Lesson objective
Place any proposed practice automation on a three-level authority scale and choose the lowest level that still does the job.
Plain-language explanation
Why this matters now
Level 4 built governance around how a practice uses AI: named owners, a permitted-use inventory, a decision record, supplier diligence. All of that assumed a tool that answers when someone asks. Level 5 is about tools that do not wait to be asked.
In this level an automation means software that takes a step, or a series of steps, without a person starting each one. Something other than a request sets it going: a fax lands in the document queue, a referral arrives in the portal, a balance ages past thirty days, a recall date falls due. The automation then reads from the practice's systems and does something: files the fax to a pending queue, drafts a letter, opens a task for the referral coordinator, stages a claim for the biller. Vendors sell these as agents, bots, workflows, copilots or "automation modules". The label does not matter. The fact that it acts does.
That changes the question. For a tool that answers, the question is whether the answer is right. For a tool that acts, the question is what it is allowed to do, in which system, and who decided.
Two public sources shape this programme's answer, and neither binds a physician practice.
- Professional guidance (CISA, with the Australian Signals Directorate's Australian Cyber Security Centre and U.S. and international partners, announced 1 May 2026; voluntary): Careful Adoption of Agentic Artificial Intelligence (AI) Services recommends "Begin with agentic AI use cases that are low-risk and non-sensitive" and "Avoid granting broad or unrestricted access, especially to sensitive data or critical systems." CISA's announcement names expanded attack surface, privilege creep, behavioral misalignment and obscure event records among the risks. The audience is chiefly developers, vendors and operators, with an emphasis on critical infrastructure and defense; it was not written for a six-physician family medicine group, and this programme relies only on the announcement page.
- Professional guidance (NIST, AI RMF 1.0, NIST AI 100-1, released 26 January 2023; voluntary, non-sector-specific): Appendix C states that human roles in decision making and oversight "need to be clearly defined and differentiated", and that human-AI configurations can span "from fully autonomous to fully manual". NIST binds no practice and there is no NIST certification. As of 2026-09-15 the AI RMF page says version 1.0 is being revised under the White House AI Action Plan, with no revised version published.
AI Coalition Network recommendation (verified 2026-09-15): no state medical board rule, no federal regulation and no payer policy recorded in this programme defines how much autonomy a practice automation may hold. The scale below is this programme's teaching design. It is not a legal or professional requirement, and nobody may be told it is.
A question to sit with
If the automation you most want to switch on made one wrong move at 4.45 pm on a Friday, would a patient, a referring office or a payer see it before anyone in your practice did, and could your practice take it back on its own?
What goes wrong
The product default becomes the practice's decision. A patient-engagement platform ships with auto-reply to portal messages, auto-send of results letters, or auto-post of demographic changes switched on, because that demonstrates well in a sales call. Nobody chose it. It is now how the front desk runs.
"It only drafts" quietly becomes "it sends". A drafting helper saves the clinical inbox forty minutes a day, so someone adds a rule to release replies that "look routine". That rule is a decision to let software speak to patients, made in a settings menu by whoever holds administrator rights.
"Inside" gets drawn in the wrong place. Staff treat a payer portal, a clearinghouse queue or a health information exchange as "our system" because they log into it daily. A step taken there reaches somebody else's records.
An internal field drives an outward event. Changing an appointment type, a recall interval or a "do not contact" flag looks like tidying the practice-management system. But the reminder engine, the recall report and the claim scrubber all read those fields, so the change decides what happens to a patient later.
One setting across every department. An automation that is harmless indexing immunisation records is not harmless in the results inbox, on a controlled-substance refill queue, or on a behavioural health schedule. Authority is chosen per use, and often per department.
The practical answer
Use three levels of authority. This scale is an AI Coalition Network recommendation, this programme's design and nothing more.
| Level | Name | What the automation does | Who acts outward | Typical practice use |
|---|---|---|---|---|
| 1 | Suggest | Reads, extracts, summarises, drafts, flags or compares, then stops | A named person decides and performs the act | Drafting a referral cover letter for the coordinator to check |
| 2 | Act inside the practice | Takes a recorded, reversible step in a system the practice owns, visible to no one outside and triggering nothing outward | Nobody; nothing leaves and nothing enters the legal record | Filing an inbound fax to a pending queue and opening a task for the right staff member |
| 3 | Act outward with a named person's approval | Prepares an outward act and waits; a named person approves that specific act | A named person, one approval at a time | Sending a rescheduling letter a practice manager has read and released |
There is deliberately no fourth level. Nothing acts outward on its own. "Outward" means anything reaching a patient, a referring or receiving practice, a payer, a clearinghouse, a laboratory, a pharmacy, a vendor or the public; anything that becomes part of the legal medical record; anything that moves money; and anything that changes a patient's access to care or to their records.
