What happens between the patient finishing speaking and the clinician reading a structured, sourced note.
This page explains iClerk by walking through a single real consultation — Mrs Patel, who arrives at your urgent care clinic with chest pain. By the end of the page you'll have seen exactly what happens between her finishing speaking and you reading the structured note, and you'll have a one-line answer to every "but how do I know it's safe?" question.
Written for two readers: you, deciding whether iClerk fits how you work; and your colleagues and the people who hold the budget. Every step ends with one sentence you can quote when you talk to them.
67-year-old woman. Walks into your urgent care clinic with central crushing chest pain that started two hours ago, radiating to her left arm. She's pale, sweaty, and a little short of breath.
Past medical history: type 2 diabetes, hypertension, hyperlipidaemia. On simvastatin and metformin. Smokes 15 cigarettes a day. On examination: BP 152/95, HR 98, sats 96% on air, chest clear, heart sounds normal. ECG shows lateral T-wave changes.
You see her, examine her, and dictate into iClerk while you go. You don't type a single word. iClerk's job is to turn what you said into a note you'd be happy to put in her record.
There are only ever two possible sources for any line iClerk writes: a document your practice has uploaded, or the underlying AI's own general medical training. Every line in the note tells you plainly which one it came from — there's no internet search, no third hidden source. For Mrs Patel today, the practice library contains:
When iClerk has a guideline-backed answer for Mrs Patel, it cites the specific uploaded document by name. When neither document fits, the line is plainly labelled [General Knowledge] — that's iClerk telling you "this is from the AI's general training, not from a document you signed off on, so apply your own judgement". The labelling is the safety feature: you always know which kind of source you're reading.
Now let's walk through the steps iClerk takes to get from one to the other.
Each step is tagged so you can tell at a glance what kind of work it's doing: Hallucination safeguard — an action against the AI making something up; Mode — a choice about what iClerk does at all; Feature — a value-add the AI couldn't produce on its own.
Before iClerk writes a word, a short AI screen reads the consultation and asks itself, using its own clinical judgement: could this presentation be any immediately life-threatening "must-not-miss" condition? For every one it identifies, the matching protocol from your library is pulled into the note — so the time-critical guidance is grounded and cited, never left to the luck of ordinary retrieval ranking and never reduced to the AI's unsourced general knowledge.
Her central crushing chest pain radiating to the left arm, with sweating, breathlessness and lateral T-wave changes, is flagged as possible acute coronary syndrome — and your ACS protocol is pulled in, so the note's plan is grounded in your guideline, not guessed.
Why it matters. Ordinary retrieval ranks documents by overall similarity to the whole dictation. A critical problem that is only a secondary thread — say, a fever mentioned in passing by a patient on chemotherapy, in a consultation otherwise about a fall — can be crowded out. This screen exists so the dangerous diagnosis is surfaced and grounded even when another problem dominates. Cost: one short AI call per consultation (a couple of seconds, a fraction of a cent); it deliberately errs toward raising a possibility rather than missing it.
Before iClerk drafts a single line of the note, a separate AI step reads only what you dictated — no documents, no protocols, no general medical training — and pulls out a structured list of facts about the patient, with the verbatim words you spoke attached as evidence for each one.
This list is the scaffold the note-writer is anchored to. The writer is allowed to use your library documents to give context and to choose differentials, but the patient-specific facts that end up in PC, History, PMH, Drug History, Social, Family History, Examination and Investigations must trace back to a quote from this list. If a fact has no quote, it doesn't make it into the note.
This step exists because the most insidious failure mode of an AI scribe is not making up a guideline (Steps 4–5 below catch that) — it's quietly inventing a patient fact. "Patient denies chest pain" when you never asked; "no shortness of breath" when you simply moved on; "no family history of cancer" because the AI saw a similar case during training. Anchoring the writer to a list of facts with verbatim quotes makes that class of failure visible and removable.
When it runs. This step is part of the Agentic Retrieval mode (an admin opt-in, see Step 4 below). With Agentic Retrieval off — the current default — the writer runs against the legacy retrieval pipeline and the fact-extraction pass does not fire. Cost is one extra short AI step (a couple of seconds and a fraction of a cent) when active.
iClerk reads what you dictated, looks at the relevant documents your practice has uploaded, and writes the structured note. By default, what you get is a digital scribe — Presenting Complaint, History, Examination, Investigations, and that's where it stops. This is regulated as a Class 1 medical device: record-keeping, no clinical advice. Lowest regulatory bar, suits a fresh tenant or a small clinic.
