The data costs for storage for the model are in the billions of dollars. This does not include the collation and processing of the data to create the model, which could also hit a billion dollars. So now we know how much it will cost, who could afford it and why would they pay it?

There’s only a few players. Big Pharma, Google, Microsoft and Amazon.

In the valuation column, the tech giants win, but in the cashflow column, big pharma wins.

Each of these players has a stake in the development of healthcare AI in different arenas, all trying to corner different segments of the market. Because the market is huge. $26.7 trillion projected by 2035.

If an AI model influences just 10% of global prescription decisions, that translates to $170 billion annual revenue shift.

So there could be a battle of the giants to see who gets there first right? Unless they team up.

Eli Lilly is building a $1 billion joint AI lab with NVIDIA and Roche launched an AI factory powered by NVIDIA in May 2026. Merck has partnered with Google and Amazon owns One Medical, PillPack and Prime Health, the infrastructure to diagnose, prescribe and deliver medications.

The friction arises because Big Pharma has all the data, but Big Tech has all the infrastructure, and neither can create the others’ tools or information.

So this is not just a black box, it is a black hole. There is no recourse for the patient in the middle and potentially the patient will be disadvantaged in a far more efficient manner than ever before.

In the above scenario, the models may not be trained to the greater good of humanity, but rather to the greater result for the company, tech or pharma. Their objectives will become the model’s objectives. The objectives become the arbiter of the result, and the supervisor of the system.

The patient asks the AI for a diagnosis. The AI is trained on the company data and objectives, what they have allowed to be in the model and what they want the model to produce. Combined, the diagnosis and recommended treatment will be influenced under those parameters. The patient thinks they have a diagnosis and treatment plan, but it has been shaped with invisible influence.

To create competition with an unbiased model, that is, one not funded by commercial interests, is effectively impossible given the amount of money involved to create a model.

Looking into the future, I can see a streamlined system where everyone is treated uniformly against the criteria set in the model. There is no deviation from the clinical path decided by the supervisor of the model. If medicine and healthcare were uniform in nature and able to be universally applied this could be a good thing. But healthcare is not uniform.

What happens when the model determines that the treatment budget has been reached? So the 1,001st patient enters their symptoms and instead of a treatment option, they are automatically told there is nothing to be done. I could easily see those types of budgetary restrictions placed into the model.

It could also get to the point where all the beds in all the hospitals are accounted for and more. How many drugs have been prescribed compared against inventory to help determine production? The potential is endless, and that’s the problem.

We are already in it. This is being developed right now as I write this. The Office of AI in Australia has no health mandate. We will be governed by it before we have decided whether we want to be. It is up to patients and citizens to start ringing the bell on where this goes, because from where I sit, and with what I know about how the systems work, this is not going to be good for patients, but it will be great for the companies.

Here is what the money says.

- Pfizer has a multi-year collaboration with Boltz to build exclusive models that target selection, structure prediction of biologics design. Science Direct. They are utilizing AI for MRNA vaccine targets. MIT are the research lab. Pfizer gets the product. Pfizer plans to spend $11 billion in 2026 in research and development utilizing AI.

- Gilead Sciences have invested in Insitro who builds AI platforms using biological data, for NASH (a serious liver disease). Gilead hopes to leverage this research to develop new drugs for the disease.

- Eli Lilly signed a $250 million collaboration with XtalPi for AI drug discovery. This is of concern as the disease and the drug they are looking for have not been stated. In 2025 A subsidiary of XtalPi, Ailux, entered a second collaboration worth $345 million for antibody development. So the total invested by Eli Lilly is $1-2 billion. To date none of the results have been made public

- Johnson and Johnson and Verily inked deals with NVIDIA for surgical AI Evelnsight

US digital startups raised $14.2billion in 2025, the highest funding invested since 2022. AI health companies collected 54% of total funding or $7.6billion. A total of $10.7billion has been invested into AI-powered health tech in 2025 alone.

