An AI scribe in
the emergency department

An emergency department note is read by more strangers, sooner, than most clinical documents. The admitting team opens it within the hour to decide what happens overnight. The GP reads it days later as the only account of what happened to their patient. If the patient re-presents at three in the morning, the next emergency clinician reads it as the entire baseline for a presentation that has already declared itself once. It is read again years later in coronial and civil proceedings, when nobody involved remembers the shift.

The conditions it is written in work against that. The presentation is undifferentiated, the history is incomplete, the clinician is holding several patients at different stages, and the interruptions arrive without warning. What gets compressed out under that pressure is predictable, and it is what later readers care about most: the differential that was considered and excluded, the reason a test was not ordered, and the specific advice given at discharge.

No documentation tool fixes crowding, access block or short staffing. What ambient capture changes is how much of the reasoning spoken aloud reaches the chart, and in emergency work that gap is wide: the differential a clinician talks through with the patient and the registrar is often the part that never gets written down.

Ambient scribes are built around a consultation that runs as one continuous conversation with one patient, bounded by a start and a stop. Emergency work breaks that assumption. The story arrives in fragments across hours: the ambulance crew's handover, the patient's own account at two in the morning, a family member arriving at three with the medication list, a reassessment after analgesia, and a consultant review at eight.

The unit of capture therefore has to be the patient rather than the session. A tool that produces one note per recording leaves a scatter of disconnected notes for one presentation, and the clinician will spend longer stitching them together than they would have spent typing. Before any trial, ask how many separate captures can attach to one patient, and how a clinician resumes an encounter hours after starting it.

The failure that matters most is what happens to a running capture when the clinician is pulled away mid-sentence. Left running in a pocket, it will record the resuscitation two bays over, a handover covering eight other patients, a phone call to a psychiatry registrar, and a family being told that someone has died. None of that belongs in a documentation tool, and some of it should never have been recorded at all.

• Start a capture, walk away for twenty minutes, and look at what the tool did with that time.
• Confirm whether capture pauses when the device locks or the app loses focus, and whether it stops on its own after a period of inactivity.
• Check that stopping takes one action that works with gloved hands and without unlocking the device.
• Ask what happens to audio captured in error, how quickly it is deleted, and whether the deletion is itself logged.
• Run the test on a night shift, when the department is thinnest and the interruptions are worst.

In most settings the note records what was found. In emergency medicine it is just as often a record of when, and the timing carries the clinical argument. Time last known well is what a stroke team works from when it considers reperfusion. The interval between a first and second troponin shows whether the chest pain pathway was actually followed. Time of arrival, time of first medical assessment, the time analgesia was given and the time antibiotics went in for suspected sepsis are discrete points a later reader uses to reconstruct the case. The Australasian Triage Scale sets a maximum waiting time for medical assessment against each category, so the clock on a presentation is already running before the treating clinician has met the patient.

Ambient capture records what people say, and clocks are rarely said aloud. The triage entry, the observation chart trend, the repeat ECG, the nursing reassessment after fluids and the exact minute the ambulance handed over all sit in the department's own systems, not in anything spoken in front of a microphone, so a draft built from audio alone is thin in precisely the place emergency notes get attacked. The timing skeleton belongs to the record system, the ambient draft carries the reasoning between those times, and the clinician reconciles the two before signing.

Ask a vendor directly whether a time appearing in a draft comes from the moment of capture or the moment of signing. A tool that stamps everything at signing will produce an entry saying a reassessment done at 0510 happened at 0730, which reads to anyone who was not there as a false statement about patient care rather than as a software defect.

The clinical claim is often that something is unchanged, or improved after treatment, and that only holds if each assessment is separately timed and attributed. A tool that merges three visits to the bedside into one smooth narrative removes the structure that made the record defensible. Give any ambient tool three short captures of the same patient at different times and see whether it produces three timed entries or one blended paragraph.

Consent to record is where emergency work departs most sharply from clinic. Many of the department's patients cannot give it at the moment they are seen, because they are intoxicated, delirious, post-ictal, in severe pain, sedated or unconscious. Some will have capacity again within hours and some will not regain it at all. Capacity is decision-specific, so a patient who cannot weigh a decision about admission may still understand a question about a tool that helps the doctor write notes, and that judgement sits with the clinician in front of them.

The legal basis for treating a patient who cannot consent does not extend to recording them. Urgent treatment without consent rests on the common law principle of necessity and on the emergency provisions in state and territory guardianship and medical treatment legislation, and it authorises what is necessary for the patient's care. A recording is neither treatment nor necessary to it, because the clinician can always document by hand. Implied consent does no work either, because a patient who cannot form a view cannot be taken to have agreed. Recording a private conversation also engages state and territory surveillance devices legislation, which is not uniform across the country.

