The AI Scribe records a consultation, transcribes it live, and then uses that transcript to help you fill in the chart it sits in — the prose under your own headings, your structured blocks, and your logs. Nothing it generates reaches a clinical record without you accepting it first.
The scribe is a smart block. Insert it the way you insert any block: type / in the chart editor and pick AI Scribe.
Recording
- Open the chart you want to document into
- Insert an AI Scribe block (
/→ AI Scribe) - Press Record and consent-check as your clinic requires
- Speech appears in the block as it is recognised
Recording follows you around the app
The recorder does not live inside the block — it runs above the whole application. You can leave the chart mid-consultation to look at the last visit, check inventory, or book a follow-up, and the recording keeps going. A floating bar appears on every other page with the two controls you need: Stop, and Back to the chart.
When you return, the block picks its own recording back up. If a take finished while you were on another page, it is folded into the block the moment you open it.
Takes add up, they don’t replace
A second recording extends the transcript rather than overwriting it. An encounter is often captured in more than one pass — an initial history, then an examination — and each take is appended in the order it happened.
One recording at a time
There is one microphone, so there is one recording. If you try to record from a second block while one is live, Zdrovia refuses by name and links you to the block that is currently recording, rather than silently taking the device.
:::info Audio is session-only The audio recording is held in your browser for playback while the block is open. It is never uploaded and is gone on reload. The transcript itself is saved with the chart — only the audio is temporary. Zdrovia warns you before you leave the page with a recording in flight. :::
Playback and correction
While the audio is still in the session, the transcript and the recording are linked to each other:
| Action | What happens |
|---|---|
| Click a word | Selects it and seeks the audio to the moment it was spoken |
| Play the audio | A highlight walks along the words as they are said |
| Click the selected word | Opens it for correction — hearing a word never costs you an accidental edit |
Correcting a word keeps its timing, so the audio still seeks correctly afterwards. To insert or delete words, switch the segment to plain-text editing; word timings are dropped only if the text actually changes.
Each take has its own small player above its own words, because each recording has its own clock.
After a reload the transcript is still there and still readable — it simply renders as plain text with no audio behind it.
Speakers
When speaker separation is on, Zdrovia groups the transcript into a paragraph per turn, so it reads like a conversation.
Speaker separation can tell voices apart; it cannot know which voice is the clinician. Zdrovia therefore does not guess — each speaker gets a chip you can rename. Naming them is worth doing: attribution is written into the stored transcript as Name: what they said, and a model reading “Patient: the knee gives way on stairs” can place it correctly, where an unattributed line has to be guessed at.
Names are stored per take. Speaker separation numbers voices independently for each recording, so “Speaker 1” in your second take is not promised to be the same person as in your first.
Fill chart
Fill chart reads the transcript against the chart the block sits in and proposes content for it. This is how the note gets written: prose proposed into your own headings, values proposed into your own blocks.
- Press Fill chart
- Zdrovia collects every fillable target in the chart, in document order
- A review dialog opens listing each proposal
- Tick the ones you want, untick the ones you don’t
- Apply — everything accepted is written in one step, so a single undo reverses it
What can be filled
| Target | What gets proposed |
|---|---|
| Headings | The paragraphs under one of your headings |
| Labelled lines | A line you started with its own label, like Symptoms: — only what follows the label is written |
| Smart blocks | The fields the block declares as spoken-aloud fields |
| Custom blocks | Every field your clinic authored on the block |
| Log records | An existing reading in a log, or new readings appended to it |
| New sections | Headings the chart does not have yet — off by default, appended at the end |
The rules that protect the record
- A human accepts every item. Nothing generated is written without being ticked.
- An empty section is ticked by default; a section that already has text is not. Accepting the defaults never overwrites something you typed — where text exists you get a diff and an unticked row.
- Writes merge, they don’t replace. Filling two fields on a block leaves the rest of it alone.
- Values that don’t fit are dropped, not forced. Prose where a number belongs, a dropdown value your clinic never authored, a malformed date, a field the block doesn’t have, or a proposal identical to what is already there — all discarded rather than written.
- Appended readings can’t carry an invented identity. Each proposed record is its own checkbox, so an invented fourth reading is dropped without losing the three real ones.
- Scored questionnaires are excluded. Their totals and severity bands are computed from item answers. A proposed total would be a score no scoring rule produced.
Guidance per heading
Zdrovia recognises the common clinical headings — the SOAP and DAP sections plus history, examination, impression, treatment, medications, advice, follow-up, referral, progress, goals, consent and allergies — and steers what belongs under each. Headings are matched exactly (ignoring numbering, casing and trailing punctuation). A heading Zdrovia doesn’t recognise works from its own name, which is safer than guessing: guidance attached to the wrong heading actively directs the wrong content into a clinical record.
The summary
Separate from the note, the block generates a summary: what you read to remember the visit at a glance — at handover, before the next appointment, or at the top of a referral.
Every generated summary starts unattested, and regenerating clears attestation again. The marker appears in the editor, in the read-only view, and — the case that matters — in the printed copy, because that is the one that ends up in a referral. Editing the summary text counts as reviewing it.
Settings
Open the settings popover on the block to configure:
| Setting | What it does |
|---|---|
| Instructions | House guidance — register, emphasis, what your clinic wants captured |
| Allow new sections | Lets Fill chart propose headings the chart doesn’t have (off by default) |
| Speaker separation | Whether voices are told apart and attributed |
| Live caption duration | How long live captions run |
Settings are stored per block, not per clinic. A chart template that ships a configured AI Scribe block carries those settings into every chart made from it — which is how you set a house default today.
Your instructions can shape how the note reads. They cannot license inventing a finding: they sit below the rules, and the rules win.
What the scribe will not do
- It will not author itself. The AI that generates chart templates cannot create an AI Scribe block. A scribe block asserts that an encounter was recorded; a generated one would assert that about a consultation that never happened.
- It will not treat the transcript as instructions. Anyone in the room can say something shaped like a command to a model. The transcript is labelled as recorded speech, and text inside it that appears to address the model is treated as something a person said.
- It will not regenerate a whole chart. Whole-chart regeneration has no per-item review and would risk silently dropping a block it failed to understand. Chart-wide generation stays where it belongs, in chart templates.
Current limits
- English only for now. The block reports the language detected but works from an English model.
- Audio does not survive a reload, so neither does word-level playback. The transcript does.
- Regimen blocks are not fillable yet.
- Appended log records all take the time of applying, even where the transcript states a time for each reading.
- A filled value looks like a typed one afterwards — no provenance marker is recorded.
- Speaker names are per take and are not carried forward between takes.
- Nothing checks a summary against a later-edited transcript. If you edit the transcript after generating, regenerate the summary.
Tips
- Name the speakers once the first take finishes — it measurably improves what Fill chart puts under Subjective.
- Record in passes. History, then examination, then plan — each take extends the same transcript, and you get to review between them.
- Put your headings in the chart template. Fill chart writes into the structure it finds, so a template with S/O/A/P headings gets you a structured note without asking for one.
- Use labelled lines (
Symptoms:,Allergies:) for fields you want filled but don’t want a whole heading for.