So if Pfizer funds a significant amount of the medical research that is published, by funding clinical trials and providing research grants, and they also fund the startups creating the AI, it will probably be fine right?

The models under construction are black boxes so we can’t see exactly what they have been trained on. If I was making a model I would hopefully have a lot about anatomy, biology, cellular biology, Most doctors learn from Gray’s Anatomy, Robbins Basic Pathology, Guyton and Hall Textbook of Medical Physiology, Kumar and Clark’s Clinical Medicine. So these are a given. From there we could go wild and publish all the medical books ever written. Now we need to go to research and publications to stay current on new medical developments.

There is currently a replication crisis in the medical and healthcare literature. This is the ability to repeat an experiment and either confirm or refute the findings of an original paper. This hardly ever happens anymore, and it has led to a muddied literature pool of what can be confirmed and considered fact, or not.

Notwithstanding this, if I was making the model I would put it all in, as much as the memory could take. For context, Pubmed, the lead repository for medical research, currently contains 40 million citations and abstracts. Over 1.5 million articles are published every year. At a glance, the abstracts alone are 12 billion words. But the models will need the full articles, hold onto your budgets.

The average full medical paper is between 4,000-6,000 words translating to 30-50 pages. So 40 million articles equates to 200 billion words. This translates to approximately 1 petabyte (a petabyte is 1,000 terabytes or 1,000,000 gigabytes). In old school language this is 20 million four drawer filing cabinets filled with pages. This equates to all of human knowledge from the beginning of recorded history to 1900. Took me a while to put it in real world context too. Safe to say, it’s really, really big and I probably won’t be able to build my own model after all. A petabyte costs $1.3million USD each and storage costs for five years is about $368,000 each.

So now we have an editing problem. Deciding what actually gets included because not everything can fit. So who decides what goes in? What and how will things be determined as essential or just nice to have? What consequences will those decisions have if the chapter on erythematous lesions (red dots by the way) is left out but that it is your complaint?

As I came to terms with the sheer size of the model I understood that the only true way to get a medical AI is through federated learning. This will not occur because as demonstrated above, the private money have private models which hold proprietary knowledge that won’t be allowed out. So any medical model that arises will only have some of the picture, not all of it. This is the core issue.

That was my practitioner’s logic applied to a technical problem I know something about. The scale of what is actually needed to deliver what the marketing claims already exists doesn’t add up, even a little bit.

Now let’s turn to an expert because I have only done small language models and the above is my logical approach to building a medical model that could truly beat everyone.

Layer 1 The training data foundation

All major textbooks, from every specialty college, in every country. All peer review literature. Pharmacology databases for drug interactions, contraindication, dosages etc.

Data needed 1 – 2 petabytes ($1.3 million each, $368,000 per year storage and operation each)

Layer 2 Passing medical exams

Medical board examinations test pattern recognition in text-based clinical scenarios. A written description of symptoms, test results etc with a multiple-choice answers. AI is apparently extraordinarily good at this because it is based on text matching, rather than training data. Board exams test knowledge, not clinical judgement.

Data needed 100 petabytes. ($1.3 million each, $368,000 per year storage and operation each)

Layer 3 Diagnosis at 87.8% (better than humans)

This was achieved with simulated electronic health record cases. Structured data with complete histories, everything properly formatted, no missing information, no ambiguity. The results were based on a system that was designed to showcase what was possible in a perfect manicured environment. All healthcare records required for complete diagnosis information of all patients.

Data needed 100-500 petabytes ($1.3 million each, $368,000 per year storage and operation each)

Layer 4 What it actually would take to beat a doctor

Multimodal input – not just text but imaging, pathology slides, audio heart sounds, video of patient movement, photographs of skin lesions, retinal scans with real time integration and the ability to ask questions of the patient. Symptoms alone mean different things in different contexts and for different populations.