Draw the practice boundary honestly. Inside means systems the practice owns and controls where a change reaches no one else and nothing downstream reads it: an internal task queue, a staging area for documents awaiting review, an internal-only note field. Outside includes the patient portal, the clearinghouse, the payer portal, the e-prescribing network, the reference laboratory interface, the fax line, the text and telephone system, the e-signature platform, the card processor, outside email, and the chart itself. Writing to the legal medical record is not an internal act, however local it feels, because the clinician's signature has to stand behind what is there.
Test the "inside" fields that drive outward events. Some practice-management fields are internal in form and outward in effect: appointment type, recall interval, provider of record, preferred contact channel, do-not-call flags, interpreter need, insurance package, and any status a worklist filters on. A step that writes one of those cannot sit at level 2, because a later automation or a colleague will act on it as though a person had checked it.
Choose the lowest level that does the job. Ask five questions of every proposed automation and take the lowest level the answers allow.
- Does the step leave the practice, enter the legal medical record, move money, or change a patient's access to care or records? If yes, the ceiling is level 3.
- Can the practice undo the step completely and alone, before anyone outside sees it and before a later step relies on it? If no, it cannot run at level 2.
- Is the step on the never list in Lesson 4? Then no level applies; it belongs to a person from the start.
- Does level 1 already capture most of the time saved? Finding, sorting and drafting are usually the slow part; releasing a checked draft takes seconds.
- Could a cross-covering colleague tell from the register alone what this automation may do? If not, the scope is too loose to run.
Write the level down. The level, the systems, the named owner and the stop go into the authority register in Lesson 3. A level that exists only in a settings screen changes the first time someone updates the platform.
Make approval survivable. A level 3 approval that takes four minutes on a desktop will be skipped by a referral coordinator clearing ninety packets a week. Before choosing level 3, confirm the approver can see the exact text, the exact recipient and the source of every figure, where they work. If not, the honest design is level 1 with a clean hand-off. Module 07 builds approval gates properly.
Remember what a level does not do. A low level does not make an output correct. A level 1 draft can carry a wrong date of birth, a wrong insurer or a wrong referring physician, and the person who releases it owns the result. An AI output is never proof: a date, an identifier, a plan name or a phone number that an automation produced is checked against the source record before anyone relies on it.
Remember this
- An automation is software that takes steps without a person starting each one; the vendor's label is irrelevant.
- Three levels: suggest, act inside the practice, act outward only with a named person's approval. Nothing acts outward alone.
- The portal, the clearinghouse, the fax line, the laboratory interface and the legal medical record are outside; so are internal fields that later automations read.
- Take the lowest level the five questions allow, and test reversibility against outsiders and downstream steps.
- CISA's 2026 agentic guidance and the NIST AI RMF are voluntary; no recorded rule sets permitted autonomy for a practice automation as of 2026-09-15.
Medical practice example
Cedar Hollow Family Medicine is a fictional seven-clinician practice with fourteen support staff in two locations. Its operations lead, Priya Nandakumar, is asked to "turn on the front-desk agent" bundled with a patient-engagement platform the practice bought for online scheduling. The demonstration shows the automation reading every portal message, answering questions about appointments and medications, updating demographics and insurance in the practice-management system, and emailing visit-summary letters, all automatically once an internal confidence score passes a threshold. Priya runs the five questions over each step. Answering a medication question is clinical and belongs to a licensed clinician, so it fails question 3 outright. An insurance package change is internal in form but the claim scrubber and the eligibility check both read it, so it fails question 2. A visit-summary letter leaves the practice and quotes clinical content, so its ceiling is level 3 and its author must be a clinician, not the platform. Demographic changes are reversible from the practice's own audit history, but the preferred-contact-channel field drives the reminder engine, so Priya splits it: address and employer at level 2, contact channel and interpreter need at level 1. She then times a week of portal traffic and finds that reading, classifying and routing messages take roughly two-thirds of the effort, while releasing a checked reply takes seconds. Her recommendation is level 2 for triage, level 1 for drafting, and no automatic sending at all. Practice owner Dr Alma Reyes asks why the confidence score is not enough. Priya answers that a confidence score is the tool's own opinion of itself, and that a wrong medication statement would reach a patient before any person in the building saw it. The register entry records level 2 for triage, level 1 for replies and demographic proposals, names Priya as automation owner and the front-desk lead as output owner for administrative text, names Dr Reyes as output owner for anything clinical, and states that automatic reply and automatic letter release stay switched off.
