"AI medical scribe" sounds like magic, and the marketing often leans into that. The reality is more reassuring and more boring: it's a pipeline of well-understood steps, each of which a doctor can inspect. Understanding that pipeline is the difference between trusting the tool blindly and using it well — knowing where it's strong, where it drifts, and where your judgment is non-negotiable.
This is a plain-language walkthrough of how an ambient scribe turns a spoken consultation into a structured note. We'll use Shifaa AI as the worked example, because being specific about the moving parts is more honest than hand-waving about "the AI." The same four stages apply to most products in this category.
The four-stage pipeline
Strip away the branding and almost every AI scribe does the same four things in sequence. The first three are automated; the fourth is you, and it's the one that matters most.
1. Audio capture in the room
It starts with a microphone — usually the one already in your phone. There are two capture modes worth distinguishing. Ambient mode listens passively to the natural back-and-forth of the consultation: your questions, the patient's answers, the whole conversation. Dictation mode is the older style, where you speak a deliberate summary to the device after or between patients. Ambient is less intrusive but depends on room acoustics; dictation is more controlled but adds a step. Either way, this stage just captures sound — no understanding has happened yet.
2. Speech-to-text (ASR)
The audio is then passed to an automatic speech recognition (ASR) model, which converts speech into a raw text transcript. In Shifaa, that model is OpenAI's Whisper, which is multilingual and copes reasonably with accents and code-switching. But this is exactly where accuracy lives or dies. A clear, close microphone in a quiet room produces a clean transcript; a noisy waiting area, overlapping speakers, or an unusual drug name can introduce errors. Whatever Whisper mishears here flows downstream — one reason the final review step is mandatory, not optional.
3. An LLM structures it into a SOAP note
A raw transcript is not a clinical note. The next stage hands that transcript to a large language model — Anthropic's Claude in Shifaa's case — which reorganises the free-flowing conversation into the familiar Subjective, Objective, Assessment, Plan structure. It pulls the history of presenting complaint into Subjective, examination findings into Objective, and so on. Crucially, the model is reorganising what was actually said; it is not supposed to invent clinical content. Both Whisper and Claude are external sub-processors, which is worth knowing about any scribe before you feed it patient audio.
4. The clinician reviews, edits and signs
This is the load-bearing step, and no honest description of an AI scribe should rush past it. The draft note is exactly that — a draft. You read it, correct anything the transcription or structuring got wrong, add the clinical reasoning that was in your head but never spoken aloud, and only then sign it. The AI does not enter anything into the record on its own authority. You are the author; the tool just gives you a head start on the typing.
The one rule that matters
An AI scribe drafts; the clinician decides. Accuracy varies with audio quality and language, the model can mishear or misorganise, and nothing reaches the patient's record until you have read, edited and signed it. Review is not a nicety — it is the step that makes the rest safe.
"Fills empty fields only" — and why it matters
Here's a design choice that sounds minor but isn't. Shifaa's scribe fills empty fields only: it populates the parts of the note you haven't written yet and never overwrites text you typed yourself. If you've already documented the examination by hand, the AI leaves that field untouched and only drafts the blanks. This matters because the most dangerous failure mode of any drafting tool is silently replacing something a human deliberately recorded — a model that can overwrite your words can quietly turn a correct note into a plausible-but-wrong one. "Empty fields only" makes the AI strictly additive: it can help, but it can't undo your documentation. Your text always wins.
Finally, set expectations honestly: a scribe is a documentation aid, not a clinician. It saves time on the administrative load around care — and that load is real. Sinsky and colleagues, writing in the Annals of Internal Medicine in 2016, found physicians spent roughly two hours on EHR and desk work for every hour of direct patient contact. But output quality tracks input: poor audio or a heavy mix of languages degrades the transcript, and structuring can occasionally misplace detail. To see how this fits a real workflow, the AI medical scribe feature page walks through it; and for where automation helps versus where a person still wins, this honest comparison of AI versus human scribes is a fair place to start.