Data needed unquantifiable but huge

Layer 5 Things that can’t be trained

The patient who says they’re fine but doesn’t look it. An elderly patient unable to articulate what’s wrong but nurses can pick on sight. The physical and emotional pattern recognition of humans.

Data needed unquantifiable but huge

Conclusion

The data does not exist in one place and some of it cannot be captured. Validating all of the data under thousands of scenarios to ensure safety is impractical at best and impossible to be frank.

Special thanks to the much smarter than me expert who actually knows how to build these things who remains anonymous at this time.

Web MD, which outlines 900 conditions and 10,000 drug interactions is only 2-5 gigabytes. It is still the most visited site on the internet for people searching for medical information or diagnosis.

Essentially, it is safe to compare the AI medical doctors as an amped up Web MD chat bot. You can type your query in and get an answer out, just like Google. Except Google will actually give you links to papers discussing your query, several options to choose to get information from and you can check the sources directly. AI models specifically trained on chosen data are black box and will not be giving you their references.

Aside from all of the ethical concerns I have raised and am still to raise surrounding these AI doctors, I think it’s safe to assume, at this point in time, that a true AI doctor is probably a few years away. There is much work to be done in published literature before all is even close to be considered well, pardon the pun.

A quote from the program caused me to pause, for a long time, “How do we reduce demand for the most expensive end of care?”

I sat with this quote for a long time. I re-watched the segment and the whole episode to make sure I heard it correctly. Was I taking it out of context? Did they mean that caring at home is saving money compared to hospitalisations and therefore is a good thing? Or was it just as it was stated. Let’s look at the numbers then.

The Aged Care Budget for 2026/2027 is $45 billion with $3.7billion allocated to care at home. In 2024/2025 4,800 people died waiting for assessment. Current wait list numbers are under contention but could be anywhere from 50,000-150,000 people waiting for assessment. It’s a whole definition thing. The tool was designed because apparently $4bilion was misallocated under the previous scheme.

There are eight classifications for funding:

Level 1 — $10,698 per year
Level 2 — $15,982 per year
Level 3 — $21,920 per year
Level 4 — $29,545 per year
Level 5 — $39,535 per year
Level 6 — $47,957 per year
Level 7 — $58,122 per year
Level 8 — $78,106 per year

The comparison made is that on average, people come in at Level 4 costing ~$30K a year compared to a residential care bed costing $117K a year so the program should be saving the taxpayer money. So how much could it really save?

4,800 people died waiting. Level one need equates to a $51,350.400.00 saving, and at level 8, the “most expensive end of care” a $374,908,800.00 saving. That’s just for those who died waiting, now let’s get to denials or ‘under-assessments’.

1,054 claims were denied and contested in 2024/2025 . At level 4 that’s $31,140,430.00 saved. So combined with assessments denied and those who died waiting the government has saved close to $100million or 2.5% of total budget. This seems a small percentage, but ask the families how small they feel it is. People being under-assessed saves billions of dollars.

Let’s compare this logic to the household budget. I am guilty of pushing my electric bill payment sometimes, right up until the last minute before I get cut off. Imagine this at scale. The power doesn’t get cut off, but people do.

I’m not explicitly saying this is the case. But if we are to be honest, it must be considered.

As an Australian, I would give up a coffee a day if it meant that our elders could receive the care they deserve. I think most Australians feel this way and would be shocked and horrified to find out how our elderly and most vulnerable, are actually being treated. It is the most honourable and respectful cause. We all stand on their shoulders. A country is judged by how they treated their most vulnerable.

Now, let’s look a little bit into the future, not too far in reality.

They replace the algorithm with an AI in a much blacker and darker box than the current algorithm. No way to comb back through the algorithm answers as can be done in the current assessment

If “How do we reduce demand for the most expensive end of care” is a core tenet of the assessment system, and you follow incentives for the people making the decisions on assessments, and the greater incentives for politicians, this situation becomes much darker, if that is even possible. There may be other reasons for these outcomes, but I don’t think incompetence is one of them.

