Lesson objective
Distinguish rules-based automation, machine learning and generative AI in practice software, and match the check you apply to the way each one fails.
Plain-language explanation
Why this matters now
"AI" is now printed on products that touch almost every desk in a physician practice: the scheduling system that sends reminders, the claim software that flags a line before it goes out, the phone system that answers after hours, the feature in the electronic health record (EHR) that drafts a reply to a portal message, and the ambient scribe that turns a visit conversation into a draft note. The label is the same. The systems underneath are different, and they fail in different ways.
The best recorded evidence about use comes from physicians, not from front-desk, billing or management staff. The American Medical Association's 2026 Physician Survey on Augmented Intelligence was fielded from January 15 to February 2, 2026, and 1,692 U.S. physicians took part. On the adoption question, answered by 1,342 physicians, 72% said they had incorporated one or more AI use cases, 9% were uncertain which AI use cases their practice offers, and 19% reported none. The AMA's headline that more than 80% of physicians use AI adds the uncertain 9% to the 72%. The AMA describes that combined 81% as "any awareness or use of AI", compared with 66% in 2024.
Read those numbers with their limits:
- The survey is self-reported by volunteer respondents, and the report gives no margin of error or weighting. The percentages are not precise national estimates.
- The 2026 wave counted qualified partial responses and asked about 17 use cases instead of 15, so the AMA itself treats comparisons with earlier years as approximate.
- It surveyed physicians. No survey of practice managers, schedulers, coders or medical assistants is recorded for this course, so nothing here tells you how many of them use AI.
Even read carefully, the direction is plain. AI is already inside the practice, often in features nobody chose one at a time. The first skill is knowing what kind of system you are looking at, because that decides how you check it.
A question to sit with
Your reminder system sends a text two days before every appointment, your claim software flags a line it expects to be denied, and a portal feature drafts a reply to a patient asking when the lab opens. All three are called AI. Which one could be wrong in a way you would not notice just by reading what it produced?
What goes wrong
Treating every "AI" label as one thing. A practice writes one instruction, "double-check the AI", and applies it the same way to a fixed reminder rule and to a drafted portal reply. The rule gets re-checked every day for no benefit, and the draft gets a glance it does not deserve.
Checking a generated draft the way you check a rule. A rule that was tested once behaves the same way the next time it meets the same facts. A generative system writes new text every time it runs. Knowing that yesterday's draft was accurate tells you very little about today's.
Reading a flag or a score as a decision. A feature that learns from past data may mark a claim as likely to be denied, or an appointment as likely to be missed. The mark is an estimate built from old patterns. It is weakest exactly where the case is unusual, and it is not a finding about this patient or this claim.
Assuming fluent wording means understanding. A drafted reply can be warm, tidy and wrong. Fluency is a property of the sentences, not of the facts behind them.
Letting the product decide who owns the result. No product category moves responsibility. The person who would have owned the work without the tool still owns it with the tool, and for anything clinical that person is a licensed clinician.
The practical answer
Sort every AI feature you meet into one of three families before you decide how to check it. Many products combine families, so sort the feature, not the product.
| Rules-based automation | Machine learning | Generative AI | |
|---|---|---|---|
| How it works | A person writes explicit if-then logic | The system learns patterns from past examples | A model trained on very large amounts of text produces new text, summaries or tables |
| Typical practice use | Send a reminder 48 hours before each visit unless the patient has opted out of texts | A score that estimates which claim lines may be denied, learned from past claims | Drafting a portal reply, summarising a payer policy you supply, drafting a visit note from a recording |
| How it fails | Consistently: the rule is wrong, or a case falls outside it | Quietly on unusual cases; drifts when the practice or payers change | Plausibly: confident, well-formed content that is false or incomplete |
| What a wrong output looks like | The same wrong result every time | A reasonable-looking flag, often with a score | A fluent, finished-looking answer |
| How you check it | Test the rule, review exceptions, control changes to it | Sample the flags, look hardest at unusual items, watch for drift | Verify every fact, date, name and figure against a source you hold |
Three questions turn the table into a habit.
- Which family produced this output? If you cannot tell, ask the vendor or the person who manages the system, and write the answer down.
- How does that family fail? Match the check to the failure. A rule needs a test. A score needs sampling. A generated draft needs line-by-line verification.
- Who owns the result? That depends on the work, not the tool. This programme uses four owners throughout, and Module 04 explains them in depth:
- the licensed treating clinician, acting within scope, for anything clinical: diagnosis, treatment, orders, prescriptions, the content of the medical record, and patient communication about clinical information;
- the clinician whose documentation supports the claim and the coding professional who assigns the codes, for diagnosis and procedure codes;
- the practice's privacy and security officials, with the practice owner responsible for the compliance programme as a whole, for decisions about patient data and safeguards;
- the practice manager or named administrative owner for schedules, reminders, forms, non-clinical letters and operational reports.
Apply that idea one feature at a time. For each AI feature in your work, write one line: what the system does, what a person does, and who decides.
| Feature | System does | Person does | Who decides |
|---|---|---|---|
| Appointment reminder rule | Sends a text to patients who have not opted out | Tests the rule when it changes; reviews the opt-out exceptions list | Practice manager |
| Denial-risk score | Marks claim lines with an estimate | Works the flagged lines and samples the unflagged ones | Billing lead, reporting to the practice manager |
| Drafted portal reply | Produces a draft | Verifies every fact, removes anything clinical for a clinician to answer, edits, sends | Front-desk lead for administrative replies; the licensed clinician for anything clinical |
One more reason to sort carefully: the same screen can show outputs from different families side by side. A claim screen may show a rule-based edit ("referral number missing") next to a learned denial score next to a generated explanation of the denial risk. Each needs its own check, even though they arrive together.
