Lesson objective
Triage a practice task on consequence to the patient, data class, verifiability and who owns the output, then route it to standard review, enhanced review, redesign or no AI before any tool is opened.
Plain-language explanation
Why this matters now
Level 1 explained how these tools produce answers, how they fail, and which duties stay with the people in the practice. Level 2 is about an ordinary Tuesday: a full schedule, a fax tray, a portal inbox and an approved assistant already open on the desktop. The first decision on any task is made before anyone types anything.
Accountability for clinical work is already settled, and it does not move because a tool was involved.
The same document lists purely administrative tasks such as patient scheduling alongside clinical documentation and decision support as current AI applications in a physician's workflow. That whole range lands on one desk in a small practice, which is why the first question in Level 2 is always what kind of task this is, not which tool to use.
A question to sit with
Two items arrive at 8:30 a.m. The first is a list of twelve patients whose Thursday appointments have to move because a clinician is away. The second is a portal message from a patient asking whether she should keep taking a medication that is making her dizzy. Both could be pasted into the same assistant in the same minute. What makes them different tasks?
What goes wrong
Choosing the tool before the task. Someone opens the assistant because it is already open, then works out what to ask it. Nobody asks whether the task should go near AI at all.
Triaging on how busy the day is. "It will save twenty minutes" is a benefit, not a triage answer. A saved twenty minutes that produces a wrong appointment letter to twelve patients costs far more than it saved.
Asking the data question after the paste. Once patient information is inside a tool, the decision about whether it should have gone there has already been made by default. The data class has to be settled before anything is typed or uploaded.
Assuming every output can be checked. Some outputs can be verified in seconds against something the practice already holds: a date on the schedule, an address in the registration record, a sentence in a payer policy the staff member pasted in. Others could only be checked by doing the whole task again.
Losing track of who owns the output. A draft moves from the front desk to the practice manager to a clinician, and each assumes one of the others is the owner. If you cannot name the person who will own the output before you start, the task is not ready.
Letting the arrival route decide the category. A question about a medication does not become administrative because it arrived in the portal inbox rather than by phone. Content, not channel, decides whether a task is clinical.
The practical answer
Ask four questions, in order, and write the answers down. The four questions are this programme's recommendation; the duties they draw on are cited beside them.
| # | Question | What you are looking for | Typical answers |
|---|---|---|---|
| 1 | Consequence: if the output is wrong and nobody notices, where does the error land? | The furthest place the output travels | An internal note; a letter to a patient; the medical record; a claim or an authorisation request; a clinical decision |
| 2 | Data class: what would the tool see? | The most sensitive item in the material | No patient information; administrative patient information; the clinical record; specially protected records |
| 3 | Verifiability: can every part of the output be checked against something you hold? | How the check is done and what it costs | Fully, against the schedule, the record or the pasted policy; partly; only by redoing the work |
| 4 | Who owns it: which qualified person owns the result? | A named person, not a job title in the abstract | The licensed treating clinician; the coding professional; the privacy official; the practice manager |
Question 1: consequence. Follow the output to its final destination, not its first. A summary written "just for the file" may be read later by someone assembling a prior authorisation request. Rate the task by where an uncaught error could finish, and be honest that a patient is usually at the end of the line.
Question 2: data class. Data class and volume are a legal question before they are a practical one.
At triage you need the headline only: what is the most sensitive thing this task would expose, and is the system approved for it? Module 13 works the classification through in detail.
Question 3: verifiability. Ask how you would prove the output right, using something already in your hands. If the honest answer is "I would have to work the whole thing out myself," AI is adding a review burden rather than removing one, and the task should be split or done without it.
Question 4: who owns it. AI drafts, summarises, transcribes, extracts and flags. It never diagnoses, triages, prescribes, signs a note, assigns a final code or speaks to a patient as though it were a licensed person.
Now combine the four answers into a route.
| Route | When it applies | What happens next |
|---|---|---|
| A: proceed, standard review | Internal consequence; data cleared for an approved system; output fully verifiable; owner named | Run the assisted-task workflow in Lesson 2; the preparer verifies every field; the owner checks the verification |
| B: proceed, enhanced review | Output reaches a patient, the record, a payer or a claim; data cleared; output fully or largely verifiable; owner named | Run the workflow; the named owner checks every fact, name, date and figure against the source before release |
| C: redesign, then triage again | Data is not cleared for any approved system, or part of the output could be verified only by redoing the work | Reduce the data, split the verifiable part from the judgement, or get the privacy official's decision; triage again |
| D: no AI | No owner can be named, or the output would be a clinical judgement, or no approved system exists for this data | Do the task without AI, and route anything clinical to the licensed clinician |
Remember this
- Triage happens before any tool is opened: consequence, data class, verifiability, owner.
- Rate consequence by where an uncaught error could finally land, which in a practice usually means a patient.
- Minimum necessary is a legal duty on the practice, and it applies to what you paste as much as to what you send.
- If the output could be verified only by redoing the work, split the task or drop the tool.
- Content, not the channel it arrived by, decides whether a task is clinical; anything clinical is the licensed clinician's.
