Lesson objective

Explain how volume, handoffs, staff turnover and fading verification change the risk of an AI-assisted practice task once several people run it across thousands of patient encounters, and convert a measured correction rate into an expected monthly count.

Plain-language explanation

Why this matters now

Level 2 taught one person to run one AI-assisted task well: decide whether the task belongs to a tool at all, write the instruction, check the output against the patient file or the payer policy, record what was done and have the right person sign it. Level 3 is about the day that task stops being a task.

A draft reply to a patient portal message, written by a medical assistant who reads the chart, checks every statement against the note and sends it to the clinician for approval, is a task. The same drafting running against every routine portal message in a four-clinician practice, produced by six people across the front desk and the clinical support team, is an operation. The software can be identical. The risk is not, because almost nothing that made the task safe survives the change without being rebuilt on purpose.

The scale is real. Professional guidance (survey, not binding): the American Medical Association's 2026 Physician Survey on Augmented Intelligence, fielded January 15 to February 2, 2026 with 1,692 US physicians, found that on the adoption question (n=1,342) 72% had incorporated one or more AI use cases, and that the average number of use cases per respondent rose from 1.1 in 2023 to 2.3 in 2026. Reported uses included documentation of billing codes, medical charts or visit notes (28%) and generation of draft responses to patient portal messages (19%). This is a self-reported survey of volunteer respondents with no stated margin of error or weighting, so read it as a direction of travel rather than a national measurement. The direction is the point: practices are not running one use case carefully, they are running several at once.

A question to sit with

A medical assistant who reads every AI-drafted portal reply against the chart catches a wrong statement about once a fortnight and fixes it before anyone sees it. Next year the practice drafts replies to every routine message, six staff share the work, and two of them started last month. How many wrong statements should the practice now expect in a month, who will find them, and what would tell anyone if they had stopped being found?

What goes wrong

Volume turns a rate into a count. A method that needs a correction in a small share of outputs feels safe at twenty outputs a week. Run it two thousand times a month and the same share is a steady stream of wrong sentences moving toward patients, charts and claims. Nothing about the method changed except how often it runs.

Handoffs strip the context that made checking work. The person who piloted the method knew which of the practice's clinicians dictate results in shorthand, which payer had changed a policy in March, and which patients have a legal guardian recorded in a place the tool never sees. The sixth person to use the method knows none of that. Each handoff passes on the output and loses the knowledge needed to judge it.

Turnover erases the reasons. A procedure records what to do. The reasons live in the head of whoever wrote it. Front-desk and clinical support roles turn over, and when the author leaves, the steps survive as rituals nobody can defend or safely change.

Shared inputs make errors arrive together. Six people working their own way make six different mistakes, and colleagues notice different things. Six people using one shared instruction, one template letter or one list of routing rules make the same mistake in every chart at once. Consistency then looks like quality when it is really a single common cause.

Checking fades when it keeps finding nothing. A reviewer who reads three hundred drafts and finds two problems learns, quite reasonably, that reading rarely pays. Reading becomes skimming, skimming becomes a glance at the first line. The step is still written in the procedure and is no longer happening.

Silent errors raise no flag. Most operations count what the tool could not do: the message it refused to draft, the field it left blank. A confidently wrong sentence is not one of those. It passes every counter the practice has, which is why a flag rate is never an error rate.

The practical answer

Separate the risks that scale multiplies from the risks that scale creates. They need different controls, and a better instruction fixes only the first kind.

Risk Present when one careful person does it? What scale does Control that addresses it
The tool states something the chart does not support Yes Multiplies it Verification against the source record, as in Level 2
Patient information reaches a tool with no agreement in place Yes Multiplies it Tool approval and access control (Module 7)
Staff run different, unrecorded versions of the method Barely Creates it One written procedure with one owner (Module 4)
Verification quietly stops while the procedure still names it Barely Creates it Sampling and quality assurance (Module 13)
One defect in a shared instruction reaches every patient it touches No Creates it Version control and staged rollout
Nobody owns the operation No Creates it A named operation owner (Lesson 2)

Do the frequency arithmetic before you widen the work. Take the correction rate the practice has actually measured, multiply it by the planned volume, and write the answer as a sentence somebody has to read out loud.

If that sentence is unacceptable, the question is not whether the tool is good enough. It is which control changes the number: a narrower set of message types, a heavier review gate, a smaller rollout, or moving the tool to a different step in the workflow.

Ask four questions of any task before you scale it. These are this programme's recommendation, not a regulator's test.

Question Why it matters If the answer worries you
How often will it run? Frequency converts a rate into a count Write the expected monthly count of wrong outputs into the memo
Who will run it, and for how long? More people and turnover mean more variation Write the procedure and its reasons down before widening use
How far is the reviewer from the source? Review needs the chart, not a summary of the chart Keep a sampled path back to the record the output describes
What do all the runs share? Shared inputs cause errors that arrive together Put every shared instruction, template and routing rule under version control

Map the work before you manage it. Professional guidance (voluntary framework): NIST's Artificial Intelligence Risk Management Framework (AI RMF 1.0, NIST AI 100-1, released January 26, 2023) says that after instituting its GOVERN outcomes, "most users of the AI RMF would start with the MAP function" and continue to MEASURE or MANAGE, iteratively. It is voluntary, binds no practice or physician, and NIST's page stated as of September 15, 2026 that AI RMF 1.0 is being revised under the White House AI Action Plan. The operational point stands on its own: understand how the work actually runs before deciding how to control it. Module 2 builds that map.

