An AI scribe for
nurses and nurse practitioners

A nursing entry is written to be acted on by someone else, usually within hours. The next person on shift, the after-hours doctor and the nurse picking up a community round all read it looking for the same things: what was assessed, what was found, what was done, how the patient responded, what is still outstanding and who has been told. The Nursing and Midwifery Board's registered nurse standards for practice require accurate, comprehensive and timely documentation of assessments, planning, decision-making, actions and evaluations. How soon an entry has to be in the system is then set locally by the health service or practice clinical records policy, and that is the deadline a documentation tool has to fit inside.

The part of a nursing record that gets read hardest afterwards is the escalation trail. When a matter reaches a coroner, a complaints body or a claim, the questions asked of nursing documentation are almost always about recognition and escalation: what was observed, at what time, who was contacted, what they were told, when they attended and what happened in between. An entry recording only that the registrar was called leaves every one of those open. An entry recording that the registrar was called at 14:20, given the observations and asked to review within the hour, answers most of them from the record itself.

Handover is the other structural pressure. The national standard on communicating for safety expects clinical handover to be structured and documented, which is why ISBAR and its variants are in use across Australian health services: the receiving clinician needs the same fields every time, and the written record has to survive the person who gave it going home. None of that has the shape of a consultation narrative, and a tool built around one clinician and one patient talking in a room will not produce it unless it is told what shape the entry takes.

Sort nursing documentation into two piles and the decision about ambient tools becomes straightforward. One pile is structured data entered into a field: observations on a track and trigger chart, fluid balance, pain scores, falls and pressure injury risk screens, wound measurements, immunisation details and the medication administration signature. The other pile is narrative: the assessment, the conversation, the education given, the clinical reasoning, the escalation and the plan. Ambient capture works on the second pile only.

The limit is not that the numbers are hard to hear. An observation entered in its field is plotted, compared against thresholds and able to trigger a response under the national standard on recognising and responding to acute deterioration, and none of that happens to the same observation written into a paragraph, which is also invisible to the person scanning the chart at three in the morning. A set of observations narrated into a progress note is worse than leaving them out, because it looks documented while doing none of the safety work.

Medication administration is the clearest case. Signing the administration record is an act performed at the time by the person who gave the medicine, tied to the checks they made in front of the patient. No draft substitutes for it. Immunisation is close behind. Reporting National Immunisation Program vaccines to the Australian Immunisation Register has been mandatory since 1 July 2021, the report carries the brand name, the dose number, the batch number and the date given, and the Register expects it within 24 hours and no later than ten working days after the vaccination. A batch number is an alphanumeric string read off the vial and it is the handle a recall works from, which makes it exactly the sort of value dictation gets wrong, so it gets typed.

Make a vendor show the workflow against your own charts rather than against a consultation demo. These stay manual in every ambient tool worth taking seriously:

• Observations entered on the chart your service actually uses, including the escalation fields that sit beside them.
• Medication administration signatures, including a second-person check where local policy requires one.
• Immunisation entries with brand, dose number, batch number, site and route, and the report to the Australian Immunisation Register.
• Wound measurements taken with a ruler, and photographs taken under the service's own clinical imaging policy.
• Risk screen scores and the care plan fields the software calculates, date stamps or schedules a review against.

A nurse practitioner consultation is structurally much closer to a medical consult than to a shift note. There is a presenting problem, a history, an examination, an assessment with the alternatives that were weighed, a plan and safety netting. A nurse practitioner providing Medicare services holds their own provider number, documents under their own name, and the entry has to stand on its own for whoever reads it next. This is the strongest case for ambient capture anywhere in nursing work, for the same reason it holds in general practice: most of the content is spoken aloud in the room and currently gets typed afterwards from memory.

The regulatory frame around this moved on 1 November 2024, when the legislated requirement for a collaborative arrangement with a medical practitioner was removed for nurse practitioners and eligible midwives providing Medicare services and prescribing under the Pharmaceutical Benefits Scheme. Consultation, referral and transfer of care did not stop being part of good practice. They are simply no longer evidenced by a standing document filed somewhere, so the clinical record is where they have to show up. An entry that names who was consulted, what was asked, what came back and when care was transferred is now doing the work the removed paperwork used to do.

Scope of practice is the other thing the note has to carry. The Board's standards require a nurse practitioner to practise within their scope, and its guidance treats scope as something the individual has to judge in context rather than something a list settles in advance. Documentation is how the judgement is demonstrated afterwards: the differentials considered and set aside, the reason a particular course was chosen, who was consulted about it and the point at which the patient was referred on. Reasoning spoken aloud during a consultation is what an ambient tool captures well, and it is the first thing lost when a note is typed at the end of a long list.