Responsible-use warning
Do not let a product's shipped defaults become the practice's authority decision. Auto-reply, auto-send and auto-post settings are frequently on out of the box because they show well in a demonstration. Every outward act must be a choice written down by a named person, and the default answer for anything that reaches a patient, enters the chart, moves money or changes access to care is that a person approves it first. A vendor's confidence score is the tool's opinion of itself and never substitutes for that approval.
Human-review requirement
The automation owner named in the register selects the level and records the reasoning. The person who owns the output decides whether that level is acceptable: the licensed treating clinician for anything clinical or anything entering the legal medical record, the practice manager or named administrative owner for schedules, letters, forms and operational records, and the practice owner for anything touching the practice's compliance obligations. They confirm that nothing leaves without a named approval, that reversibility was tested against downstream steps rather than assumed, and that the approver can realistically give that approval where they work.
Practical exercise
Use a fictional practice, or your own team with no patient data of any kind and no real names.
- List six operational tasks your practice would like an automation to handle, from at least four of these areas: front desk and scheduling, registration and eligibility, the clinical inbox, referral coordination, document handling, and the billing office.
- For each, write the event that starts it and the last step it would take before stopping.
- Mark every system each task touches as inside the practice or outside it. Treat the patient portal, the clearinghouse, payer portals, the laboratory interface, the fax line, the telephone and text system and the legal medical record as outside.
- List any practice-management field the task would write that a later automation or worklist reads, such as appointment type, recall interval, preferred contact channel or insurance package.
- Apply the five questions in order to each task, and write down which question, if any, caps the level.
- Assign each task level 1, level 2 or level 3, or decide it should not be automated at all, with one sentence of justification.
- For every level 2 task, describe how the practice would restore the prior state from its own records, and how long that would take on a busy morning.
- For every level 3 task, name the approving role, state whether the approver must be a licensed clinician, and describe what they would see on screen before approving.
- Identify one task where the product would default to more authority than you chose, and name the setting that must be switched off and who controls it.
- For the highest-authority task, describe the worst plausible mistake at 4.45 pm on a Friday, who outside the practice would see it first, and what the practice would do on Monday.
Knowledge check
-
What are the three levels of automation authority taught in this lesson?
Answer guidance: Level 1 suggest, where the automation reads, drafts or flags and stops; level 2 act inside the practice, a recorded and reversible step in a practice-owned system that triggers nothing outward; and level 3 act outward only with a named person's approval of that specific act. No level acts outward on its own.
-
Why is writing to the legal medical record never treated as an internal act?
Answer guidance: Because the chart is the record a clinician signs and stands behind, and because payers, other clinicians and the patient rely on it. A write there is outward in effect even though it happens inside the practice's own system, so it needs a named person's approval.
-
Give three examples of practice-management fields that look internal but drive outward events.
Answer guidance: Appointment type or duration, recall interval, preferred contact channel, do-not-call or do-not-text flags, interpreter need, provider of record, and insurance package. Reminder engines, recall reports, worklists and claim scrubbers all read these fields, so a wrong value acts later.
-
What does CISA recommend about where an organisation starts with agentic AI, and does it bind a physician practice?
Answer guidance: CISA's 1 May 2026 joint guidance recommends beginning with low-risk, non-sensitive use cases and avoiding broad or unrestricted access to sensitive data or critical systems. It is voluntary guidance aimed mainly at developers, vendors and operators, and binds no physician practice.
Sources and further reading
- Cybersecurity and Infrastructure Security Agency (CISA), Australian Signals Directorate's Australian Cyber Security Centre, and U.S. and international partners — Careful Adoption of Agentic Artificial Intelligence (AI) Services (2026-05-01). Accessed 2026-09-15.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023-01-26). Accessed 2026-09-15.
- OWASP Foundation, OWASP GenAI Security Project — OWASP Top 10 for Agentic Applications for 2026 (2025-12-09). Accessed 2026-09-15.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.