Remember this

  • Scale multiplies risks you already know and creates new ones: version drift, fading verification, errors that arrive together, and an operation nobody owns.
  • Convert your own measured correction rate into an expected monthly count and write it in the rollout memo before you widen the work.
  • A flag rate is not an error rate, because a confidently wrong sentence raises no flag.
  • Every shared instruction, template and routing rule needs a version, an owner and a way back to the last good one.
  • If honest review cannot keep pace, narrow the operation rather than loosen the sign-off.

Medical practice example

Larkspur Orthopedic Associates PLLC is a fictional seven-clinician practice with a busy referral desk. Referral coordinator Priya Raman builds a method that summarises incoming referral packets into a structured intake note: reason for referral, imaging already performed, the referring clinician and the payer named on the face sheet. For two months she runs it on her own desk, reads every packet behind every summary, and revises the instruction three times, once because a scanned fax put the payer name in a header the tool ignored. Her supervisor spot-checks and the clinicians are pleased, so the practice extends the method to all four coordinators and to the surgical scheduling team, about 900 packets a month. Nine months later two things have happened that nobody decided. First, the instruction has forked. Most staff use the text Priya pasted into the team chat in week one; two use a version a colleague edited to pull the referring clinician from the signature block instead of the letterhead. Nothing records which packets were processed with which. Second, the reading behind the summary has thinned. The procedure still says to open the packet before accepting the summary, but on heavy days most coordinators accept the structured fields and open the packet only when something looks odd. In month ten a scheduler notices an intake note naming a referring clinician who had left the referring group two years earlier. Priya traces it: the edited version reads the signature block, and one referring group signs its letters with the practice founder regardless of who saw the patient. Every packet from that group processed with the edited version carries the wrong referring clinician, and the practice cannot say from its records how many packets that is or which months they fall in. Rebuilding the answer from chat timestamps and staff memory takes two days. Priya's method was sound. The operation built around it had no version control, no sampling of the reading step and nobody who owned it.

Responsible-use warning

Do not read a successful pilot as evidence that a method is safe to scale. A pilot run by the person who designed it, on work they know, with checking they do themselves, tests the method under the one condition that disappears the moment it is handed over. Version drift, fading verification, a single shared defect reaching every patient it touches and outcomes nobody owns are all invisible in a pilot, because one careful operator suppresses all four. Ask what breaks at six operators and two thousand outputs a month before the rollout, not after it.

Human-review requirement

The qualified person who owns each output still owns it at scale. The licensed treating clinician owns the content of the medical record and any patient communication carrying clinical information; the clinician whose documentation supports the claim and the coding professional who assigns codes own the codes; the practice manager owns schedules, forms, letters and operational reports. At scale that person judges two further things: whether the review actually performed still matches the review the procedure describes, and whether volume is inside what the named reviewers can genuinely read. If it is not, the answer is to narrow the scope, add reviewers or pause the rollout, never to sign a record that overstates what was checked.

Practical exercise

Use fictional or fully anonymised material only. No real patient information.

  1. Choose one AI-assisted task your practice already performs or plans to perform: portal message drafts, referral intake summaries, prior authorization packet assembly, recall letters or visit-note drafting. Write the steps as they are actually done, not as the policy describes them, in no more than one page.
  2. Record the current volume in outputs per month, and the volume the task would reach if every eligible encounter or message went through it.
  3. Take your last 20 reviewed outputs and count how many needed any correction. Calculate the rate, multiply it by the projected volume, and write the result as one sentence naming the expected monthly count of outputs containing a correction.
  4. List every shared input the task depends on: instructions, templates, letter shells, routing rules, tool settings, reference documents. Beside each, write who owns it today and how a change to it would be recorded.
  5. Identify every step that exists only to catch an error. For each one, write down what evidence, other than asking the person, would show whether that step is still happening.
  6. Name the single shared input whose defect would reach the most patients at once, and estimate how many patients that would be in one month.
  7. Using the risk table in this lesson, mark each risk in your task as multiplied by scale or created by scale, and note which control in the right-hand column your practice does not yet have.
  8. Write the three questions you would need answered before widening the task, and note which later module of this level takes each one up.

Knowledge check

  1. Which risks does scale multiply in a practice AI task, and which does scale create?

    Answer guidance: Scale multiplies risks already present when one careful person does the work, such as the tool stating something the chart does not support, or patient information reaching an unapproved tool. It creates risks that barely exist in a single task: staff running different unrecorded versions, verification quietly stopping, one shared defect reaching every patient it touches, and an operation nobody owns.

  2. A pilot found 6 corrections in 150 drafts and the practice plans 1,800 drafts a month. What belongs in the rollout memo?

    Answer guidance: The measured rate is 6 divided by 150, which is 4%, and 1,800 multiplied by 0.04 is 72. The memo should state that the practice expects about 72 drafts a month to contain something a reviewer must correct, and that the review design has to be capable of finding all of them.

  3. Why is the number of outputs a tool flags never the same as the number of outputs that are wrong?

    Answer guidance: A flag counts what the tool itself could not handle, such as a message it declined to draft or a field it left blank. A confidently wrong sentence produces no flag at all, so it passes every counter the practice has. Only comparing outputs against the source record can reveal it.

  4. What did the AMA 2026 survey report about adoption, and what are its limits?

    Answer guidance: On the adoption question with a base of 1,342 respondents, 72% had incorporated one or more AI use cases and 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 fielded between January 15 and February 2, 2026, with no stated margin of error or weighting, so the percentages are not precise national estimates.

Sources and further reading