Medical practice example
Cedar Hollow Family Medicine is a fictional five-clinician practice. On a Tuesday morning the practice manager, Rosa Delgado, has four tasks and an approved AI assistant covered by a business associate agreement. She triages before opening it. Task one: letters for twelve patients whose Thursday slots must move. Consequence: a letter reaching patients. Data: names, old and new appointment times, clinic address. Verifiable: fully, against the schedule. Owner: Rosa. Route B. Task two: a portal message from a fictional patient asking whether she should stop a medication that is making her dizzy. Owner: the licensed treating clinician, Dr. Amara Boateng. Route D for the clinical answer; the assistant drafts only a neutral acknowledgement that the message has gone to the clinical team, which Rosa reads before sending. Task three: a summary of a payer's published appeals policy, pasted in full into the prompt, so the billing team knows the deadline and the address. Consequence: internal note, though an appeal will rely on it. Data: no patient information at all. Verifiable: fully, against the pasted text. Owner: the billing lead, Priya Nair. Route A, with a note that the deadline must be read back from the policy itself. Task four: a request from a new nurse practitioner for the assistant to "look over" a scanned outside record and say what matters. Rosa stops at question 3: nobody could check that output without reading the whole record, and question 4 puts anything clinical with the clinician anyway. Route D, replaced in Module 05 by a checklist-driven extraction the practice owns. The four triages take Rosa seven minutes, and she writes each one-line entry before any work starts.
Responsible-use warning
Do not let an approved system make triage feel unnecessary. Approval means the practice has cleared a system for certain data and certain uses. It does not mean every task suits it, that every output can be checked, or that someone owns the result. The two failures that cost most are sending a task whose output could only be verified by redoing it, and treating a clinical question as administrative because it arrived in an administrative queue. Triage the task, not the tool, and write the route down before you begin.
Human-review requirement
The person who will own the output confirms the triage before work starts. For anything clinical, including any communication that pertains to clinical information, that is the licensed treating clinician; for administrative outputs it is the practice manager or named administrative owner; for coding it is the coding professional working from the clinician's documentation; for questions about what data may go where it is the privacy official. That person checks that consequence was followed to its final destination, that the data class is right and cleared, that a practical verification method exists, and that their own name is on the task. Any of them may move a task to a stricter route. AI never assigns the route.
Practical exercise
Use fictional tasks and fictional patients only. Do not use real patient information.
- List eight tasks from a typical week in your role: for example an appointment letter, a portal message, a referral packet, a payer policy question, a records request, a recall list, a form and a report for the owner.
- Replace every real detail with a fictional one before you write anything down.
- For each task, answer the consequence question by following the output to its final destination, not its first.
- Record the data class of the most sensitive item each task would expose, and whether an approved system covers it.
- Describe in one sentence how each output would be verified, then mark it fully, partly or only by redoing the work.
- Name the qualified person who would own each output: licensed clinician, coding professional, privacy official or practice manager.
- Assign a route, A to D, using the decision table in this lesson.
- For every Route C task, rewrite it as two tasks: the verifiable part AI may prepare, and the judgement that stays with the qualified person.
- For every Route D task, write one sentence naming which of the four answers ruled AI out, so the decision itself is on file.
- Ask a colleague to triage the same eight tasks without seeing your answers, then compare. For each disagreement, note which question caused it and what the owner of the output would say. Record the results in your workbook.
Knowledge check
-
What are the four triage questions, and when are they asked?
Answer guidance: Consequence (where an uncaught error would land), data class (what the tool would see), verifiability (how every part of the output would be checked) and who owns the result. They are asked and written down before any tool is opened.
-
What does the FSMB 2024 document say about who is responsible for AI use, and whom does that statement bind?
Answer guidance: It says state medical boards do not regulate tools or technologies, only the licensed physicians that use those tools, and that the physician is ultimately responsible for the use of AI. It is FSMB guidance to boards, binding only where a state board adopts rules or applies its existing disciplinary authority.
-
How does the minimum necessary standard apply at triage?
Answer guidance: 45 CFR 164.502(b)(1) requires reasonable efforts to limit PHI to the minimum necessary for the purpose, so the data class question asks what the tool would see before anything is pasted, not afterwards.
-
Why does a task whose output can be verified only by redoing the work go to redesign rather than straight to AI?
Answer guidance: Because the check would cost as much as the task, so AI adds risk without saving review time. The task is split into a verifiable part AI may prepare and a judgement the qualified person makes, then triaged again.
Sources and further reading
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.
- Federation of State Medical Boards — Navigating the Responsible and Ethical Incorporation of Artificial Intelligence into Clinical Practice. Accessed 2026-09-15.
- American Medical Association (AMA), House of Delegates — Augmented Intelligence Development, Deployment, and Use in Health Care (AMA Board of Trustees report and policy adopted November 12, 2024) (2024-11-12). Accessed 2026-09-15.
- U.S. Department of Health and Human Services (Electronic Code of Federal Regulations) — 45 CFR 164.502(b) Standard: Minimum necessary; 45 CFR 164.514(d) Implementation specifications. Accessed 2026-09-15.