AI in Daily Medical Practice — Module 2 of 6
Handout — self-paced study material
How this module works. Take the short Preview Quiz first — five minutes, ungraded. Then work through this handout (or the companion slide deck) at your own pace. Section C is a worked exercise rather than plain reading — do it actively, with your own answers, before checking the answer key. Finish with the Post-Test.
By the end of this module you will be able to:
Across the CME materials reviewed for this curriculum, one theme was more consistent than any other: physicians spend a disproportionate share of clinic time on charting rather than on patients. Exact figures vary by setting and shouldn't be taken as precise — what's consistent is the direction and the frustration behind it. That's the problem "ambient documentation" tools are built to address.
Ambient documentation (sometimes called an "AI scribe") listens to — or reads a transcript of — a clinical encounter and drafts a structured note automatically, so the physician can stay engaged with the patient instead of typing during the visit. The often-used phrase for this is "keyboard liberation": the goal isn't the technology itself, it's getting the physician's attention back on the person in the room.
This is deliberately the first practical skill in this course, ahead of diagnostic AI in Module 3, for two reasons. First, it is comparatively low-risk: a documentation error is usually easier to catch and correct than a missed diagnosis. Second, it produces a fast, visible win — most physicians feel the time saved within a single week — which is exactly the kind of low-stakes success that builds trust before higher-stakes applications.
The basic pipeline is short: audio (or an existing transcript) goes in, speech is converted to text if needed, and a language model — the same underlying technology from Module 1 — restructures that free-flowing conversation into a formatted clinical note (often a SOAP-style draft). The draft is then handed to you for review before it becomes part of the record.
Every failure mode from Module 1 Section B applies here directly, because the same underlying technology is doing the work:
| Failure mode | What it looks like in a drafted note |
|---|---|
| Hallucination | A detail appears in the note that was never actually said in the encounter — a symptom, a value, a plan. |
| Misattribution | Something the patient said gets written as the physician's assessment, or vice versa. |
| Silent omission | The model judges a detail "unimportant" and drops it — this is especially dangerous for things like allergies, mentioned once in passing. |
| Plausible-value filling | When information was genuinely missing from the encounter, the model may still generate a plausible-sounding value rather than leaving a blank. |
| Mishearing | A spoken dosage or number is transcribed incorrectly — "20 mg" heard as "200 mg" — and then carried through with full confidence. |
None of this makes ambient documentation unsafe to use. It makes Section C — verification — non-negotiable.
Below is a short, fictional follow-up visit. Read the transcript excerpt, then read the AI-drafted note that followed it. Before reading further, make your own list of what's wrong — then compare it to the answer key.
Transcript excerpt — follow-up visit, known hypertension
AI-drafted note
Subjective: Patient reports improvement in symptoms since medication adjustment. Taking lisinopril 20 mg daily as prescribed. Denies headache, cough, and ankle swelling.
Objective: BP 128/82 today, improved from prior visit.
Assessment: Hypertension, well-controlled on current regimen.
Plan: Continue current dose. Follow up in 2 months. No known drug allergies.
Answer key — four errors were planted
If you caught three or four of these on your own pass, you're already applying the right instinct. If you caught fewer, that's precisely why this exercise exists — re-read the transcript against the draft once more before moving on.
A verification checklist you can actually use
There's a second, less obvious reason careful verification matters: many documentation tools use clinician-reviewed notes to improve future versions of the model, subject to the vendor's specific data and consent policies. A sloppy or uncorrected note doesn't just sit wrong in one chart — depending on the product, it can become a small piece of the training signal for the next version of the tool, for every clinician using it. Verification isn't just protecting this patient's record; it's protecting the quality of the tool itself over time.
A self-directed worksheet, same format as Module 1's Section D. Work through it on your own — your answers here will matter directly when you're actually selecting a tool, not just studying one.
1. Does the tool need to function with intermittent or no internet connection during a consultation, or is reliable connectivity a safe assumption in your setting?
2. What language(s) do your encounters actually happen in, and does the tool draft notes accurately in those languages — not just transcribe them?
3. Where is the audio or transcript data stored, for how long, and who besides you can access it?
4. What does the tool actually cost on an ongoing basis, and is that sustainable for your practice specifically — not just affordable once?
5. Does the tool need to integrate with an existing record system, or can it work as a standalone step in your current workflow?
| Layer | What this module put there |
|---|---|
| KNOW | What ambient documentation tools do and why they're the starting use case; the basic capture-to-draft pipeline; the five documentation-specific failure modes and how they map to Module 1's core failure modes; the documentation feedback loop. |
| THINK | A trained verification habit — practiced directly on a worked example — plus a reusable checklist for every AI-drafted note you review from here on. |
| FIND | A first pass at judging documentation-tool fit for your own setting — connectivity, language, data storage, cost, integration. |