AI for a small medical practice: where it fits, and what it must not decide
In a small medical practice the hours go to notes, phones, intake, and the paperwork between the chart and the payer. AI fits in all four, at prices that start under $50 a clinician a month for the notes. It does not fit anywhere a clinician decides, and the rules about patient data come before any tool.
AI fits a small medical practice in four places: ambient scribes that draft the note during the visit, phone and intake tools that book and route, the paperwork between the chart and the payer, and the routine messages. Scribe plans for independent practices run $39 to $120 a clinician a month, and write-back to the EHR starts around $120 (Commure). It must not decide anything clinical: diagnosis, triage severity, medication, or coverage. The system lays the inputs out; the clinician decides. A signed Business Associate Agreement comes before any of it.
Where the hours go
Ask a physician in a three-provider practice where the day went and the answer is not patients. It is the note finished at nine at night, the front desk on hold with a payer, the intake form retyped from a clipboard, and the portal messages that arrive faster than anyone can answer them. Intuit's 2026 AI Impact Report found adoption highest in routine work and lowest where human judgment is essential, and that is the right map for a practice: the routine work is where the hours are, and the judgment is the part nobody should be trying to automate.
1. The note: ambient scribes
An ambient scribe listens to the visit and drafts the note, and it is the one category with real outcome evidence. A study published in JAMA Network Open in October 2025 enrolled clinicians across six health systems and included 263 who completed surveys before and after; burnout fell from 51.9 percent to 38.8 percent after 30 days with an ambient scribe (PubMed record, UChicago Medicine summary). That is a measured result from a controlled rollout, and it is the number to hold vendor claims against.
| Tier | Per clinician per month | What it includes | Examples named |
|---|---|---|---|
| Free | $0, with volume limits | Transcription, copy and paste into the EHR | Doximity Scribe, Heidi Health, Nabla, Tali |
| Independent practice | $39 to $120 | Ambient transcription and note generation, some with coding | Freed, Commure Scribe, Tali |
| Mid-market | $120 to $300 | One-click write-back into the patient record, team management | Heidi Health, Commure |
| Enterprise | $400 to $700 and up | Custom implementation, IT integration, a sales process | DAX Copilot, Suki, Abridge |
Tiers and examples from Commure's April 24, 2026 pricing guide. Commure is a vendor in its own comparison. The same guide puts a human scribe at $32,000 to $42,000 a year per provider.
Three checks. Does the vendor sign a Business Associate Agreement? Does the note write back into your EHR, or does someone paste it? And what does it do when it is unsure: flag the gap, or fill it? A scribe that invents a plausible sentence where it did not hear one is a liability with a subscription.
2. The phones and intake
The second place is the front desk. An AI receptionist answers, books into the schedule, and takes messages; an intake tool reads the form into demographics instead of someone retyping it. The value is in the integration with the practice management system, and the hand-off list is the safety. Anything with a symptom in it goes to a person. Anything that sounds urgent goes to a person immediately. The tool can book a physical; it cannot decide whether chest pain waits until Thursday.
We wrote a buying guide for AI receptionists in dental offices, and every check in it applies to a medical practice with the hand-off bar set higher.
3. The paperwork between the chart and the payer
Eligibility checks, prior authorizations, claim scrubbing, denials worked by hand. This is where a small practice loses money quietly, because the work is specific to each payer and each practice and no product covers all of it. It is also where the gap between systems lives: the chart on one side, the clearinghouse on the other, and a person in the middle copying fields. That gap is what custom work is for, with the same rules as everything else: minimum data, a stop on uncertainty, and a person on every denial.
What it must never decide
- Diagnosis. The scribe drafts what was said. It does not conclude.
- Triage severity. The receptionist books and routes. It does not rank urgency.
- Medication. Nothing that touches a prescription is automated end to end.
- Coverage and denials. The system prepares the appeal. A person decides to send it and what it says.
The pattern is the same in every row. The system gathers, drafts, and lays out. A clinician decides. A tool that blurs that line is not saving time; it is moving responsibility somewhere nobody can see it.
Find the edge where the hours go.
The ten questions ask about a normal week: what gets retyped, what gets rebuilt, what crosses systems. For a practice the answer is usually the note, the phones, or the payer, and the result says which.
The rules come first
None of the above is allowed to touch patient data until the vendor has signed a Business Associate Agreement, and the consumer tools do not sign one. HIPAA Journal is explicit that OpenAI will not enter into a BAA for the Free, Plus, Team, or Enterprise versions of ChatGPT, and that the healthcare product introduced in January 2026 is aimed at hospitals and regulated environments. Anthropic's BAA covers its first-party API and HIPAA-activated Enterprise plans, and not the Free, Pro, Max, or Team plans. The cloud platforms have their own BAA programs. The rule for a small practice is short: no BAA, no patient data, including the note a nurse pastes into a chatbot to tidy it up. Our four HIPAA checks are the one-page version.