The CEOs and developers of AI doctors seem not to understand that it is not information that is the problem. It’s resources. I would suggest that none of them have spent half an hour on the phone trying to find a bed for a patient who has to be moved out of emergency urgently because there are fifteen ambulances ramped waiting for patients to get into the emergency room. When you eventually find the bed you then have to find transport to that bed. Another half an hour on the phone. Once both are secured you run down to ED to tell the staff because you can’t get through on the phone, only to find that the bed isn’t needed anymore because the patient has died waiting. True story.

That is the real world of healthcare and having a doctor in your pocket will not help you find the bed.

Information is not the problem. We have had all the information we can stand since the advent of the internet and life expectancy has not moved that far. The graphs below tell the story.

Life expectancy is demonstrated compared with the availability of the internet. In Australia life expectancy increased by 5-6 years over 1995-2024. The US is worse and they created the internet. Life expectancy gained 4 years from 1995-2024 but they have the lowest life expectancy among the high-income countries despite having twice as much per capita spent on healthcare.

Chronic disease is the most interesting finding. In 2023, 76.4% of US adults reported one or more chronic conditions, that’s 194 million people. Among young adults, arguably the cohort with the highest engagement in technology, chronic conditions increased from 52.5% to 59.5% between 2013 and 2023, within the era of the maximum internet health information availability.

Devil’s advocate may say well that’s because we could find the disease because we had access to all the information and now people know what’s wrong with them.

Knowing what’s wrong and getting care are two very different beasts. Let’s look at expenditure. The funding increased as the disease burden increased. If information was the answer, both should have decreased.

In Australia wait times for emergency surgery, treatment needed to sustain life, are reaching record breaking wait times in New South Wales, with just 2 in 3 patients seen within recommended National timeframes. Elective surgery, treatment that would improve quality of life and increase longevity had 100,678 people on it in March 2025, waiting for treatment for a diagnosis which had long been established. Semi-urgent surgery wait times were 65 days, up 9 days from the year before.

On paper this may not seem too bad. Now imagine yourself being in chronic pain, unable to eat or sleep and tell me how long those days are.

The AI evangelists tells us AI will train more doctors, if in the future doctors are even deemed necessary after the AI models improve. In the meantime, the AI needs to pick up a shovel and build the hospitals that will house all the patients it diagnoses.

We are in trouble.

Not because the technology can do what it does, but for what this actually means as we are steered by companies, rather than the people, that this affects.

Let’s apply this for just a minute onto the current healthcare system, notwithstanding the article suggested a total overhaul of it noting “significant barriers to implementation remain”, which will be necessary for this to work, but let’s just play for a second.

A doctor with a medical degree, and a provider number, in the nearer than you think future, will type all of your symptoms, and maybe your presentation and some medical history, into the oracle and get the answer of what ails you and what treatment is recommended. The doctor will put in a claim for this to the publicly funded healthcare system and get paid for using their AI.

This is an oversimplification however, it’s not really if you look into the future.

Now, let’s look a little further into the future.

The AI is owned by big health. They funded its development, the information it was trained on, and rated the responses it gave.

So whose drug treatment will it recommend? What clinical pathway will it recommend? The practitioner will hold the liability at this point in time, but it will be led by the AI. Public health policy and private insurers will influence treatment recommendations based on a cost benefit analysis. Sorry, that’s how it is.

Let’s look at the medical research that the AI is trained on.

At present, in 2026, there is a replication crisis in scientific medical literature. This means that papers are published based upon research and experiments and published as the guiding light and answer to a specific question, or finding, from the experiments.

We used to then repeat the experiment to make sure it was true and stood up to scrutiny.

Due to funding constraints, this barely happens now. Published medical papers are accepted as gospel and some have been proven to be outright false by those who bother to replicate.

The medical AI is being trained on all of it as gospel.

