Most of what a language model can touch in your day is the work that stretches past the last patient: the note, the in-basket, the prior authorization appeal, the letter of medical necessity, the portal message that needs a plain-language reply. The judgment in the exam room is yours. The typing does not have to be.
Two lines do not move. AI does not make diagnostic, treatment, or dosing decisions, and protected health information never goes into a consumer chatbot. A general assistant has no Business Associate Agreement, its medical knowledge stops at a training cutoff, and it will invent a guideline or a citation with the same fluent confidence it uses for a real one.
So the working split looks like this. Ambient scribes and EHR-integrated drafting tools your organization has licensed handle anything with patient data in it. General assistants handle everything else: the appeal built from the payer's published policy, the fictional teaching case, the recertification study plan, the committee memo. And every output gets the read you would give a resident's note before you cosign it.
Quick wins this week
- Paste a payer's published medical policy and a de-identified summary of the clinical picture, and ask for a prior authorization appeal that quotes the policy's own criteria back to the reviewer.
- Turn a dense guideline update into a one-page briefing for your group, then check every line against the source document before you circulate it.
- Ask for a plain-language explanation of a common procedure at a sixth-grade reading level for your after-visit instruction templates, and verify each clinical statement against your organization's approved materials.
- Draft the peer-review response, the committee memo, or the recommendation letter you have been avoiding, described without any patient identifiers.
- Build a study plan for your board's recertification exam or longitudinal assessment from the board's published content outline.
What AI can do for physicians, task by task
Ambient documentation and note drafting inside the EHR
If your organization has licensed an ambient scribe (Abridge, Ambience, Nabla, Suki, Microsoft's Dragon Copilot, or your EHR vendor's own feature), learn three things: the consent script, where the draft lands, and how to correct it before it becomes the note. Read the assessment and plan word by word before signing. Drafts misattribute who said what, drop pertinent negatives, inflate the history, and occasionally document an exam element that never happened. Under the Cures Act open-notes rules, the patient will read what you sign.
In-basket and patient portal message replies
Several EHRs can now draft replies to portal messages if your organization has turned the feature on. Treat the draft as a starting point written by someone who has not read the chart carefully: check that it answers the actual question, does not offer guidance you would not give, and does not promise a refill, a referral, or a result you have not reviewed. Some states, California among them, require a disclosure when generative AI writes clinical communications to patients that a licensed clinician has not reviewed, so review is the rule, not the exception.
Prior authorization appeals and letters of medical necessity
Give the assistant the payer's published policy, the criteria you believe are met, and a fully de-identified summary of the clinical picture, then ask for a letter that maps each criterion to the documentation. Verify every guideline or study the letter cites by opening it; fabricated references are common, and a reviewer who catches one discounts the whole letter. If the letter needs real patient detail, write that part inside an approved tool, not a consumer chatbot.
Evidence lookups and literature review
Use clinician-focused evidence tools that cite their sources (OpenEvidence is one; UpToDate and other references have added AI layers, so check what your institution licenses), or use Perplexity to find candidate papers, then read the actual paper. Models overstate effect sizes, blur the study population, and quote conclusions the abstract never made. For a stack of articles you already have, NotebookLM answers only from what you upload and shows the passage, which is the behavior you want.
Patient instructions and education at a readable level
Describe the procedure or condition in general terms, with no identifiers, and ask for instructions at a specific reading level with a 'call us if' section and one teach-back question. Check each clinical statement against your organization's approved education content and your own practice; models restate outdated instructions with total confidence. Keep medication guidance general and leave doses to the prescription and the pharmacist's counseling.
Documentation self-audits against E/M levels
Since 2021, office visit levels rest on medical decision making or total time, and the same framework now covers most other settings. Paste a fictional or fully de-identified note and ask the model to list the problems addressed, the data reviewed or ordered, and the risk elements it can find, quoting the note for each, without assigning a level. Then level it yourself. Never let a tool suggest adding something that was not done; the goal is accurate documentation of real work, not a higher code.
Teaching, board recertification, and CME
Ask for fictional cases at a stated learner level with branching questions, or for a study plan built from your board's published blueprint and your weak areas. Have the model mark any rationale it is not certain about, then check those against a current reference. For longitudinal assessment programs, the question bank from the board and your specialty society remains the source of truth for what is tested.
Committee work, quality projects, and the writing nobody has time for
De-identified M&M summaries, PDSA write-ups, policy drafts, abstract edits, and the slide deck for grand rounds are all fair game. Copilot inside your organization's Microsoft tenant is usually the sanctioned place for this if IT has enabled it, because your documents stay inside the tenant. Gamma turns an outline into a first-draft deck in minutes; the content still needs your review for accuracy and attribution.
Prompts for physicians
Replace the bracketed placeholders, paste into any chat assistant, and iterate on the result.
