Rolling out an AI scribe across your practice

The first pushback never comes from the software. It comes from your own team.

I ran a multidisciplinary clinic in New Zealand. The people who dug their heels in were the ones who'd written notes the same way for years. Some of them reckoned the tool was there to replace them. It isn't. It drafts the note. They read it, fix it, sign it.

An AI scribe is software that listens to a consult and writes up the clinical note, so a clinician can check and sign it instead of typing it out after hours.

So a rollout needs a process, not a memo. Consent, then policy, then the software setup, then training, then someone checking back. Skip a step and the risk doesn't go away. It shows up later, usually in front of a patient.

The short version. Pay for the follow-up. Most practices buy the software, run one training session and hope. Give your team two to three weeks to get comfortable, pick one champion who actually wants the job, and write a 30, 60 or 90 day hypercare window into the project before you sign. That's the bit that decides whether any of the rest sticks.

What's on this page, and what isn't

Haven't picked a scribe yet? AI scribes for Australian clinics: how to choose one is the buying guide. What to compare, what to ask a vendor, what to check before you sign. This page picks up the day after you've signed.

After the general running order for any AI tool, rather than a scribe specifically? That's the AI implementation roadmap. This page is the scribe version of the same job.

And the rules underneath all of it, such as RACGP's new F11 standard and the vendor indemnity clause, sit on privacy, regulation and ethics in AU and NZ. Both are quoted in full and in plain words over there, so I'll point you at them rather than say it twice.

Why this needs a process

The federal Health Department told The Guardian in July 2026 that AI scribes "have little oversight", and found "a significant variation in how clinicians and practices obtained consent from patients" (The Guardian).

Australia's Privacy Commissioner, Carly Kind, went harder in a May 2026 speech. She pointed to "deficiencies in GPs' implementation of AI scribe consent protocols, and the absence of disclosure around scribes in privacy policies" (The Guardian).

None of that is a reason to skip a scribe. It's a reason to have a plan before you switch one on.

Consent

Informed consent is the patient knowing the tool is listening, knowing roughly what happens to the recording, and saying yes anyway. A sign on the wall isn't it.

Avant, a medical defence organisation, gives three reasons to get it: "1. to comply with privacy obligations, 2. to avoid inadvertently breaching surveillance devices legislation, 3. to be transparent about the use of AI and enhance trust" (Avant).

RACGP is blunt about it: "GPs must obtain consent from their patients prior to using an AI scribe during consultations, with some medical defence organisations (MDOs) advising that they obtain their patients' consent in writing" (RACGP).

Avant also publishes a script for that first chat with a patient: "I have started using an AI software tool that listens to our consultation and then summarises the information into a clinical record... The recording is stored securely and when I am finished the draft is deleted. Do you have any questions or concerns? If you are happy to go ahead, I will be using the tool for this appointment and future appointments, but you can decline its use at any time" (Avant).

Steal it. Say it out loud a few times first so it sounds like you and not like a policy.

After that first conversation you don't need to do the whole thing again. A verbal yes, written into the record, is enough at later visits, according to Avant (Avant).

Two ways this goes wrong

A sign in the waiting room is not consent. MDA National's case study is a patient who had no idea a scribe was recording and threatened to complain. That's the exact thing a waiting-room sign is meant to prevent, and doesn't (MDA National).

The other way is quietly making it compulsory. One Melbourne psychiatrist wouldn't take patients who declined an AI scribe, The Guardian reported (The Guardian). Patients get to say no. Have a plan for when they do, so the consult still runs properly, Avant advises (Avant).

Policy

RACGP's fact sheet asks for three things. Write a practice policy on using AI scribes. Update your privacy policy. Review both as you go (RACGP). Start by deciding, as a practice, which scribes are allowed. Then say so in the privacy policy patients can actually read.

New Zealand is a bit more work. The Privacy Act's Information Privacy Principles cover AI tools there, and the Privacy Commissioner wants a Privacy Impact Assessment done before you adopt one (NZ Privacy Commissioner). Do it and your policy has a documented starting point. Read the Medical Council of New Zealand's March 2026 guidance on AI in patient care too (MCNZ), and check your own board's code if you're not a doctor.

Name one person to keep an eye on scribe use across the practice, Avant advises (Avant). Someone owns it, or nobody does.

Two things sit next to the policy. Something plain and patient-facing, and the script above is a fine start. And a short internal list of which tools are actually approved. ClinicComply, an Australian compliance template vendor, sets that list up as approved tools with a risk tier on each, reviewed on a schedule (ClinicComply). You don't need its template. A page saying what's approved, who approved it and when it was last looked at does the same job.

Getting it into your systems

Check your practice management system, or PMS, is on the vendor's own published list before you buy. Our post on choosing a scribe goes through what each vendor confirms on its own pages.

RACGP's fact sheet also flags the boring physical stuff: a decent microphone, a second screen, internet that holds up (RACGP).

"Integrates with" can mean three very different things, and each one changes the cost and the timeline.

So ask which one it is, whether anything has to be built, and who's doing the building. For example, joining a CRM like HubSpot to a practice management system such as Cliniko can be quick or slow depending entirely on that answer.

Training, and the champion

Here's the bit people rush.

An AI champion is the person who carries the tool day to day. First to learn it, first to get asked about it, and the one who tells you when it isn't working.