Now to the data, your healthcare data, being entered into the AI and then used as further training. Patients do train doctors in the real world. It becomes part of their knowledge on an individual basis. A practitioner will learn how to speak to patients by speaking to patients, will learn treatment and diagnosis by seeing patients, treating them and diagnosing them. They learn from their mistakes. Patients do die through this process. Humans make mistakes, learn from that knowledge and apply it to their working knowledge and practice. Will the AI be told that the patient died because of their recommendation, and if so will the AI take responsibility and learn? Who determines that?

An AI will make these mistakes faster and on a larger scale than any human possibly could. As it rolls out to the entire healthcare system you will potentially have a young trainee registrar who cheated on their exams, in charge of all of the medical and treatment decisions made, at a scale that is barely fathomable from here.

So with a current 87.8% accuracy determined in the laboratory simulation on electronic health records, not a clinical setting, means in the US that’s approximately 298.5 million people who would get correctly diagnosed. So 41.5 million people will be misdiagnosed. In Australia 23.8 million people will be correctly diagnosed and 3.3 million misdiagnosed. For comparison, a team of human physicians has an approximately 85.6 diagnostic accuracy. So the AI shows slight improvement.

Even if the AI accuracy rate increases, this is the scale, the rapid scale, of adverse outcomes. It’s happening in real time today, not in the future.

Today, the public is using publicly available AI models like Google and Chat GPT at an approximate 25% rate (66 million people in the US) and then hide their usage when they see the doctor. Users report it is easier to use AI than see a doctor and they use the information from the AI to decide whether they should even see the doctor.

Deployed into the clinical context, out of the 66 million using this method, based on the model accuracy in the lab, 8 million people will be misdiagnosed by the AIs currently in the marketplace.

As a society we have gained much from technological advances and AI is the most disruptive. We must govern and guide this correctly, for the benefit of the people, not to create a tool that makes quantifying people more rapidly and more accurately than ever before. If you think it’s difficult now to get approval for a procedure, wait until you have to argue with the AI that has never seen you and doesn’t know who is waiting on you to come home. The potential for it to base a decision on quantity not quality is our responsibility. We must advocate for the humans in the middle, not the technology that will decide whether or not we get care.

For governments, the WHO recommended investment in public or not-for-profit AI infrastructure, enact laws and regulations that uphold human rights standards and establish clear legal liability for AI harms. Establish regulatory bodies for assessing AI applications in healthcare, including registration requirements, audits and pre-certification programs. Specifically, for governments to create ovesnight mechanisms that are independent of commercial interests.

For developers and tech companies, they should engage a diverse range of stakeholders, not just data scientists and engineers but healthcare professionals, patients and civil society. Tools should be designed for specific purposes not just implement general purpose tools not fit for purpose. Training data should be transparent with the models having specific limitations. AI must be tested in specific clinical contexts before deployment.

For healthcare providers, human oversight must be maintained. Ultimate responsibility for clinical decisions should be made by a human. Information consent should be obtained from patients if AI is used in their care in any capacity. AI should ensure equitable access to care and not exacerbate existing health inequities.

The five key themes across all domains were safety and risk, privacy and data, transparency and governance, clinical quality and efficacy, equity and bias.

So where does Australia land against these recommendations?

WHO said governments must establish regulatory bodies for healthcare AI. Australia has no dedicated healthcare AI regulatory body. The Office of AI announced in July 2026 will not be online until April 2027 and to date, no health mandate has been announced.

WHO said data must only be used where there is an appropriate legal basis. In Australia, My Health Record is currently uploading pathology and imaging files to the portal with medications to follow shortly. Patients have not been fully informed of the risk to their personal health data and are opted in.

WHO said humans must maintain oversight over clinical decisions. Currently AI scribes are deployed in 40% of general practices without TGA regulation, recording verbatim, confidential patient doctor conversations and transcribing them, with every detail forming part of the model.

WHO said independent oversight must be in place. In Australia 70% of consultations respondents said a national oversight body was needed but as yet has not been created for healthcare.