If your administrator turns on Decision Support, two extra sections appear automatically:
Now iClerk is suggesting things, not just recording them. That moves it from the lowest-risk medical-device class into a higher one — the exact paperwork burden depends on the country, but every major regulator (US FDA, EU MDR, Health Canada, TGA, PMDA and so on) treats it as a higher-risk tier. Your institution signs off a clinical-safety case; your insurer is informed; your training plan changes. The same step is being done by iClerk — only the prompt is different — but the regulatory weight is heavier, and that should be a deliberate institutional decision, not a tick-box.
Cost note. Flipping Decision Support on or off doesn't add any extra AI work — same one AI step, just a different instruction. The cost is regulatory burden, not the bill.
Step 1 stops the writer drifting away from the transcript while it's drafting. This step is the belt-and-braces underneath it: after the note has been written, a different AI reads every history section line by line and asks a single question — is each fact-bearing line supported by either the transcript or the facts Step 1 extracted from it?
Anything unsupported is stripped out and replaced with Not documented in transcript. The note never reaches the clinician with a fact in it that the clinician didn't actually say.
This is a different check from the citation checks below. The citation checks (Steps 4, 5, 7) verify that the source of clinical reasoning is a real document in your library. This step verifies that the description of the patient is rooted in the consultation that just happened. They're checking different things, both important.
The judge runs on the history sections (PC, HPC, PMH, Drug History, Allergies, Social History, Family History, Examination, Investigations) — the parts of the note that describe the patient. It does not run on Differentials or Initial Plan, because those are reasoning sections that are allowed to step beyond the transcript (and have their own citation checks in Steps 4–7 below).
Honest caveat — when the judge fails open. If the AI running this check goes down, iClerk does not refuse the note. It logs the failure and shows you the note as the writer produced it, with the upstream checks (Step 1's transcript anchor, plus the citation checks below) still in force. Hard-failing on a transient outage would be a worse clinician experience than a single missed grounding check, given Step 1 has already done the heavy lifting. This trade-off is reported in the audit log so a safety officer can see when it happens.
When it runs. This step is part of the Agentic Retrieval mode (an admin opt-in). It pairs with Step 1's fact extraction — they run as a set or not at all. With Agentic Retrieval off, the transcript-grounding judge does not fire and the writer's draft is shown as-is, with the citation safeguards (Steps 4–5, and Step 7 if enabled) still in force. Cost when active is one extra AI step per consultation (a couple of seconds, a few cents).
This is the single most important safety feature for the "but couldn't the AI just make things up?" question. The answer, with this on, is structurally no — and here's why.
When iClerk writes the bullet "Aspirin 300 mg PO stat.", it is told not to write the source name itself. A separate, narrower step then chooses the source — but its only options are the documents that were actually pulled up for Mrs Patel today. For her, that's Adult_and_Paediatric_Differential_Diagnoses, ACS_Acute_Management, or "General Knowledge" if neither fits. Nothing else.
ACS_Acute_Management · Adult_and_Paediatric_Differential_Diagnoses · [General Knowledge]iClerk is literally not given the keyboard to type a fabricated guideline name. This is different from "we trained the AI not to lie" — this is "we don't give the AI the option".
An alternative shape — agentic retrieval (admin opt-in). Your administrator can switch iClerk into an experimental mode where the writer model does its own retrieval: instead of a pre-narrowed shortlist, it's handed a tool called search_kb and a manifest of the documents it's allowed to search. It then decides — per clinical theme — which documents to query, runs the searches itself, and cites the chunks it actually pulled back. The closed-list guarantee still holds: any citation that doesn't trace to a chunk the model fetched this turn is collapsed to [General Knowledge] before the note is shown. The mode is off by default and currently sits behind an admin-only flag.
Cost note. Turning this on adds two small extra AI steps to each consultation — a few extra seconds of latency and a few extra pence per note. Most clinics decide that's worth it. If you want to test whether the simpler safety net (Step 5 below) is enough on its own, you can switch this off for a trial week, sample some notes, and see if any fabricated guideline names slip through.
On top of Step 4, iClerk runs a final check on every source name in the finished note. It compares each name, letter by letter, against the list of documents in your practice's knowledge base. If a name doesn't match, it's automatically rewritten as "General Knowledge" so you know that bullet isn't backed by a guideline you signed off on.
This check runs every single time, on every single note. It doesn't use any AI — it's just a name comparison, the kind of rule a regulator can read in plain English and verify works. It's the always-on belt-and-braces underneath Step 4.
Clickable citations. Once a citation has passed Steps 4 and 5, the [Source: X] tag in the finished note is a link. Click it and iClerk opens the actual document the citation points to — scrolled directly to the passage the AI used, with the matching sentences highlighted. You don't have to take iClerk's word for the citation; you can read the source. This is the same view your administrator uses to manage the knowledge base, surfaced inline at the point of clinical reading.
This step is a feature, not a hallucination safeguard. The other numbered steps in this walkthrough defend against the AI inventing facts or citations; this one adds something the AI couldn't produce on its own — the institution's accumulated clinical experience. Including it here in time order because that's when it runs, not because it belongs in the safeguard cascade.
