How-To Guide

How to Use AI for Medical Practice in 2026: Charting, Coding, Patient Comms & Compliance

Q: Is it HIPAA-compliant to paste patient notes into ChatGPT?

Not by default. Consumer ChatGPT, Claude.ai, and Gemini are not covered by a Business Associate Agreement (BAA) and should never receive PHI. For any protected information, you must use a BAA-backed plan: OpenAI has a BAA path under ChatGPT Enterprise and the API with a signed BAA, Anthropic offers BAAs for eligible Claude for Work customers, Microsoft 365 Copilot runs under the enterprise BAA inside your tenant, and dedicated medical scribes (Abridge, DAX Copilot, Suki, Freed, Heidi, Augmedix) are BAA-covered by design.

Q: Will ambient AI scribes replace transcription?

They already have for most outpatient visits. Abridge, DAX Copilot, Suki, Freed, and Heidi generate SOAP notes in real time and are integrated with Epic, Cerner/Oracle Health, athena, and eClinicalWorks. Clinician burnout studies from 2025-2026 show 1.5-2.5 hours per day of charting time returned when an ambient scribe is used properly with physician review.

Q: Can AI do my E/M coding for me?

It can draft the code and explain the MDM level, but CMS still requires a human billing professional or provider to attest. The 2026 CMS guidance (MLN Matters MM13412 and later) specifically calls out that auto-generated codes must be reviewed. Treat AI as a coder's assistant, not the coder of record.

Q: What about AI for patient-facing communication?

Safe and valuable for pre-visit instructions, post-visit summaries at a fifth-grade reading level, portal replies drafted for provider review, and medication adherence nudges — always reviewed by a licensed clinician before sending. Never let AI auto-reply to clinical portal messages without human sign-off; that's the single biggest liability exposure.

Q: Do I need a Model Card or AI disclosure for my patients?

Increasingly yes. Several states (California AB 3030 as amended, New York S1331A, Texas SB815) now require patient disclosure when generative AI is used in clinical communication. A one-paragraph notice in your intake paperwork plus a line in the after-visit summary satisfies most current requirements; check your state medical board for specifics.

Published April 25, 2026 · 14 min read

TL;DR

Ambient scribes return 1.5–2.5 clinician hours per day; use one.
Never paste PHI into consumer chat. Only BAA-covered tools touch patient data.
Ten prompts cover chart prep, SOAP drafting, E/M coding assist, prior-auth letters, denial appeals, and patient-facing comms.
2026 state laws (CA AB 3030, NY S1331A, TX SB815) now require patient-facing AI disclosure in many clinical communications.
Minimum small-practice stack: ambient scribe (Abridge, Freed, Heidi, or Suki) + a BAA-covered writing assistant for non-PHI admin work.

Why 2026 is the year medical practices adopt AI

The 2025 AMA Physician Burnout Survey reported 53% of physicians citing documentation burden as their top burnout driver. In parallel, peer-reviewed studies from Kaiser Permanente, Stanford, and UCSD showed 25–40% reductions in after-hours EHR time ("pajama time") when ambient scribes were deployed properly. The technology is no longer experimental. Every major EHR — Epic, Oracle Health (Cerner), athena, eClinicalWorks, NextGen — now has native or first-party AI integration partners under BAA.

What AI is now genuinely good at in a medical practice: generating a first-draft SOAP note from an ambient recording, summarizing a chart before a visit, drafting prior-authorization letters, writing appeal letters for denied claims, translating discharge instructions into plain language and into Spanish or other languages, and composing thoughtful replies to patient-portal messages for clinician review. What it is still bad at: final clinical decisions, high-stakes differential diagnosis without human oversight, and any scenario that needs accurate temporal or dose-specific drug information (always double-check via Lexicomp or UpToDate).

The 2026 practice stack

Need	BAA-covered tools	Typical cost
Ambient scribe	Abridge, DAX Copilot, Suki, Freed, Heidi, Augmedix	$99–$299/provider/mo
Chart summary / prep	Epic ED Chart Summary (MyChart Cortex), Oracle Clinical AI, athena Intelligence	Bundled in EHR
Portal message drafting	Epic In Basket AI, Oracle Clinical AI Inbox	Bundled
Admin writing (non-PHI)	Happycapy Pro, Claude for Work (with BAA), ChatGPT Enterprise (with BAA)	$20–$60/user/mo
Prior auth + denial	Cohere Health, Rhyme, XpertCoding	Varies by payer volume

For the non-PHI admin writing layer, Happycapy Pro at $20/month handles the policy drafts, patient education handouts, marketing, and staff communications that do not touch patient data. Pair it with a BAA-covered ambient scribe for the clinical work.

10 prompts every clinician should keep in 2026

1. Pre-visit chart summary (BAA tool only)

You are my rooming nurse. Summarize the attached chart (problem list, med list, last 3 visit notes, most recent labs, imaging) into a 5-bullet pre-visit brief for my next patient: 1. Active issues most relevant to today's chief complaint 2. Open care gaps (overdue screens, immunizations, chronic care metrics) 3. Medication red flags (interactions, renal/hepatic dose issues, controlled substances) 4. Social determinants that may affect adherence 5. One suggested talking point for the first 30 seconds of the visit Chief complaint: [CC]. Max 120 words total.

