Lesson objective
Describe a medical practice audience by role, by the AI-assisted output each role produces in an ordinary week, and by the named person who owns that output, then plan a session around the interruptions a working practice imposes.
Plain-language explanation
Why this matters now
Levels 1 to 5 taught you to use AI on medical practice work, to write the rules a practice runs on, to govern vendors and access, and to put automations behind gates a named person can actually operate. Level 6 asks something different. It asks you to stand in a break room with a whiteboard nobody has cleaned, teach other practice staff to do those things, then leave and never see their screens again.
An instructor who teaches a rule wrongly makes the error once for every learner. In a medical practice that error does not stay in the room. It reaches a chart, a claim, a portal message, or a patient sitting on an exam table.
The room is not one audience. A single half-hour session at a three-physician family practice can hold an owner physician who signs every note, a nurse practitioner who joined in the spring, a nurse who runs the triage line, two medical assistants who room patients and take vitals, a front desk lead who registers and schedules, a referral and prior authorisation coordinator, a biller, a certified coder who works two afternoons a week, and the practice manager who bought the software and is also the named privacy official. Those people do different work, are supervised differently, and hold different things.
They are also not starting from zero. Survey research (not binding): American Medical Association, Center for Digital Health and AI, 2026 Physician Survey on Augmented Intelligence, published March 2026 (day not stated). It was fielded 15 January to 2 February 2026 and 1,692 U.S. physicians participated, with full completes and qualified partial responses, and percentages calculated on question-level bases. On the adoption question, with a base of 1,342, 72% reported incorporating one or more AI use cases, 9% were uncertain which AI use cases their practice offers, and 19% reported none. Those three account for the whole base: 72 plus 9 plus 19 makes 100. The average number of use cases rose from 1.1 in 2023 to 2.3 in 2026. It is a self-reported survey of volunteer respondents with no margin of error or weighting stated, and the 2026 wave added use cases and counted partials, so AMA itself calls year-on-year comparisons approximate.
Read the 9% again, because it is the most useful number in the whole survey for an instructor. Nearly one in ten physicians on that base could not say which AI their own practice offers. Your room will contain people using something they cannot name, inside a workflow nobody wrote down.
A question to sit with
A medical assistant has been told by the practice manager to attend. She is eating at the table, she has three rooms to turn before the one o'clock, and she will stand up at 12:55 whether you have finished or not. What single thing must she carry out of the room for the half hour to have been worth her lunch?
What goes wrong
1. The definitions opening. Ten minutes on what a large language model is, delivered to a front desk lead who has been drafting appointment reminders with one since spring and to a physician who does not care. Both switch off, for opposite reasons.
2. One role's example for everybody. A claim-edit example serves a room holding a nurse, two medical assistants and a triage line, none of whom touch claims. They hear about someone else's job for half an hour.
3. The clinical line ignored. A demonstration shows a tool summarising a fictional chart and the instructor says "so you would tell the patient that". In the room sit a medical assistant and a front desk lead for whom relaying a clinical statement to a patient as a finding is precisely the thing they must never do. Half the room has just been shown work that is not theirs.
4. The clock treated as negotiable. The session is built for forty-five minutes and gets twenty-six, because the 11:40 patient needed a wound check. Everything after minute twenty was the part that mattered.
5. Ownership left vague. The session says "someone should review it" and never names which role owns which output. In a practice, ownership is specific: clinical content and the signature belong to the licensed treating clinician, the final code rests on the clinician's documentation and the coding professional's judgement, and administrative output belongs to a named administrative owner.
6. Attendance counted as success. The manager wants to hear that everyone signed in. What protects a patient is the check three weeks later at 4:40 on a Friday, when nobody is watching.
The practical answer
This programme holds no education or learning-science research. Nothing below is a claim about how adults learn, what they retain, how long they retain it, or whether anyone has a learning style. It is instructional practice, labelled as practice, built on how a practice day is actually organised, which everyone in the room can check against their own week.