Remember this
- "AI" covers three families (rules-based automation, machine learning and generative AI) that fail in different ways.
- Match the check to the failure: test rules, sample scores and flags, verify generated content line by line.
- A score is the system's own estimate from past patterns, not a decision about this patient or claim.
- A survey figure means nothing without its sponsor, sample and date; the AMA's 72% and 81% are volunteer physician responses.
- The tool never changes who owns the work, and anything clinical belongs to a licensed clinician.
Medical practice example
Cedar Hollow Family Medicine is a fictional four-clinician practice. Its practice manager, Dana Whitfield, is asked by the owner to list every feature described as AI that touches the front desk and billing, and to say how each is checked. She finds three. The first is the appointment reminder rule. It is supposed to skip any patient who has opted out of texts. Dana tests it on 40 fictional test patients, 6 of whom are marked as opted out, and finds that 3 of the 6 would still receive a text: the rule reads an old preference field that the practice stopped updating last year. The rule is rules-based automation, so it fails the same way every time. She has the vendor point it to the current field, re-runs the 40-patient test, and gets zero texts to opted-out patients. The second is a denial-risk score in the claim software. It learned from past claims and marked a claim for a fictional patient as low risk, even though the referral number for a new payer was missing. The claim looked like thousands of earlier ones, so the score was reassuring and wrong. Dana asks the billing lead to work every claim for that payer by hand until the pattern is understood. The third is a portal feature that drafts replies. A patient asked when blood can be drawn, and the draft said the lab is open "Saturdays from 8 to noon." Cedar Hollow has no Saturday hours. Nothing in the practice record said so; the sentence was generated because it sounded right. The front-desk lead checks the posted hours list, corrects the reply and sends it. Dana's one-page note lists each feature, its family, its check and its owner.
Responsible-use warning
Do not let a reassuring score or a polished draft shorten your review. A score is the system's estimate from past patterns and is least reliable on the unusual case that matters most. A fluent generated reply can contain an opening time, a name or a date that nobody supplied and no record supports. Neither is evidence. Before you act on an AI output, name the kind of system that produced it and apply the check that matches its failure: test the rule, sample the score, or verify every fact in the draft.
Human-review requirement
The owner of the work decides whether each AI feature may be used and how it is checked. The practice manager or named administrative owner owns reminder rules, schedules, forms and operational reports. The billing lead, under the practice manager, owns how denial flags are worked, while codes stay with the documenting clinician and the coder. The licensed treating clinician owns any reply or document with clinical content. The reviewer confirms each feature's family was identified, the check matched its failure, and no output was accepted on the strength of a score or fluent wording alone.
Practical exercise
Map the AI features in one workflow you know, using fictional or anonymised examples only. Do not paste patient information anywhere for this exercise.
- Choose one recurring task, such as appointment reminders for one clinic day, a batch of claims for one payer, or a morning of portal messages.
- List every feature in the tools you use for that task that is described as AI, smart, automatic, predictive or suggested.
- For each feature, decide its family: rules-based automation, machine learning or generative AI. If you cannot tell, write "unknown" and the name of the person you will ask.
- For each feature, write in one sentence how that family fails.
- For each feature, write the check that matches the failure: a rule test, a sample of flags, or line-by-line verification of generated content.
- Complete a four-column table: feature, what the system does, what a person does, who decides.
- Name the owner using this lesson's four owners: licensed treating clinician, documenting clinician and coder, privacy and security officials with the practice owner, or practice manager.
- Circle any feature whose current check does not match its family, and write one change you would propose to your manager.
- Keep the table with your workbook notes; you will reuse it when you build a risk matrix later in this Level.
Knowledge check
-
Why should an appointment reminder rule and a drafted portal reply not be checked the same way?
Answer guidance: The reminder rule is explicit logic that fails consistently, so it is tested and its exceptions reviewed. A generative draft is new text each time and fails plausibly, so every fact, date and name in each draft must be verified against a source the practice holds.
-
What does a denial-risk score on a claim line tell you, and what does it not tell you?
Answer guidance: It is the system's estimate, learned from past claims, of how likely the line is to be denied. It is not a finding about this claim, and it is least reliable on unusual claims, such as a new payer, that do not resemble past data.
-
Why is the AMA's "more than 80%" figure not the share of physicians who use AI tools?
Answer guidance: The combined 81% adds the 72% who incorporated at least one AI use case to the 9% who were uncertain which use cases their practice offers. The AMA describes it as any awareness or use of AI, and the survey is a self-reported volunteer sample with no margin of error.
-
According to NIST AI RMF 1.0 Appendix C, what must be true of human roles, and does that bind a practice?
Answer guidance: Human roles in decision making and oversight need to be clearly defined and differentiated, and configurations can range from fully autonomous to fully manual. The framework is voluntary; it binds no practice, and it is being revised as of September 2026.
Sources and further reading
- American Medical Association, Center for Digital Health and AI — 2026 Physician Survey on Augmented Intelligence (2026-03-01). 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.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (2024-07-26). Accessed 2026-09-15.