Prior authorization appeal that quotes the payer's own policy
Act as a physician advisor who writes prior authorization appeals. Request: [DRUG, TEST, OR PROCEDURE]. Payer policy: [PASTE THE PAYER'S PUBLISHED MEDICAL POLICY OR CRITERIA]. De-identified clinical summary (no names, dates, or identifiers): [DE-IDENTIFIED CLINICAL SUMMARY]. Prior treatments and outcomes: [WHAT HAS ALREADY BEEN TRIED]. Write an appeal letter that takes each criterion in the policy in order, quotes it, and then states the facts from the summary that satisfy it. Do not cite any guideline, study, or source beyond the policy text and the summary I provided. If a criterion is not clearly met by the summary, say so in a separate note to me rather than papering over it. Under 500 words, formal, no adjectives about the payer.
Tip: The 'say so in a separate note' line is where the value is. A model that fills gaps will get you a second denial.
Portal message reply for a common non-urgent question
You draft patient portal replies for a [SPECIALTY] practice. The patient's question, with all identifying detail removed: [FICTIONAL OR DE-IDENTIFIED QUESTION]. Our standard guidance on this topic: [YOUR STANDARD GUIDANCE OR PROTOCOL TEXT]. Write a reply under 120 words at an eighth-grade reading level. Use only the guidance I provided; do not add clinical advice, doses, or reassurance beyond it. Include one line on what should prompt a call or a visit. End with a plain sentence that the message was reviewed by the clinician. Then list any part of the question the guidance does not cover so I can answer it myself.
Tip: Keep the real reply inside the EHR's own draft feature if your organization has it. Use this version to build your standard-guidance library.
Guideline update briefing for the group
Act as a medical editor. Below is the text of a guideline update relevant to [SPECIALTY]: [PASTE THE GUIDELINE TEXT OR EXECUTIVE SUMMARY]. Write a one-page briefing for practicing clinicians with these sections: what changed from the prior version, what did not change, what a clinician would do differently on Monday, and open questions the guideline itself acknowledges. Quote the guideline for every recommendation you summarize, with the section number. Do not add recommendations from your general knowledge. Finish with a list of every statement I should verify against the full document before sharing.
Tip: Paste the actual text. Asking a model what a guideline says from memory gets you the previous version, or a blend of several.
Fictional teaching case for residents
You are an attending physician building a teaching case. Create a fictional patient (no real data) with [PRESENTATION OR DIAGNOSIS] for learners at the [LEARNER LEVEL] level in [SPECIALTY OR SETTING]. Structure it as a three-stage unfolding case: initial presentation with vitals and history, then results, then a turn in the course. After each stage, give three questions that test clinical reasoning rather than recall, with model answers. Mark any clinical statement you are not certain is current with the tag [VERIFY] so I can check it against a reference before I use the case.
Tip: Search the output for the VERIFY tag before you teach it. Then read what it did not mark, because that is where the wrong-but-confident lines hide.
Board recertification study plan
Act as a study coach for the [BOARD OR ASSESSMENT NAME]. Official content outline: [PASTE THE BLUEPRINT OR CONTENT PERCENTAGES]. My weakest areas: [WEAK AREAS]. Time available: [HOURS PER WEEK] hours a week until [EXAM OR WINDOW DATE]. I see patients [CLINICAL SCHEDULE]. Build a week-by-week plan that weights time by the blueprint and my weak areas, uses spaced review, and includes a short self-check each week. Then write five practice questions on my weakest domain with rationales, marking any rationale you are not fully certain about so I can verify it in a current reference.
Tip: Question banks from your board or specialty society stay the source of truth. The model is for scheduling and warm-up, not for deciding what is true.
Letter of medical necessity for equipment or therapy
Act as a physician drafting a letter of medical necessity for [DEVICE, SERVICE, OR THERAPY]. Coverage criteria, pasted from the payer or the relevant Medicare coverage determination: [PASTE THE COVERAGE CRITERIA]. De-identified clinical facts: [FUNCTIONAL STATUS, DIAGNOSIS IN GENERAL TERMS, PRIOR ALTERNATIVES TRIED]. Write the letter so that each coverage criterion is addressed in its own short paragraph with the supporting fact. Do not invent measurements, dates, or test results; leave a bracket where I need to insert a real value inside my approved system. Under 400 words, plain clinical language, no sources beyond what I pasted.
Tip: The brackets it leaves behind are your checklist. Fill them in the EHR or an approved tool, never by pasting chart data back into the chatbot.
Want a prompt for something else? Use the Prompt Builder.
Skills to build
Treating the model as a fluent stranger, not a colleague
Why: It writes like the sharpest fellow you ever worked with and knows less than an intern about what happened last month. Fluency is not evidence, and confident wrongness is the failure mode that reaches patients.
How: For every clinical claim in an output, ask where it came from. If the answer is not a source you can open, it goes through your reference, the primary literature, or a colleague before it changes anything.
Knowing the PHI line and the BAA line
Why: HIPAA's Safe Harbor list of 18 identifiers includes dates of service, ages over 89, device serial numbers, and full-face photos, and a rare diagnosis plus a small town can re-identify a 'de-identified' case.
How: Keep the identifier list next to your workstation. Before any prompt, rewrite the situation as a fictional case, and ask your compliance office which tools carry a BAA and what each is approved for.
