It’s seven o’clock and the waiting room emptied hours ago. The GP is still at her desk, in the quiet, working through a prior-authorisation form for a medication she has already decided her patient needs. She knows the answer. She has known it since the consult. What she is doing now is not medicine. It is finding the right field, in the right portal, to say a thing she already knows to a system that will make her say it three more ways before it agrees.

She trained for years to be the person in the room who decides. Somewhere in the last decade the job quietly rearranged itself so that the deciding takes twenty minutes and the paperwork about the deciding takes the evening. Every night it recedes a little further — the reason she started.

The day and a half

This is not a story about one tired doctor. It’s the shape of the week for a lot of them.

The American Medical Association’s 2024 survey of a thousand physicians found prior authorisation alone eats about thirteen hours per physician per week. Widen the lens past prior auth to the whole clerical load (documentation, the inbox, coding queries, the forms), and an NHS study of 966 staff put it at roughly a third of the working day. The inbox by itself keeps growing: primary-care inbox time rose about 24% across 2019 to 2023, according to the Annals of Family Medicine. None of that is care. All of it lands on people who trained to give care, and it lands after hours, because the care has to come first.

Call it a day and a half a week of arguing with forms. That’s the number that matters, because it’s the number that made her stop enjoying the job.

The person on the other side of the form

The efficiency conversation usually skips this part. The 7pm form is not only the clinician’s problem. Someone is waiting on the other side of it.

In that same AMA survey, 95% of physicians said prior authorisation delays patient care, and 79% said it can push a patient toward abandoning treatment altogether. The delay is not administrative. It has a body. And the referral that should have carried a patient onward often just… doesn’t: Healthwatch England’s 2025 review found around 14% of referrals are delayed, lost, rejected, or never processed, and of the people who hit a referral problem, only 7% were satisfied with how it was handled.

So the forms cost twice. They cost the clinician her evening, and they cost the patient the care the clinician already decided to give. The gap between “the doctor decided” and “the system let it happen” is where both of those costs live.

Why the obvious fix is the wrong one

If you’ve read anything about AI in healthcare this year, you’ve read about the ambient scribe — the microphone in the consult room that listens and writes the note. It’s the most-marketed thing in the sector, and the evidence for it is genuinely good. Olson and colleagues, publishing in JAMA Network Open in 2025, found clinician burnout fell from 51.9% to 38.8% within thirty days of an ambient scribe rollout across six US health systems. A sober 2026 JAMA study by Rotenstein and colleagues put the floor honestly: about sixteen minutes of documentation time saved per eight-hour day, thirteen minutes of total EHR time, no rise in billing denials. Modest but real, and worth having.

But nobody selling it will lead with the catch. The scribe is a closed, in-product feature. You rent it. It writes its notes into your practice-management system through a private partnership, and it exposes nothing an agent of yours can reach or reconfigure. The scribes (Lyrebird, Heidi) and the analytics tools like Cubiko are consumers of your data, not doors into it. You can buy them. You cannot build on them.

That matters because the reachable win, the one you can actually shape around your bottleneck, lives one layer down, in the operational system where your data already sits. And whether AI pays back for your practice is decided less by how clever the model is than by which tier your core system sits in.

The honest map looks like this. If you’re an allied-health practice on Cliniko, MediRecords, Nookal, Coreplus or Halaxy, you’re in luck: these expose clean, self-serve provider APIs — REST and webhooks, FHIR, proper read-and-write — that an agent can reach without a partner program. That’s the open layer. One rung down are the dominant GP and specialist clinical systems, and they are gated. Best Practice and Zedmed reach their data only through a FHIR facade via Halo Connect; MedicalDirector Helix through Telstra Health’s Smart API+; Genie and Gentu through the Genie Partner API. Real, but partner-onboarding required. The secure-messaging and claiming rails — HealthLink, Medical-Objects, Tyro Health — are gated the same way. Below that again is the national layer: My Health Record and Medicare/ECLIPSE, reachable only through PRODA and NASH certificates and conformance testing: genuine, but high-barrier. And the enterprise hospital systems, Epic and Oracle Health, are org-gated: not something a mid-sized practice self-serves at all.