A substitute decision-maker can consent where one is identified and their authority is established, which is realistic for a planned review and rarely realistic in the middle of an emergency assessment. The relative standing at the bedside at two in the morning is not automatically the person responsible. The position that holds up under pressure is to leave the tool off while capacity is impaired and revisit it if the patient clears.

• Default to not capturing while capacity is impaired, and document the encounter the ordinary way.
• Ask again if the patient clears during the stay, and capture the reassessment rather than trying to recover the first assessment.
• Where the patient is competent, ask before the first capture and record the outcome, including a refusal.
• Do not start a capture on a family member's agreement alone unless their authority has been established.
• Do not let a decision made at triage run for the whole stay, because whoever does the next assessment decides again.

The resus bay is the clearest exclusion. Several people speak at once in closed-loop calls, monitors alarm, and compressions and suction change the room's acoustics, so transcription quality collapses at exactly the moment when there is nobody able to consent. In most departments the record of a resuscitation is a timed structured form kept by a designated scribe, which is a different document from a narrative note and is not improved by an audio draft.

Other exclusions are less obvious and matter as much: forensic and sexual assault examinations, patients under police guard, assessments of a patient detained under state or territory mental health legislation, and conversations with a family after a death. In each of these the presence of a recording changes the encounter, and the recording becomes an artefact that can be sought later by people nobody in the room had in mind.

The mental health presentation is the exclusion most likely to be applied wrongly, because a patient's legal status can change during the stay. Someone who arrived voluntarily may be detained under the relevant state or territory legislation four hours later, and someone brought in involuntarily may settle and be assessed as competent by morning. What governs is the status at the time of the assessment, so the clinician doing the assessing has to check it rather than trust what the tracking board said at triage.

The department needs a written list of where the tool is not used, agreed before rollout, and a stop that anyone in the room can invoke. The practical test is whether a nurse who walks into a bay can say to stop recording, have it stop, and lose nothing from the clinical record because the clinician documents by hand instead. If stopping means unlocking a device and finding the right screen, it will not happen fast enough to matter.

A single emergency presentation is commonly touched by a triage nurse, a resident, a registrar, a consultant and a registrar from the admitting team. Each forms an impression, and the record has to show whose impression it was and when. Every practitioner in that chain carries record-keeping obligations under their own National Board's code of conduct, and the Medical Board's Good medical practice code expects records that are accurate, up to date, legible and written so that another practitioner can understand them. A record that cannot say which clinician made a given assessment, and when, fails that test however well it reads.

A draft assembled from audio captured across a shift will attribute everything to whoever happens to run the review and sign it, and nothing on the face of the note will look wrong. Consultant review is a distinct clinical event with its own weight, particularly where it changes disposition, and it belongs in the record as its own timed entry signed by the consultant rather than folded into the registrar's note. A consultant should be able to see and correct any entry that carries their name.

Shift handover is a documentation event in its own right, and the Communicating for Safety Standard in the National Safety and Quality Health Service Standards expects clinical handover to follow a structured process and to be documented rather than left to memory. Capturing the handover round itself raises a different consent problem, because it happens away from the patient and covers many patients in one conversation. Drafting a per-patient handover entry from that patient's own encounter record keeps consent, attribution and the record aligned.

The note leaves the department by two routes, up into the hospital when the patient is admitted, and out to the GP and the patient when they go home. The admitting team wants the working diagnosis, what has been excluded and by what means, and what is still pending. The GP wants what was found, what changed in the medications, what to watch for, and what follow-up has been arranged and by whom.

The discharge conversation is the most closely examined part of emergency documentation. Return precautions decide whether a patient who comes back was safety-netted or sent away, and advice recorded only as discharge advice given will not help anyone later. What belongs in the note is the specific advice, who received it, and whether they were in a state to take it in. A patient who is still intoxicated at discharge, has nobody at home, or has no transport at four in the morning is a different discharge from the same diagnosis handled in daylight.

That conversation happens out loud, which makes it good raw material for a draft, and it is the part compressed into one line when the note is typed an hour later. The same applies when a patient leaves before treatment is complete, where the record has to show what they were told about the risk and what was offered before they went.

Every emergency draft should also state what was still pending when it was written, including results not back, imaging not reported and the review not yet done. Most emergency notes are written before the story has finished, and saying so inside the note is what makes it safe for the next clinician to rely on.