Reviewing generated notes like a cosigner
Why: Ambient drafts fail by omission and by plausible invention: the dropped negative, the exam element that was not performed, the medication the patient mentioned but does not take. Your signature makes it your record and the basis of the claim.
How: Read the assessment and plan first, then the exam, then the history, against your memory of the visit. Correct the draft in the tool so it learns your style, and never batch-sign.
Prompting with the payer's, the board's, or the guideline's own text
Why: A model reasoning from memory drifts toward generic medicine. A model reasoning from the actual policy, blueprint, or guideline text quotes the right criterion and stays inside it.
How: Build a folder of source documents you reuse: payer policies, coverage determinations, your board's blueprint, your organization's patient-education standards. Paste the relevant one at the top of every prompt.
Understanding what your board, the FDA, and your carrier expect
Why: The FDA has published guidance on when clinical decision support software counts as a regulated device, the AMA's policy uses the term augmented intelligence on purpose, state medical boards are issuing their own guidance, and some states now require disclosure of AI-written patient communications.
How: Read your state board's current statement on AI, ask your malpractice carrier what documentation it wants when AI assists, and follow your organization's consent script for ambient recording, which matters most in all-party consent states.
Tools worth knowing
Microsoft Copilot
AI inside Word, Excel, Outlook, and Teams, with your company's data protections.
Claude
A careful writing and analysis assistant that shines on long documents.
ChatGPT
The general-purpose AI assistant most of your coworkers already use.
Perplexity
An answer engine that cites its sources, built for research rather than chat.
NotebookLM
A research notebook that only answers from the sources you give it, with citations.
Cautions for physicians
Consumer chatbots carry no Business Associate Agreement, and a prompt containing any of the 18 HIPAA identifiers is a disclosure of PHI. Never paste confidential patient data into a consumer AI tool unless your organization has approved that tool and signed the agreements. Ambient scribes, EHR-integrated drafting, and enterprise chat under a BAA are the places for anything with a patient in it; everything else gets a fictional or fully de-identified version.
A language model can list possibilities, but it has not examined the patient, does not know today's labeling, makes arithmetic errors, and states outdated guidelines with confidence. If you use one to check your own thinking on a fictional or de-identified case, every idea it surfaces gets verified in a source you trust before it touches a plan. Dosing comes from the reference and the pharmacist, and the decision is yours under your license.
An ambient draft that documents a review of systems that was not performed, or a plan the patient never heard, is a documentation error the moment you sign it and a billing problem the moment it supports a level of service. Under open notes the patient reads it too. Read every generated line against what happened, and know how your organization expects AI assistance to be disclosed in the record.
Models invent plausible references with real journal names and fake page numbers, and they blend guideline versions from different years. Open every citation before it goes into a letter, a lecture, or a manuscript. For anything time-sensitive, paste the current source text into the prompt instead of asking the model to recall it.
California requires a disclaimer on generative-AI patient communications a clinician has not reviewed, other states are following, and ambient recording needs consent under state recording laws. Bias is documented: a 2023 study in npj Digital Medicine by Omiye and colleagues found large language models repeating outdated race-based clinical assumptions. Read generated material with your actual patient population in mind, and treat your organization's AI policy as part of your scope of practice.
Your 30-day plan
- Week 1: Ask compliance or informatics which AI tools are approved, which carry a BAA, and what may go into each. Read the 18-identifier list. Use a general assistant for two non-patient tasks: a committee memo and a CME plan.
- Week 2: Build one prior authorization appeal from a payer's published policy and a de-identified summary, verify every citation, and compare it with your last hand-written appeal. Draft two patient-instruction templates and check them against approved materials.
- Week 3: If your organization has an ambient scribe or portal-reply drafting, complete the training, learn the consent script, and use it for a week while reading every draft before signing. Note the error types you catch.
- Week 4: Load your specialty's most-used guideline and two payer policies into NotebookLM and use it for a week of lookups. Write one fictional teaching case and use it with a learner.
- End of month: List the three tasks where AI saved real time, the one place it was confidently wrong, and what you now want IT or your group to pilot next.
Frequently asked questions
Will AI replace physicians?
Can I use ChatGPT to write my clinical notes?
Is it legal to use AI for prior authorization appeals?
Do I have to tell patients I used AI?
Which AI tools are HIPAA compliant?
Terms used on this page
Related roles
- AI for NursesAI can take the writing, summarizing, and studying load off your shift without ever touching a patient record, as long as you know where the HIPAA line is and stay on the right side of it.
- AI for PharmacistsPharmacy has a writing problem hiding inside a verification problem. AI can take the first draft of nearly everything you write, as long as the drug reference, the PDMP, and your own judgment stay in charge of every decision.
- AI for Healthcare AdministratorsAdministrators drown in documents: policies, payer correspondence, board decks, survey prep, budget narratives. AI drafts and summarizes all of it, if you keep PHI inside covered systems and set the rules for your organization before someone else sets them for you.
- AI for Medical Coders and BillersAI can read a note faster than you can, and it will also cheerfully assign a code that was deleted two years ago. Used well, it speeds up the reading, the denial letters, and the research; the code selection and the compliance stay yours.