2. SOAP note QA after ambient draft

The ambient scribe produced the attached SOAP draft. Before I sign it: 1. Flag anything in the HPI that isn't supported by the recorded transcript. 2. Check the MDM level for E/M 2021+ criteria (problem complexity, data reviewed, risk). 3. Identify any prescription changes mentioned in the A/P that aren't in the current med list — I need to reconcile before signing. 4. Suggest any ICD-10 codes that match the documented assessment but aren't yet captured. Do not add clinical findings that aren't in the transcript.

3. Patient-friendly after-visit summary

Rewrite the attached A/P at a 5th-grade reading level for the patient's portal: - Plain language, second person - What the patient should do today, this week, and before the next visit - Warning signs that mean they should call us or go to the ED - Any meds started/stopped/changed with the reason in one sentence each - Next appointment and what will be checked then Keep to 250 words. Add a Spanish translation after the English version.

4. Prior authorization letter

Draft a prior-authorization letter to [PAYER] for [MEDICATION/PROCEDURE]. Pull from the attached chart: - Diagnosis with ICD-10 - Prior therapies tried, duration, outcome - Reason requested therapy is medically necessary - Clinical guideline citation (HEDIS, specialty society, FDA label) - Risks of delay / non-treatment Format for fax submission. Keep to one page. Do not invent clinical history that isn't in the chart.

5. Denial appeal letter

Payer [X] denied claim [Y] for [REASON: medical necessity / experimental / non-covered]. Draft a first-level appeal citing: - The specific denial reason and my counter-argument - Relevant medical record evidence with date references - Applicable coverage policy or LCD - Clinical guideline support (include journal + year) - A closing ask with clear timeline expectations Tone: professional, firm, not emotional. One page max. Ready to send on our letterhead.

6. Portal reply draft (for clinician review)

A patient sent this portal message (attached). Draft a reply for my review: - Address the question directly - Set expectations on timeline (meds take X days to work, labs back in Y days) - List the red flags that mean call/ED - Close with a clear next step (no action needed / book visit / await labs) - Never commit to a diagnosis or prescription change; escalate to me Do not send. Mark as DRAFT at the top.

7. Refill review triage

I have 18 refill requests this morning. Triage them into 3 buckets using the attached list + chart snippets: - AUTO-OK: chronic stable med, last visit <12 months, labs within monitoring interval — auto-approve - OK-WITH-FOLLOWUP: stable but overdue visit or lab — approve this cycle, flag for nurse outreach - NEEDS MD REVIEW: controlled substance, dose change warranted, missed monitoring, or new red flag Output a table with patient initials, med, bucket, and one-line reason.

8. Policy or handout draft (non-PHI, Happycapy-safe)

Write a 1-page patient handout on [TOPIC, e.g., managing type 2 diabetes with diet]: - 5th-8th grade reading level - Plain language, friendly but clinical - What it is, why it matters, 5 things to do at home, when to call us - Include a "pantry swap" or similar concrete action box - Cite sources at the bottom (CDC, ADA, AHA) Do not include any patient-identifying info. This is a general handout.

9. Staff meeting agenda and summary

Draft an agenda for our 45-minute weekly clinical huddle given these topics: - No-show rate trend - Quality metric gaps (diabetic eye exams, colorectal screening) - New payer policy changes - Staffing for holiday schedule - Patient feedback themes from last 2 weeks Include time boxes, owners, and a "decisions needed today" list at the top. After the meeting, I'll paste the notes back and you'll produce the summary + action items.

10. Online review and reputation replies

Draft HIPAA-compliant public replies to the 6 attached Google / Healthgrades / Zocdoc reviews. Rules: - Never confirm or deny the reviewer is a patient - Never reference any specific visit, diagnosis, or care detail - For positives: warm, specific-to-the-general-theme (bedside manner, wait time), short - For negatives: empathetic, invite offline contact via office manager, do not relitigate - For clinical complaints: one line acknowledging, one line redirecting to clinical director Each reply under 60 words. Ready to post.

Compliance checkpoints you cannot skip

BAA on file. Any tool handling PHI must be BAA-covered. Keep a one-page vendor matrix with BAA effective date, data residency, and breach-notification terms.
Patient disclosure. Add a plain-language AI statement to your intake paperwork: "We use AI tools to help draft notes and communications; all content is reviewed by a licensed clinician before it affects your care."
State-specific rules. California AB 3030, New York S1331A, Texas SB815 — confirm your state's current language.
Bias audits. If you use AI for triage or risk scoring, document the audit cadence. ONC and the 2026 HTI-2 rule require transparency on decision-support outputs.
Medical-record integrity. CMS and state boards require that the clinician of record review and sign AI-generated content. Never let a scribe auto-sign.

Frequently asked questions

Is it HIPAA-compliant to paste patient notes into ChatGPT?

Not by default. Consumer ChatGPT, Claude.ai, and Gemini are not covered by a Business Associate Agreement (BAA) and should never receive PHI. For any protected information, you must use a BAA-backed plan: OpenAI has a BAA path under ChatGPT Enterprise and the API with a signed BAA, Anthropic offers BAAs for eligible Claude for Work customers, Microsoft 365 Copilot runs under the enterprise BAA inside your tenant, and dedicated medical scribes (Abridge, DAX Copilot, Suki, Freed, Heidi, Augmedix) are BAA-covered by design.

Will ambient AI scribes replace transcription?