It seems we all know the what, or this could just be me, but the how is an afterthought. AI has shot first and now we are left asking the questions.

The Australian Government has done nothing concrete to address valid concerns raised by patients surrounding the implementation of AI into healthcare. Technologies like AI Scribes are already implemented in 40% of general practices, without oversight, or policy to oversee.

The Office of AI will not even come into effect until April 2027, yet the market is forging ahead without governance. By the time the Government determines policy, all of Australians’ healthcare data will be in some form of a model, either public models or private in-house models not available for scrutiny.

This affects every domain of healthcare, data sovereignty, clinical safety, billing integrity and service distribution. Everything.

Without governance, these systems are currently, in real time, deciding your future before we as a community have decided what we want that to look like. We are being told what is happening, not determining what should happen.

We are not being governed. We are being instructed.

Medical practitioners use short hand in their note taking. This is for a number of reasons but one of them is to protect patient privacy. Every detail of the consultation is not recorded in the consultation note, only matters that are directly relevant to the presenting problem, or that form part of the patient history of their medical condition. It’s not explicit, it’s implicit to the practitioner and a trained eye. Anyone who wants to find out what’s happening with a patient cannot easily decipher this just from the notes.

The AI scribes have changed this forever. The full conversation is recorded in real time.

The full conversation.

Every detail, concern, life history, opinion expressed is recorded by the scribe and transcribed verbatim at the end of the consultation.

Every detail.

Discussions between doctors and patients are the most intimate of all conversations. Very personal details are discussed and reassurance is sought from the practitioner. Patients are usually worried and may have various ideas about what’s wrong with them, things that doctors know aren’t possible, but patients don’t and so they express them.

History could judge these expressions.

Let’s say in a consultation with your doctor for flu like symptoms you casually mention that your right knee has been a bit sore lately. This is in 2026. It’s not really affecting your activities of daily living (ADL for short) so you’ve just been icing it and using anti-inflammatories from time to time. In 2027 your right knee has become worse and you go back to doctor. You’ve got private health insurance now, had it for 6 months, and think maybe if there’s something wrong with the knee you can do something about it. The surgeon you see recommends an arthroscopy to examine the knee. Your health insurer says no, it’s pre-existing. You say it’s not because it only became an issue this year. Your health insurer examines your records with your general practitioner who print out your entire history with the practice, including all consultation notes.

There it is. In 2026 you casually mentioned your right knee was a bit sore.

It’s pre-existing.

You never went to see your doctor about your right knee until 2027. You did not seek treatment for your right knee until 2027. But it in 2026 it is noted as being sore. It will fall under the clinical pathway of knee symptomatology that leads to disease of the joint leading to intervention.

This is just an example of how medical information could be used in the future as it is recorded today. What about political opinions that may be expressed routinely in a consultation, nothing extreme just a point of view about a specific topic, like vaccines or birth control. History could judge these positions and they are accurately documented as to what and when it was expressed.

This is not hypothetical. The technology is deployed and has been deployed since 2024. The conversations are being recorded. The transcriptions are being stored. And nobody has defined what they can be used for. What have you spoken about with your doctor for the last two years?

There is a place for increased efficiency for the consultation process. Recording and transcribing the entire conversation is not it.

As a human medical biller, all I need is your item number history to know your medical history and complaints. What triggers a review? The clinical pathway.

The Modernising My Health Record (Sharing by Default) Act 2025 has passed and came into effect on 1 July 2026. Pathology and diagnostic imaging reports are now being uploaded to My Health Record by default. In January 2026 the Australian Government announced that medicines would be next. These alone give a diagnosis. A series of oncology items tells you the patient has cancer, a visit to a specialist and allied health points to a physical injury or complaint ie shoulder, knee. The Government has committed over $598 million over two years to get this done.

The consent to upload this information is implicit, but ACT, NSW and Queensland have introduced laws that require patient’s consent. The data is collected anyway through Medicare. The report may be missing, but the item numbers tell you everything you need to know.