Best practice: describe the room by role, by the output that role produces with AI help, and by the named person who owns that output. The ownership column is not a preference. It follows licensure, delegation and who is accountable for the record and the claim.
| Role in the room | AI-assisted output in an ordinary week | Who owns the output | The check that matters most | What loses them |
|---|---|---|---|---|
| Physician, NP or PA | Drafted progress note from ambient capture, chart summary, drafted portal reply | The licensed treating clinician, for everything clinical and for the signature | Their own reading of the record before signing, including where AI captured the transcription | Tool mechanics for work they will delegate anyway |
| Nurse or triage nurse | Drafted call-back note, drafted patient instruction text | The licensed clinician for clinical content; the nurse within their own licensed scope | That nothing leaves the practice stating a clinical finding no clinician has made | Billing examples and vendor tours |
| Medical assistant | Drafted rooming note, drafted intake summary, form pre-fill | The clinician for clinical content; the practice manager for administrative text | That a drafted line is never relayed to a patient as a clinical finding | Being shown work their role does not cover |
| Front desk and registration | Drafted reminder, drafted scheduling message, insurance field capture | The practice manager or named administrative owner | That the message matches the record, the consent and the channel, and that only the minimum necessary information travels | Being treated as an afterthought when they use the tool most |
| Referral and prior authorisation coordinator | Drafted request letter, packet assembly, status summaries | The named administrative owner, with clinical content from the clinician | That every clinical statement in the packet traces to the clinician's own documentation | A session with no packet in sight |
| Biller and coding professional | Flagged claim mismatches, drafted appeal text, documentation queries | The coding professional for code assignment, working from the clinician's documentation | That the record supports what is submitted, and that nothing goes out unreviewed | General prompting with no claim on the screen |
| Practice manager, privacy or security official | Policy text, vendor records, operational reports | The practice owner, with the privacy and security officials | That a written rule exists and somebody reviews against it | Detail with no decision attached |
Best practice: ask the room before you build it. A four-question form sent a week ahead can ask each person's role, which AI tool they use if any, what they use it for, and one time it produced something wrong. Ask for no patient identifier and no clinical detail. The answers describe that room on that day, and are never reported as a statistic about medicine.
Best practice: build for the clock you will actually get. Plan a thirty-minute slot as twenty-two minutes of content plus eight minutes of slippage, and put the thing that must survive in the first ten. Write it so that if you are cut off at minute eighteen, the behaviour has already been practised once by every role present.
Best practice: give each role something to do, not something to watch. An experienced user walks the room through the check they run. A non-user names the one condition that would have to be true before they would try it. A sceptic is asked what would have to go wrong for their worry to be justified.
Guidance (voluntary, not binding): HHS-OIG General Compliance Program Guidance, dated 11-06-2023. OIG states the GCPG "is voluntary guidance" and "is not binding on any individual or entity", and it names training and education among the seven elements of a compliance program. It also says the owner or CEO of a small entity is ultimately responsible for compliance with Federal health care program requirements, and names individual and small-group physician practices as small entities. The GCPG says nothing about artificial intelligence, so never attribute an AI-specific recommendation to it. What it gives an instructor is the honest frame: a practice that bills Federal health care programs already has a reason to train its workforce, and your session can be part of that, at the practice's choice.
Remember this
- The room holds clinical and administrative people doing different work under different supervision; design by role, by output and by named owner.
- AMA's 2026 figures come from 1,692 physicians fielded 15 January to 2 February 2026, with a base of 1,342 on the adoption question; quote no percentage without its base and date.
- 72% reported at least one use case, 9% were uncertain which their practice offers, 19% none; the 9% is the room you are walking into.
- An instructor gives no clinical advice, no triage, no coding determination and no payer interpretation from the front of the room.
- Plan thirty minutes as twenty-two plus slippage, and put the behaviour that matters in the first ten.
Medical practice example
Renata Villalobos, a fictional practice manager and training lead at the fictional Cedar Hollow Family Medicine, is given a thirty-minute lunch slot to introduce a drafting assistant the practice has licensed for administrative text. Her first draft opens with eight minutes on how language models work. She cuts it. The sign-in sheet has thirteen names: 2 physicians, 1 nurse practitioner, 1 triage nurse, 3 medical assistants, 2 front desk staff, 1 referral coordinator, 1 biller, 1 part-time coder and 1 practice manager, which makes thirteen. She sends a four-question form a week ahead asking role, tool used if any, what for, and one time it was wrong, with an explicit instruction to include no patient name, medical record number, date of birth, account number or clinical detail. Ten people reply. Four use something daily, two occasionally and four not at all, which makes ten. Three describe a drafted reminder that named a follow-up the fictional patient had never been scheduled for. Renata rebuilds the session as three tracks. The front desk staff and the referral coordinator practise one behaviour: before a drafted message is sent, the patient, the appointment, anything clinical it names and the contact channel are checked against the record, and the check is noted with a name. The medical assistants and the triage nurse practise a different one: when a drafted summary states something clinical, it goes to the clinician and is never relayed to a patient as a finding. The physicians and the nurse practitioner get four minutes on what a signature affirms. When the owner, Dr Aurelio Banks, asks what share of practices their size use AI, Renata says she has no verified figure for practices of that size, gives the AMA 2026 numbers with the 1,342 base on the adoption question and the 15 January to 2 February 2026 fielding window, notes that 9% of that base could not say which AI their own practice offers, and moves on. The 11:40 patient runs long and the session starts at 12:08. Because Renata built twenty-two minutes of content into a thirty-minute slot and put the practice round first, every role performs its behaviour once before the medical assistants stand up at 12:55.