I’ll say the uncomfortable part plainly, because a buyer deserves it: “we integrate with standard APIs” is hand-waving if your core system is Best Practice or MedicalDirector, because those are not clean REST you flip on. If a system can’t expose its data or its events, the use case built on it stalls. That’s not a reason to walk away. It’s the first thing an honest diagnostic tells you, before anyone promises a build.

What actually changes

So put the scribe to one side and look at the 7pm form, which nobody markets because it isn’t glamorous.

An assistant sitting on the practice-management layer can read the clinical note and the plan and draft the prior-authorisation request — the fields filled, the justification assembled from what’s already in the record. Electronic prior auth saves roughly fourteen minutes per authorisation, on CAQH’s numbers. The message triage copilot reads the day’s patient inbox and sorts it by urgency: in the SCPMG deployment reported in JAMA Network Open, across 3.03 million messages, the time for a high-acuity message to be first read fell from about twenty-two hours to about five, with topic-classification accuracy of 81% against 44% for the legacy system. An intake sorter reads inbound referrals so the 14% don’t fall down the hole. A coding assistant drafts the MBS item; an Easy-ICD trial in JMIR in 2025 cut median coding time on longer notes by 46%.

In every one of those, the machine drafts and the clinician decides. This is not a slogan; it is the architecture, and it is load-bearing. A wrong MBS item is not a typo: it’s a Medicare compliance event, with PSR and clawback exposure, and the liability lands on the billing practitioner. So no code is ever silently finalised. The clinician confirms every code before it’s claimed. No prior-auth is auto-submitted, and no message or referral is actioned without a human. AHPRA’s August 2024 guidance permits AI that assists when a clinician reviews and approves the output, with consent and retained accountability, and that’s exactly the line this sits behind. Every draft-to-sign-off step is timestamped, retained in your own tenancy, on Australian soil, never used to train anyone else’s model, and exportable if AHPRA or Medicare ever ask.

What changes is not who decides. It’s whether the deciding still has to happen at 7pm.

The honest ledger

Two things I won’t dress up. First, RMAI is new to this sector. The numbers above are independent, peer-reviewed reference benchmarks — not our client results. The strongest evidence is enterprise-scale; small-practice figures are more modest, and we recalibrate to a small-practice cost base in the diagnostic rather than borrowing an American hospital’s ROI. Second, the payback is conditional. Hard-dollar admin savings bank faster than reclaimed clinician time, and reclaimed time erodes if you just pour more admin into it. The tool alone doesn’t pay you back; the tool plus a redesigned workflow does.

With that said: a build typically ships in three to six weeks in the $10k–$60k band, quoted fixed-scope after a free discovery call on your own numbers, and, directionally rather than as a promise, pays back in roughly four to eighteen months at realistic practice volumes. If you want to know what the day-and-a-half is worth to you, do the arithmetic yourself: your clinicians’ admin hours, multiplied out, valued at what an hour of their clinical time is actually worth. For most practices that lands in six figures a year. That figure is illustrative — your own, not a benchmark — and it is the number the 7pm form is quietly costing you.

Where to start

You don’t need a transformation programme to find out whether this is real for you. You need one honest look at where your data actually lives, and which tier your core system sits in.

Two ways in, both low-risk:

• Read the map first. We’ve written a plain-English brief on the Australian healthcare software landscape — which of your systems can actually be reached, by what mechanism, and what “gated” really costs you. → https://realmindsai.com.au/guides/healthcare/
• Book a free 30-minute discovery call. We’ll take one real path — a prior-auth, a referral intake, a message queue — name the API tier underneath it, and show you the highest-value place to close the gap, before anyone promises a build. → https://outlook.office.com/book/[email protected]/?ismsaljsauthenabled

There’s a form open on a screen somewhere in your practice tonight, being filled in by the most expensive and most exhausted person in the building. The only question worth asking is whether the system is ever going to fill in the parts that were never hers to do.