What we built for a clinic
Our patient presentation automation was built for an orthodontic clinic whose staff assembled the same patient presentations by hand from intake PDFs and clinical photos. The system runs locally on the clinic workstation, reads the clinic's own template, resolves the fields and photo views into the right slots, and produces a deterministic presentation with an audit. Protected health information never leaves the workstation, working files are wiped after delivery, and every missing input stays visible as a warning instead of being filled in. The operator confirms patient identity and decides whether the deck is ready. The record makes no claim about labor saved or clinic adoption, because we do not have a measured number to make it with.
That build is the pattern for a medical practice in one sentence: the repeatable assembly becomes a pipeline, and judgment stays visible.
How to start
- Name the edge. Notes, phones, intake, or payer paperwork. Where do the after-hours hours go?
- Write the four rules before the first demo: BAA, minimum data, where the data lives, what the tool may not decide.
- Buy where a product exists. Scribes and receptionists have mature products; test them against the rules and the hand-off list.
- Build the gap. The chart-to-payer work is custom, and it is scoped after someone has watched a week of it.
A practice does not need to become a technology company. It needs the note written by ten, the phone answered, and the clinician deciding everything a clinician should.
What to hold on to
- Four places, one rule. Notes, phones, intake, and payer paperwork. The system drafts and lays out; the clinician decides.
- Scribes have real evidence. Burnout fell from 51.9 to 38.8 percent in 30 days across 263 clinicians in a JAMA Network Open study.
- Independent-practice scribe plans start under fifty a month. Write-back into the EHR starts around the mid-market tier. Free tiers paste.
- No BAA, no patient data. Consumer chatbots do not sign one. That includes the note someone pastes in to tidy it.
- Our clinic build ran local. PHI stayed on the workstation, gaps stayed visible, the operator decided. No outcome is claimed.
Questions owners ask us
How can a small medical practice use AI?
In four places: an ambient scribe that drafts the note during the visit, phone and intake tools that book and route, automation of the paperwork between the chart and the payer, and routine patient messages. Every tool needs a signed Business Associate Agreement, minimum data access, and a hand-off to a person for anything clinical.
How much does an AI medical scribe cost for a small practice?
Free tiers exist with volume limits and paste-only output. Independent-practice plans run $39 to $120 a clinician a month. Plans with one-click write-back into the EHR run $120 to $300. Enterprise platforms run $400 to $700 and up and require a sales process. A human scribe costs $32,000 to $42,000 a year per provider (Commure).
Do AI scribes actually reduce physician burnout?
A study published in JAMA Network Open in October 2025 included 263 clinicians across six health systems and found burnout fell from 51.9 percent to 38.8 percent after 30 days with an ambient scribe (PubMed). That is the measured benchmark to hold vendor claims against.
What should AI never decide in a medical practice?
Diagnosis, triage severity, medication, and coverage or denial decisions. The system can draft the note, book the appointment, and prepare the appeal. A clinician or a named staff member decides. Any tool that blurs that line moves responsibility where nobody can see it.
Can staff use ChatGPT for patient notes?
Not the consumer versions. OpenAI does not sign a Business Associate Agreement for the Free, Plus, Team, or Enterprise versions of ChatGPT (HIPAA Journal), so pasting a note that contains patient information into them is an impermissible disclosure. A practice needs a tool under a signed BAA, and a written policy that says so.
What has Prometheus built for a clinic?
A local-first patient presentation system for an orthodontic clinic. It assembles presentations from intake PDFs and clinical photos on the clinic workstation, keeps missing inputs visible as warnings, wipes working files after delivery, and leaves identity confirmation and the final call with the operator. The record makes no claim about labor saved.
Sources
Prices are quoted from a published 2026 guide, the burnout figures from a peer-reviewed study, the BAA facts from the vendors and HIPAA Journal, and the build from our record. We have no relationship with any vendor named. Checked September 5, 2026.
- Commure, "AI Medical Scribe Pricing 2026". April 24, 2026. Tiers, examples, and the human-scribe comparison. Commure is a vendor.
- "Use of Ambient AI Scribes to Reduce Administrative Burden and Professional Burnout," JAMA Network Open, October 2025 (PubMed record), with the UChicago Medicine summary. 263 clinicians, six health systems, burnout 51.9 to 38.8 percent at 30 days.
- HIPAA Journal, "Is ChatGPT HIPAA Compliant?". January 13, 2026. OpenAI's BAA position by product tier.
- Anthropic, "Business Associate Agreements (BAA) for Commercial Customers". Which Claude products the BAA covers.
- Intuit, "2026 AI Impact Report". May 12, 2026. Adoption highest in routine work, lowest where judgment is essential.
- Patient presentation automation, our implementation record.