Responsible-use warning
Never open a practice session with a definition of machine learning, and never show one role's work to a room of seven roles. Never demonstrate a task to a medical assistant or a front desk lead that would have them state a clinical finding to a patient, and never leave ownership as "someone". Never quote an AMA percentage without its base and its January to February 2026 fielding window, and never answer a question about practices of a particular size by inventing a figure you do not have.
Human-review requirement
The instructor owns the accuracy of every figure and every role description in the session material. Before delivery, the practice owner reviews the role and ownership table and confirms it matches how that practice actually delegates and supervises. The named privacy official confirms the pre-session form asks for no patient identifier and no clinical detail. Any clinical content appearing in the session is reviewed by a licensed clinician, and the clinician, not the instructor, owns every clinical statement made in the room and every signature that follows it.
Practical exercise
Profile a fictional practice room by role, output and owner. Use a fictional practice, fictional staff and fictional form answers throughout, and no real practice, patient or payer.
- Write a fictional sign-in sheet of ten to eighteen people for a thirty-minute session at a fictional primary care or single-specialty practice. Record each person's role.
- Count the attendees by role and confirm the counts add to the sheet total. Show the addition.
- For each role, record one AI-assisted output that role could plausibly produce in an ordinary week at that practice.
- For each output, name the owner: the licensed treating clinician for clinical content and the signature, the coding professional working from the clinician's documentation for a code, or the practice manager or another named administrative owner for administrative output.
- For each role, write the single check that matters most before that output reaches a patient, a payer or the record.
- Mark any task on your list that you must not demonstrate to a particular role, and write one sentence saying why, naming the clinical line rather than a rule you cannot cite.
- Write one sentence you may truthfully say about the AMA 2026 adoption figures: who published it, who was asked, when, and what it does not show.
- Write the exact sentence you will say when an owner physician asks what share of practices their size use AI.
- Draft a four-question pre-session form that asks for no patient name, medical record number, date of birth, account number or clinical detail.
- Write a thirty-minute plan as segments and minutes totalling twenty-two minutes of content, show the addition, and say in one line which behaviour survives if you are cut off at minute eighteen.
- Strike any sentence in your plan that asserts a finding about how adults learn, and relabel the rest as best practice or as this programme's recommendation.
Knowledge check
-
How should an instructor describe a medical practice audience when designing a session?
Answer guidance: By role, by the AI-assisted output each role produces in an ordinary week, and by the named person who owns that output under licensure, delegation and the practice's own written rules.
-
What must accompany any AMA 2026 survey figure quoted in a session?
Answer guidance: Its question-level base, such as the 1,342 respondents on the adoption question, the 15 January to 2 February 2026 fielding window, and the fact that it is a self-reported volunteer survey with no stated margin of error.
-
Why is the 9% who were uncertain which AI use cases their practice offers useful to an instructor?
Answer guidance: Because it describes the room directly: some learners are already using a tool they cannot name, inside a workflow nobody has written down, so a session must ask what is in use before it teaches anything.
-
What may an instructor never do from the front of the room?
Answer guidance: Give clinical advice, triage a patient, make a coding determination for a real patient or claim, or interpret a specific payer's policy as binding, because none of those belong to a teaching relationship.
Sources and further reading
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.
- American Medical Association, Center for Digital Health and AI — 2026 Physician Survey on Augmented Intelligence (2026-03-01). Accessed 2026-09-15.
- U.S. Department of Health and Human Services, Office of Inspector General (HHS-OIG) — General Compliance Program Guidance (GCPG) (2023-11-06). Accessed 2026-09-15.
- ECRI — Misuse of AI chatbots tops annual list of health technology hazards (2026-01-21). Accessed 2026-09-15.