The first wave of hospital AI answered questions. The next wave finishes the job. For Australian health services, the evidence now supports getting started.
If you have used a chatbot, you know the first generation of hospital AI. You ask, it answers, and the work the question was about remains exactly where it was. The systems now arriving are different in kind. An AI agent is software given a goal rather than a question. It breaks the goal into steps, carries the steps out across the hospital’s own systems, checks its results and asks a human when judgement is needed. Researchers reviewing the field describe this as the defining shift: from tools that respond to tools that initiate.1
For readers new to clinical AI, one concrete example carries most of the idea. “Get Mrs Chen discharge-ready by Friday” is a goal. Meeting it means a medication review, a discharge summary, an aged-care referral and booked transport. A chatbot can describe that list. An agent works through it overnight and hands a clinician the completed pieces to approve.
Capability is measurable, and the numbers are encouraging. MedAgentBench, a Stanford benchmark, set frontier AI models 300 physician-written tasks inside a simulated medical record. The best completed roughly seven in ten.2 Two years ago the honest answer would have been near zero. Seven in ten is not a licence for autonomy; it is ample capability for supervised work, and the trajectory suggests the figure will keep rising.
Application · 01
Every Australian clinician can name the forms. PBS authority scripts, sought by phone or through a portal the RACGP says was not designed around its users.3 NDIS evidence bundles. Aged-care assessments. Private-insurer eligibility checks. None of this workload requires a clinical decision, yet nearly all of it consumes clinical time. In the United States, where the equivalent burden runs to about 13 hours per physician each week, health services already deploy agents that read a funder’s decision, gather the supporting evidence from the record, draft the response for sign-off and track every deadline.4 The Australian paperwork is different in name and identical in shape. This is the natural first deployment, and the reason to feel comfortable about it is structural: the agent’s output is a draft on a form, reviewed by a person, with every error recoverable before it reaches a patient.
Application · 02
The prize in discharge coordination is unusually well documented here. A South Australian government report presented to the nation’s health ministers in September 2025 counted close to 2,500 older patients in acute hospital beds awaiting an aged-care place, a queue costing roughly $1 billion a year.5 The Australian Medical Association calls it exit block.6 Those beds represent the capacity every state is trying to build with concrete and cranes, and a meaningful share of them can be released with coordination alone.
The building blocks exist. A model published in Communications Medicine predicted next-day discharge with an AUROC of 0.87 and projected whole-of-hospital discharge numbers within about 5 to 9 per cent.7 Wire a forecast like that to an agent and the overnight hours become productive: pharmacy review queued, draft summary prepared, the My Aged Care or NDIS referral assembled, transport held. A formal trial linking agent-coordinated discharge to shorter stays has yet to be published, which argues for measuring as you deploy rather than for waiting. The components are individually proven, the downside is small, and the cost of the status quo is a billion dollars a year and rising.
Application · 03
Radiology shows what happens when AI is allowed to act on a workflow instead of just commenting on it. Worklists have historically run first-in, first-out, treating a routine follow-up and an unread pulmonary embolism as equals. When AI reorders the queue by what the images likely contain, urgent findings reach a radiologist sooner: one hospital study recorded report turnaround for embolism-positive scans falling from 59.9 to 47.6 minutes.8 A systematic review of 38 studies found the same pattern across stroke, haemorrhage and chest disease, with a telling detail: systems that actively re-sorted the worklist cut turnaround by 43.7 per cent, while passive pop-up alerts managed 7.6.9 Acting beats flagging, by a factor of about six.
Australia is not importing this capability; it is exporting it. Annalise CXR, built in Sydney by Harrison.ai with I-MED, reads 124 findings on a chest x-ray. In a Lancet Digital Health study it improved radiologist accuracy on 102 of them, and it now operates across the country’s largest private radiology network and in health systems abroad.10 Local proof, at national scale, from a local company.
Agents do the paperwork; clinicians keep the decisions. That division holds up clinically, industrially and, under the TGA’s rules, legally.
The plan–act–verify loop, the three numbers that make the Australian case, and where a health service can safely start.
Clear regulation is an adoption asset, and Australia now has it. TGA guidance finalised in February 2026 draws the line by intended purpose: software that shapes diagnosis or treatment is a medical device and belongs on the ARTG, while software confined to administrative work sits outside the scheme.11 A board can apply that test in a sentence. If the agent’s output changes what happens to a patient, it is a device with a defined approval path; if it changes what happens to a form, it can be deployed under ordinary governance today.
Supervision, likewise, is best understood as a design feature rather than a brake. A long research tradition, going back to Bainbridge’s work on automation in the 1980s, shows that automating routine work changes the human role rather than removing it: people shift from doing to checking.12 Hospitals that design for that shift from day one, with audit trails showing what an agent read, did and queued for approval, get the productivity without surrendering the judgement. Early evidence suggests clinicians prefer this arrangement to the status quo of doing the paperwork themselves.1
Few countries are better positioned. Australia has a home-grown clinical AI industry with published, peer-reviewed results. It has a regulator that has already drawn the line adopters need. It has national standards, single-payer infrastructure and, in exit block and clinician burnout, pressure points that are documented to the dollar. Amalthēa Health is developing a Clinical AI solution that comes ready for supervised administrative work today, and each capability gain widens what supervision can safely delegate. We have built in the governance muscle, the data plumbing and it’ll fit into your existing workflow. The hospitals that thrive in the next decade will be the ones that use AI to manage the administrative load. We are now reviewing requests for early access, join the waitlist.