Lesson objective

Run an assisted practice task through seven fixed steps, from assembling verified source material to a named owner's review, so that the output is checked against a source rather than trusted because it reads well.

Plain-language explanation

Why this matters now

Triage says whether a task may use AI and on what terms. This lesson is what happens next: the same seven steps, in the same order, every time, so that two people doing the same task produce work a reviewer can check the same way.

The reason the steps are fixed is that the failure mode is quiet. A wrong appointment date does not look wrong. A confidently written sentence about a payer's deadline looks exactly like a right one.

The same document names the human side. Its Human-AI Configuration risk includes people "experiencing algorithmic aversion, automation bias, over-reliance, or emotional entanglement" with these systems, and Section 2.7 describes automation bias as "excessive deference to automated systems," which "can exacerbate other risks of GAI, such as risks of confabulation." A workflow that requires a check against a source is how a practice designs around that.

A question to sit with

A scheduler produces a perfect-looking recall letter in forty seconds and spends ten minutes checking it against the recall list. A colleague produces the same letter in forty seconds and sends it. Six months later, which of the two can show what happened?

What goes wrong

Starting with the instruction instead of the material. Someone asks the tool for a letter about a patient, relying on the model to supply facts it has no way to know, instead of pasting the facts from the practice's own systems.

Letting the draft become the source. The draft is edited, the edit is edited, and by the third round nobody has looked back at the schedule or the payer policy the work started from.

Reading for tone instead of checking for facts. The reviewer fixes a clumsy sentence and never verifies the date, the address or the amount.

Checking the parts that are easy to check. The name and the date get verified because they are quick; the sentence describing the payer's deadline does not, because that would mean opening the policy.

No named owner at the end. The output goes out under the practice's name with nobody's review behind it.

Skipping the record because the task was small. Small tasks are the ones that repeat, which makes them the ones worth having a record of.

The practical answer

Seven steps. They are this programme's recommendation, not a legal or professional requirement; the duties they serve are cited where they apply.

Step What you do What must be true before you move on
1. Name the output and its owner Write one sentence saying what will exist at the end and who owns it A named person, not a role in the abstract
2. Assemble the source material Gather the facts from the practice's own systems and documents Only what the task needs, and nothing more
3. Write the instruction State the task, the audience, the format, and the constraints The instruction says to use only the supplied material
4. Generate the draft Run it in the approved system The system is approved for this data class
5. Verify field by field Check every fact, name, date, figure and quoted rule against the source Every checkable item has been checked, not sampled
6. Owner's review The named owner reads and takes responsibility for the output Anything clinical has gone to the licensed clinician
7. Record Log the task, the system, the source material, the verification and the reviewer The entry is written now, not reconstructed later

Step 2 in detail: the source material is the point. The strongest single habit in practice AI work is supplying the facts rather than asking for them. Paste the payer's published policy text, the relevant extract of the form, the schedule lines, the registration fields. Everything the model did not receive, it may invent. Everything it did receive, you can check it against. Minimum necessary applies here: 45 CFR 164.502(b)(1) requires reasonable efforts to limit PHI to the minimum necessary to accomplish the intended purpose, so the source packet is built by adding what the task needs, not by removing what it does not.

Step 3 in detail: constraints that keep the output in its lane. Four constraints belong in almost every practice instruction:

  1. Use only the material supplied; if something is missing, say so rather than filling it in.
  2. Produce administrative content only; give no clinical advice, no assessment of symptoms and no statement about whether a treatment is appropriate.
  3. Do not state any rule, deadline, policy or figure that is not in the supplied material.
  4. Mark anything uncertain, so the reviewer can see where to look.

Module 02 builds these into a full prompt architecture. At this stage the point is only that constraints are part of the workflow, not a personal style.

Step 5 in detail: verification is a list, not a feeling. Turn the output into a list of checkable items and tick each one against its source. A useful default for practice work:

Item type Checked against
Patient name, date of birth, address The registration record
Appointment date, time, clinician, location The schedule
Payer name, plan type, member identifier The insurance record
A deadline, an address for submission, a form name The payer's published policy text you supplied
A statement of what the practice will do The practice's own policy or the owner's instruction
Anything clinical Not verified by the preparer at all; routed to the licensed clinician

An AI output is never proof of anything. A citation, a rule, a code reference or a figure that came out of a model is checked against the primary source before anyone acts on it, however plausible the explanation beside it looks.

Step 6 in detail: the owner's review is where the work becomes the practice's.

Professional guidance (United States; Joint Commission and Coalition for Health AI, released September 17, 2025; voluntary, written mainly for hospitals and health systems, so its application to a small practice is by analogy): the guidance treats health AI tools as including clinical documentation and patient communication as well as administrative services such as "revenue cycle management, coding, prior authorization." Its third element says that if PHI is involved, organisations should execute appropriate business associate agreements, apply the minimum necessary standard, and ensure uses comply with the HIPAA Privacy, Security and Breach Notification Rules. Its fourth element says monitoring should be risk-based: tools that inform or drive clinical decisions "should be checked more often," while tools that help with administrative or documentation tasks "can be checked periodically." Nothing in it is required of a practice.