Every experienced clinician carries pattern-recognition that isn't in any guideline. "Elderly diabetic women can have silent MIs." "Anyone over 50 with new back pain — exclude AAA." "On rituximab + hypertension + neuro symptoms — think PRES." These are pearls.
You write your pearls into iClerk once. Each one has a trigger ("elderly + diabetic + atypical chest pain") and the advice ("low threshold for repeat troponin, don't be falsely reassured by a single normal value"). They're personal — your colleague's pearls are different, and yours improve as you accumulate experience.
If your administrator enables Pearl Relevance, iClerk runs an extra AI step after the note is written. It reads the structured note, looks at every pearl you've saved, and surfaces any that match. Surfaced pearls appear at the bottom of the note, plainly marked as your own teaching — never confused with the AI's reasoning.
Cost note. One extra AI step per consultation — a couple of seconds and a few pence. Worth turning on as soon as a few clinicians have started writing pearls; not worth turning on until then.
Steps 4 and 5 stop iClerk from citing a document that doesn't exist. This step catches a subtler problem: citing a real document, but for the wrong reason. For example, a chunk from a chest-pain document gets attached to a heart-failure bullet because both mention "ECG" and "shortness of breath".
With this on, iClerk runs a second AI that re-reads each recommendation alongside the document chunk it was attributed to and asks one question: does this chunk actually support this bullet? If the second AI says no, the bullet is marked as off-target — the source name stays visible, but the clinician sees clearly that the system doesn't fully back the match.
Why it ships off by default. In our testing, Steps 4 and 5 already catch the citations that matter for routine clinical cases. Turning the second AI on adds latency and cost without changing what most clinicians see. Worth turning on if:
For every recommendation in the differentials and plan sections, iClerk shows a small four-cell coloured indicator next to the citation:
It's pure visual — no extra AI work — and it lets the clinician see at a glance whether to trust each line, without having to click through to the source.
Fair question — and the answer is that there are two different kinds of certainty at play, and Steps 4–5 only guarantee one of them:
So the confidence indicator isn't admitting a weakness in Steps 4 or 5 — it's giving you a second piece of information Steps 4 and 5 can't give you. The same transparency principle as the "General Knowledge" label: tell the clinician everything we know about each line, including how confident we are. We'd rather show "partial" honestly than show "strong" and be wrong.
The four tiers can only be distinguished when Step 7 is on, because Step 7 is the AI that judges content-fit. With Step 7 off, the indicator effectively collapses to two states:
So if you want the indicator to discriminate honestly between strong and partial citations, turn Step 7 on. With Step 7 off, treat the indicator as "cited / not cited" and rely on Steps 4–5 for the source-name guarantee. Both setups are valid; the question is just how much detail you want surfaced. (A cleaner fix would be to label cited-bullets as "Verified name" rather than "Strong" when Step 7 is off — that's a small UI change worth doing if you adopt this product story.)
iClerk's running cost per consultation depends on how many of the optional safety layers your administrator has turned on. Three reference points:
All three setups produce a clinically usable consultation note. The difference is how many independent safety checks sit between the AI's first draft and what the clinician sees.
iClerk is a digital scribe with optional clinical decision support. Configured as a scribe, it sits in the lowest-risk medical-device class in every major regulatory framework — the lightest paperwork bar, an immediate documentation productivity gain. With decision support switched on, the same product moves into the higher-risk tier, the deploying institution completes the clinical-safety process that applies in its country, and differentials + initial plans appear in every note.
What makes iClerk safe to use clinically: every recommendation is plainly tagged with its source. Either the specific document the institution uploaded — by name — or the label "General Knowledge" when the suggestion is coming from the AI's general medical training. There are only ever two possible sources, and the clinician always sees which one. Nothing hidden. If a clinical decision is ever questioned later, the system points to exactly where that suggestion came from. And the AI is structurally prevented from making up a guideline name — it physically can't type one the institution didn't give it.
For patient facts — the description of this patient, today — iClerk's Agentic Retrieval mode adds two more safeguards: a separate AI step that reads only the dictation and extracts a list of facts with the verbatim words attached before the note is written, and a second AI step that, after the note is written, strips any line of the patient history that isn't supported by the dictation and replaces it with "Not documented in transcript". The "AI quietly added a symptom no-one asked about" failure mode is closed off at the source when these two are running. Agentic Retrieval is an admin opt-in today — off in the default configuration.
On top of that, clinicians can save their teaching points and have iClerk surface them automatically at the right moment. Institutional knowledge accumulates over time, instead of leaving when staff do.
Cost is a few cents per consultation. Latency is a few seconds. The risk being managed is "AI hallucinating a guideline" — designed out at the architecture level, not just at the prompt level.