Pick the champion at the start, and pick a volunteer. Mandating it doesn't work. The right person usually has a genuine interest in the technology anyway, and they'd have put their hand up without being asked. Take responsibility as the owner for making sure people get trained, and don't force anyone to use it. The Australian Physiotherapy Association says the same, and puts the adjustment period at two to three weeks so staff can settle in and tune their templates (Australian Physiotherapy Association, InMotion).

Two to three weeks matches what I'd expect. Bigger practice, longer run. More people, at more levels, all coming along at their own pace.

On the shape of the rollout: if you can, take everyone at once. Same training, everyone at the same level, and the admin stays simple. If that would cause chaos, start with a couple of keen champions instead and let them pull the rest along.

That works because people notice. A colleague walking out on time instead of doing notes at 9pm gets attention fast. It's the same pattern you see selling anything: a few take it on early, most follow once it's proven, and a handful hold out until last.

Whichever way you go, say it out loud on day one. What's changing, why, and what you expect. Some people find change harder than others, so give them more, not less. Break it into small steps and record how to do each one. A two minute screen recording they can play again beats a training session they half remember.

Heidi, one of the scribe vendors, names internal champions in its own adoption advice. Its guide puts typical clinician activation at 20% to 40% (Heidi). That's Heidi's number, not independently checked, and worth sitting with before you assume the whole team is using the thing by week one.

Hypercare, the bit nobody pays for, and where Opal helps

Hypercare is a fixed window after go-live where one named person has to check the tool is actually being used, answer the questions that only turn up in real life, and fix what got missed. Thirty, sixty or ninety days, agreed before the project starts.

The step people skip is this one, and they always regret it. Training gets rushed. The project gets delivered. Everyone washes their hands of it and nobody checks back.

So put the window in the scope. If a consultant is running the rollout, put it in the price too, and say what it's for. It's the cost owners least expect, because you pay for the project and you pay for the software, and then you don't pay for the adoption. So you're rolling the dice on whether it sticks. Lose that roll and everything you spent earlier goes with it.

If your rollout needs hands-on help, AI agents is the work we do to get a tool like this running properly across a whole team.

When use stays low months later

If the scribe was bought and barely anyone touches it, ask the practitioners first. Some of them won't want to use AI, and that's fine. Nobody should be forced. What you're really looking for is a gap between what they think and what they've told you.

Then ask whether the thing is easy to use. Ease of use is the main barrier to adopting any new system, CRMs like HubSpot included.

Checking back

Avant's checklist keeps it simple. Check AI-generated notes after every consult, name someone to monitor use across the practice, and update your privacy and cyber security policies (Avant).

If you're a general practice chasing RACGP accreditation, Criterion F11 asks for a risk check before you implement and a monitoring process afterwards. It isn't assessable yet and the transition date hasn't been announced (RACGP). Not chasing accreditation? It doesn't apply to you. The whole criterion is quoted and translated on our privacy and regulation page.

One more thing to do once, not monthly. Open the vendor's terms and read the indemnity clause, because a broad one can sit outside what your own insurer will cover. That page covers what those clauses actually do to you, with Avant's warning quoted in full.

The rollout checklist

Consent

• Use Avant's script, or your own version of it, at the first consult.
• Write the verbal yes into the record at every visit after that.
• Have a plan for patients who say no.

Policy

• Decide which scribes are approved.
• Update the privacy policy so it says you use AI.
• Name one person to monitor it.
• Keep a short list of approved tools.

Setup and training

• Check your PMS is on the vendor's own list.
• Check the hardware: microphone, second screen, internet.
• Pick everyone at once or champions first, and say which and why.
• Pick the champion at the start, and take a volunteer.
• Give it two to three weeks before you judge how it's going.
• Record short guides, and give extra help to the people who need it.

After go-live

• Agree a 30, 60 or 90 day hypercare window, in the scope and in the price.
• Keep checking AI notes for accuracy well past week one.
• Set the policy review date now, not later.

For where AI fits across a clinic more broadly, such as admin, documentation, patient messaging and marketing, see AI in healthcare administration.

FAQ: the questions I get asked

Do patients have to consent, and does it have to be in writing?

Yes, you need consent. RACGP says GPs must get it before using an AI scribe, and notes some medical defence organisations recommend getting it in writing (RACGP). AHPRA lists informed consent as one of five principles for every registered practitioner using AI (AHPRA), so it isn't only a GP question. Avant's line is that a proper first conversation, then a verbal yes written into the record at later visits, is enough (Avant).

What goes in an AI scribe policy?

RACGP builds it around three things. Write a policy on which scribes get used and how. Update your privacy policy so it discloses them. Review it as you go (RACGP). Avant adds naming someone to monitor it (Avant). For general practices chasing RACGP accreditation, the new F11 criterion, not assessable yet, asks for a risk check before you implement (RACGP).

How do we train staff without losing clinic time?

Decide upfront: everyone at once, or a few champions first. Then break it into small steps and record short guides people can come back to. Heidi's own adoption guide, one vendor's account of its own rollouts, names internal champions as a factor in getting there (Heidi). Give it two to three weeks before you decide it isn't working (Australian Physiotherapy Association).

How long should we keep checking in after go-live?

Longer than you think. Agree a hypercare window of 30, 60 or 90 days before the project starts, name the person who owns it, and put it in the scope and the price. Let's say you go live in March and nobody looks at it again until the renewal invoice in September. By then people have drifted back to typing their own notes, and you've paid for a year of something nobody uses.