Remember this

  • Supply the facts; do not ask the model to recall them. What it did not receive, it may invent.
  • Constraints are part of the workflow: supplied material only, administrative content only, no unsourced rules, uncertainty marked.
  • Verify field by field against a source, not by reading for tone; fluent wrong text reads exactly like right text.
  • An AI output is never proof; check any rule, figure or citation against the primary source first.
  • A named owner reviews before release, and anything clinical goes to the licensed clinician.

Medical practice example

Brightfield Pediatrics is a fictional four-clinician practice. Its referral coordinator, Daniel Okoro, has to write to 23 families whose fictional insurer, Vantage Mutual, has moved its referral submission address and shortened its appeal window. Step 1: the output is one letter template plus a merge list, owned by the practice manager, Lily Chen. Step 2: Daniel assembles the source material, which is the insurer's published bulletin, pasted in full, and a merge list holding only family name, address and the child's referral reference. He leaves out dates of birth and visit reasons, which the letter does not need. Step 3: his instruction says to use only the pasted bulletin and the merge fields, to produce administrative content only with no clinical statements, to state no deadline or address not present in the bulletin, and to mark anything uncertain. Step 4: he runs it in the practice's approved assistant, which is covered by a business associate agreement. Step 5: the draft is good, but verification finds two problems. The letter says appeals must be filed "within 30 days"; the bulletin says 45 days from the date on the denial notice. The draft also adds a friendly sentence saying the family "should not delay the referral for medical reasons," which is a clinical statement nobody asked for and which the constraint should have prevented. Daniel corrects the deadline against the bulletin and deletes the clinical sentence. Step 6: Lily reviews the corrected template, checks the submission address against the bulletin herself, and approves it; because nothing in it is clinical, no clinician review is needed, and she notes that if a family replies with a clinical question it goes straight to a clinician. Step 7: Daniel logs the task, the system, the two source documents, the two corrections and Lily's name. The whole job takes 40 minutes, of which 18 are verification.

Responsible-use warning

Do not let the draft replace the source. After the first round of editing it is easy to work only from the document on screen, and from that point nothing is being checked against anything. Verification means going back to the schedule, the registration record, the insurance record or the pasted policy text for every checkable item, not sampling the ones that are quick. A fluent sentence stating a deadline, an address or a rule that was never in your source material is the single most common way a wrong fact leaves a practice under its own name.

Human-review requirement

The named owner reviews before release and takes responsibility for the output: the practice manager or named administrative owner for letters, forms, schedules and operational reports; the licensed treating clinician for anything that pertains to clinical information, including any patient communication that touches symptoms, medication or treatment; the coding professional working from the clinician's documentation for anything that will support a claim; the privacy official for what data may enter which system. The reviewer checks the field-by-field verification rather than repeating it blind, confirms nothing clinical was produced by an administrative route, and confirms the record entry exists.

Practical exercise

Use fictional practices, fictional payers and fictional patients only.

  1. Choose one recurring administrative task from your own work and describe it in one sentence, with a fictional practice standing in for yours.
  2. Write step 1: the output in one sentence, and the named owner.
  3. Build the source packet for step 2. List every document and field you would supply, then cross out anything the task does not need and explain each deletion in four words or fewer.
  4. Write the step 3 instruction, including all four constraints from this lesson in your own words.
  5. Produce a fictional draft output yourself, deliberately including one invented fact, one clinical sentence that should not be there, and one correct but unsourced figure.
  6. Build the step 5 verification list: every checkable item in the draft, with the source you would check it against.
  7. Run your own verification and mark what you would have caught and what you might have missed.
  8. Write the step 6 review note the named owner would sign, in two sentences.
  9. Write the step 7 record entry for the task on the AI Use Record.
  10. Repeat steps 5 to 7 with a colleague verifying your draft instead of you, and note every item they checked that you did not.

Knowledge check

  1. Why does the workflow supply source material rather than asking the model to recall facts?

    Answer guidance: Because anything the model did not receive it may invent, and anything it did receive can be checked against a source. NIST AI 600-1 notes confabulations are a natural result of how generative models are designed and may include logic or citations that purport to justify an incorrect answer.

  2. What four constraints belong in almost every practice instruction?

    Answer guidance: Use only the supplied material and say what is missing; produce administrative content only with no clinical advice; state no rule, deadline, policy or figure that is not in the supplied material; and mark anything uncertain.

  3. What does AMA policy 2(b) say about content generated on a physician's behalf, and does it bind a practice?

    Answer guidance: It says AI cannot augment, create or otherwise generate records, communications or other content on behalf of a physician without that physician's consent and final review. It is AMA advocacy policy adopted November 12, 2024, not a legal requirement.

  4. What does the Joint Commission and CHAI guidance say about how often AI tools should be checked?

    Answer guidance: That monitoring should be risk-based: tools that inform or drive clinical decisions should be checked more often, while tools that help with administrative or documentation tasks can be checked periodically. The guidance is voluntary and written mainly for hospitals and health systems.

Sources and further reading