Safety netting is a smaller gain and an easy one to miss. The instruction about what to do if symptoms worsen, where to go and when to come back is usually given while the patient is already standing, and it is a common omission from a note written up hours later. Ambient capture picks it up because it happened in the room. It still gets checked at review, because the wording has to be what the clinician actually said.

Nurse-led work in Australian primary care covers chronic condition reviews, wound clinics, immunisation sessions, spirometry and health assessments for older people. These are long appointments with a great deal of talk in them, and the same line runs through them: the conversation, the education and the agreed plan can be drafted from the encounter, while the spirometry values, the recall interval and the care plan review date cannot.

Where a service is provided on behalf of a general practitioner, the record has to make the arrangement visible: which nurse provided the service, which doctor it was provided on behalf of, and that whatever the item requires of the doctor personally was actually done. Health assessments are the common trap. Under those items a practice nurse may assist by collecting information and giving the patient information about recommended interventions at the medical practitioner's direction, and the medical practitioner still has to see the patient as part of the service and remains responsible for it. A draft that blends the two into one voice creates a claiming problem as well as a clinical one, so the review step puts the attribution back in.

Chronic condition management planning has its own shape: the patient's goals, the agreed actions, who is responsible for each of them and when the plan gets reviewed. Most of that is generated out loud in the review appointment, and a draft that keeps the patient's own words holds the entry to what was actually agreed rather than to what the template proposes. The plan itself is still completed in the structured fields of the practice software.

Nurse-led telephone work sits outside the ambient case. Recording a phone call brings state and territory listening and surveillance devices legislation into the decision and the rules differ between jurisdictions, so treat it as a separate policy question with its own advice rather than as an extension of the consent given for an in-room consultation.

Clinical software attributes an entry to the user who is logged in, and the audit log records who wrote what and when. The questions to put to a vendor about a scribe sitting alongside that record are narrow: does every clinician hold their own account in the scribe as well as in the clinical software, does the signed entry carry the identity of the person who reviewed it, and does the audit trail show that individual rather than a shared practice login.

Three failure modes turn up in real practices. A shared login inside the scribe means every draft in the trail belongs to one identity, which destroys the evidence of who reviewed what. A nurse's draft pasted into the record under a doctor's login attributes a nursing assessment to a clinician who did not perform it. A single blended entry for a visit where the nurse did the observations and the injection and the doctor did the consultation leaves nobody able to say who assessed what, which is precisely the question asked when something is missed.

The clean pattern is one entry per clinician, each reviewed and signed by the person who did that part of the work. Where a service genuinely wants a single entry, the text has to name which clinician did which part. A scribe will follow whichever pattern the practice has already settled on and it will not settle one, so the choice belongs in the documentation policy and in the software's entry templates rather than with whoever happens to be rostered that day.

An unsigned draft is not yet the clinical entry, but it is health information the practice holds, which puts it inside the same privacy obligations as the signed note and within reach of a subpoena or a patient's request for access. A practice should know how long unsigned drafts live, who can see them and what happens to one that is never signed, because a draft abandoned halfway through a shift is the case most documentation policies do not cover.

The usual primary care sequence puts the nurse first. Observations, a dressing, an injection, a screening question set, and then the doctor joins or sees the patient afterwards. Consent to record given to the nurse covers the encounter the patient understood themselves to be in. Patients do not assume it follows them into a different conversation with a different clinician, and a recording that quietly continues across that handover is hard to defend later.

The workable habit is that whoever is in the room owns the explanation, and gives it in plain words before anything starts: a tool is being used to help write the notes, it listens to the conversation, the nurse checks and signs whatever it drafts, and the patient can decline without anything else about the visit changing.

Access should follow who was actually present. A nurse who was in the room for the first ten minutes has a reason to see that material. A nurse who left before an examination began does not. The Board's code of conduct for nurses puts the principle plainly: nurses access records only when they are professionally involved in the person's care and authorised to do so, and a scribe's library of recordings and drafts is a record store like any other. A tool that gives every clinician access to every recording has widened the audience for a consultation well past what the patient agreed to, so check it inside the product rather than assuming.

The handover points are where a team gets this wrong:

• The clinician who starts the capture is the one who explains it, and the explanation comes before the recording starts.
• Name the recording again in one sentence when a second clinician joins the room.
• Decide in advance whether capture continues while the nurse steps out for equipment or a result, and default to stopping it. If it continues, it belongs to the clinician who stayed.
• Record the agreement and record a refusal, so the next nurse on the roster is not asking the same question at the next visit.
• Treat interpreters, carers and family members in the room as people also being captured, and tell them.