refill requests and questions about hours or parking. An AI phone agent connected to the EHR can complete those on the call instead of taking a message.
MedPhone answers on the first ring around the clock, identifies the patient against the EHR, books against real provider availability, takes refill requests into the provider inbox as a task for a clinician to approve, writes a call summary to the chart, and transfers clinical, urgent or unusual calls under rules the practice sets. At Synergy Medical, a family medicine practice in Michigan on athenahealth, 65 to 75 percent of calls are now handled without reaching staff, and we estimate the front desk got 10 to 15 hours a week back.
The limits are real. It depends on the EHR connection, it does not give clinical advice, and it does not do insurance eligibility checks or prior authorization.
Clinical documentation and ambient scribes
Ambient scribes listen to the visit and draft the note. The best published evidence comes from The Permanente Medical Group: in the first ten weeks, 3,442 physicians used one across 303,266 encounters, and most users spent about an hour less a day at the keyboard. The authors also said the output still needs review for accuracy, and that is the part to take seriously. The clinician signs the note, so the clinician owns every word.
The problem they address is large. Primary care physicians in one Annals of Family Medicine study spent 5.9 hours a day in the EHR, 1.4 of them after clinic. If that describes your clinicians, a scribe may be a better first purchase than anything on the phone.
Revenue cycle: claims, coding and denials
Claim scrubbing, checking a claim against payer rules before it goes out, has been automated with rules for years and works. Newer tools suggest codes from the visit note and sort denials by likely cause. These are useful in the hands of billing staff who can tell a good suggestion from a bad one, and risky as a way to code without them. An unreviewed coding suggestion is an audit finding waiting for a date.
Insurance eligibility verification sits here too, in clearinghouse and practice management tools, not in the phone agent. A phone agent can capture insurance at intake and answer which plans you accept from practice settings; whether a plan covers today's visit is a separate system.
The patient portal inbox
Drafting replies to portal messages is one of the faster growing uses; 19 percent of physicians in the AMA survey reported it. It is early. A draft can be wrong in fact or tone, and patients read portal replies as coming from their clinician. Treat it as a first draft that a clinician edits and sends, never as an autoresponder.
Inbox work was already about 85 minutes a day for the physicians in the Annals of Family Medicine study. Refill requests that arrive as structured tasks, with the patient, medication and pharmacy already identified, do not shrink the approval work, but they cut the back and forth around it.
Why is prior authorization still mostly manual?
Because the bottleneck sits with payers. In an AMA survey of 1,000 physicians in late 2024, practices completed 39 prior authorization requests per physician per week, physicians and their staff spent 13 hours a week on them, and 40 percent of physicians had staff working only on prior authorizations.
AI tools on the practice side can gather the chart documentation a request needs and track status. The bigger change is regulatory. The CMS Interoperability and Prior Authorization Final Rule requires impacted payers to decide urgent requests within 72 hours and standard ones within seven calendar days, with most operational provisions starting in January 2026 and the prior authorization API generally due by January 1, 2027. Until payers are on the other end of that API, much of the work stays on the fax and the portal.
Clinical AI: decision support, diagnosis and triage
Clinical AI exists and some of it is good, but it is specialty specific and much of it is regulated as a medical device. The FDA keeps a public list of AI enabled devices it has authorized, and about three quarters of the entries are radiology tools. If you are a radiology or cardiology group, that list is worth reading. If you are a family practice, almost none of it is a front office decision.
Symptom triage is where I would be most careful. A 2015 BMJ audit of 23 online symptom checkers found they gave appropriate triage advice in 57 percent of standardized cases. Tools have improved since, but a front desk product that decides how urgent a symptom is has taken on clinical judgment. Urgent calls belong with a nurse line or the on call clinician. The most an AI phone agent should do with a symptom is hand it to one of them faster.
What AI should a small or private practice try first?
One administrative problem, measured before and after. Not a strategy.
- Pick the number that hurts most: abandoned calls, after hours charting, days in accounts receivable, or the refill backlog
- Measure it for two weeks so you have a baseline that is not a feeling
- Choose the one tool aimed at that number. For most small practices that is the phones or documentation
- Pilot it on a slice of the work, and review real output weekly for the first month
- Decide on the evidence, then move to the next problem
How to evaluate any AI tool for a practice
The same six questions work for a scribe, a phone agent or a billing tool.
- Will you sign a business associate agreement, and is patient data ever used to train your models?
- Who reviews the output before it reaches a patient or the chart, and how does an error surface?
- What exactly can you read and write in our EHR, and is it live on our EHR today or in development?
- How is it priced: per clinician, per seat, per call or per minute, and what does a heavy month cost?
- Can we speak to a practice like ours that has run it for at least three months?
- What happens to our data if we leave?
My own answers to the first and fourth: MedPhone signs a BAA, encrypts PHI at rest with AES 256 and in transit with TLS 1.3, never uses patient data to train models and audit logs every action. It is priced per practice on call minutes, from $150 a month, with no setup fee or contract. On the third, it is live in production on athenahealth; on other EHRs, ask what is live.
See how MedPhone handles a call
Sources
- AMA, More than 80 percent of physicians use AI professionally, March 2026
- Tierney et al., Ambient Artificial Intelligence Scribes to Alleviate the Burden of Clinical Documentation, NEJM Catalyst, 2024
- Arndt et al., Tethered to the EHR, Annals of Family Medicine, 2017
- AMA, Fixing prior auth: nearly 40 prior authorizations a week is way too many
- CMS, Interoperability and Prior Authorization Final Rule, January 2024
- FDA, List of Artificial Intelligence Enabled Medical Devices
- Semigran et al., Evaluation of symptom checkers for self diagnosis and triage, BMJ, 2015
Questions people ask
How is AI used in medical practices today?
Mostly for administrative work: answering and booking phone calls, drafting visit notes, checking claims and drafting replies to patient messages. Clinical uses exist, mainly in imaging, but in a typical outpatient practice the value in 2026 is in reducing front office and documentation work.
Can AI be used in healthcare safely?
Yes, when it handles a defined task, a person checks the output, and the vendor signs a business associate agreement. The risk rises sharply when AI is asked to make clinical judgments, such as deciding how urgent a symptom is, without a clinician involved.
What are examples of AI agents in healthcare?
An AI phone agent that books appointments and routes refill requests, an ambient scribe that drafts the visit note, and tools that sort claim denials or gather documentation for prior authorization. Each does one job and hands anything outside it to a person.
What AI should a small medical practice use first?
The tool aimed at your worst measured problem. For most small practices that is either the phones, if calls go unanswered, or documentation, if clinicians chart late into the evening. Measure for two weeks first, then pilot one tool.
Is AI for medical practices HIPAA compliant?
Some of it is, and the label alone proves nothing. Ask for a signed business associate agreement, encryption at rest and in transit, audit logs, and a written answer on whether patient data trains the vendor's models.
Will AI replace doctors or front desk staff?
No. AI takes over specific tasks, such as routine calls or the first draft of a note, and leaves judgment, examination and in person work with people. Practices that use it well usually redeploy staff time rather than cut roles.
Written by
Agni Patel is the founder and CEO of MedPhone, which builds an AI phone agent for medical practices. He works on one row of the table below and has tried to mark the others as fairly as his own.