Private Health Funds have enough data, and access to data, to give them a complete picture of your medical history based on claiming alone. Private health funds are not allowed to access My Health Record without the patient’s permission. They cannot access it directly as they are not registered healthcare providers under the My Health Records Act framework.

They are allowed to request records from hospitals and providers as part of the audit process. This was the objection of the ABC piece, that funds are over reaching and requesting more information than required to substantiate a claim.

My Health Record is becoming a comprehensive real-time record of every Australian’s health history. Patients share by default. They must actively opt out rather than opt in. And funds, through the audit process, will find ways to use that data.

AI will too.

AI will be able to scan all these records, match against the clinical pathway, and find exceptions at speeds unimaginable to us. A great deal of them will be within range or able to be substantiated. Many will not. This will then lead to another segment of work to be completed by humans, eventually.

Right now, AI can discern anomalies. It can also halt payment and send automatic please explain letters. This may be a correct or incorrect judgement. The onus is on the provider. This is where it becomes problematic in a number of ways.

The provider must now substantiate the claim. This is usually routine as most of the paperwork and information is within the patient record. The provider sends in the information. More often than not, more information is requested. The provider has it and sends it off. The claim is approved and payment is made.

This can take over 60 days to complete. So the provider is now almost 90-120 days from receiving payment for a service that would normally be paid in 7-45 days. The impact on providers is not just the administrative burden, it is also the financial burden.

All practices and hospitals are on thin margins. A delay in cashflow can have real effects on the business.

If you add the scale of an AI determining this, the financial effects to providers could be catastrophic.

The ABC story highlighted the plight of what’s happening now to providers, and to patient health data being shared on demand. My Health Record will be the most comprehensive patient health database in Australian history. AI is being deployed without governance. If I can judge your history just by looking at item numbers, imagine, really imagine, what AI will do with the complete record.

Back in the late 1990s as the internet was emerging, the same rhetoric was abundant. The internet would change business forever, the cost savings would be enormous, and everyone would lose their jobs. This didn’t happen. What happened was an economic catastrophe with the tech bubble bursting and people losing billions of dollars collectively.

The internet did change everything but not in the way people thought at the time, and it happened much more slowly than people had anticipated. Bottlenecks that the masses didn’t see began to emerge and the practical application of the technology slowed everything down, for nearly a decade.

Some people did lose their jobs as internet efficiencies began impacting banking and finance and many other sectors. The population began to adopt the new technology and internet banking, bill paying etc all became the norm. This took about a decade. Today we are still left with a split in the population of those who can easily use the internet for day-to-day life and others who are unable, and unwilling.

The internet created a digital divide, people who could use it and people who couldn’t. That divide still exists 30 years later. AI will create its own divide. But in healthcare, unlike banking, the people left behind by the technology aren’t just inconvenienced. They’re the patients, the practitioners and the administrators whose care, livelihoods and data are affected by systems they don’t understand, didn’t consent to and have no recourse against.

AI is being implemented right now, in more ways than are immediately obvious to the outside observer, and users of healthcare. The governance policy of these implementations is lacking with the Australian Government Office of AI not even in effect until 2027. Meanwhile tools like scribes, recording intimate patient information for transcription, are rolling out to 40% of general practices. Patients consent by default rather than fully understanding what is happening to their data.

The questions we didn’t ask ourselves about the internet in 1996 were who governs it, who owns it, who carries the risk. These questions were answered over time once the key players like Google and Microsoft emerged and drove the market. Governments eventually began to shape their own internet and digital policies. But the horse had already bolted. We are in the same position now with AI.

AI 2040 is a detailed policy scenario and recommendations from the AI Futures Project, written by researchers including Ryan Greenblatt from Anthropic. The central argument is that superintelligence be delayed until 2040 with international agreements put in place between the two superpowers, China and US. It also calls for AI research to be done in public and allowing multiple companies to develop safely together.

It outlines potential scenarios along the way including how to handle explosive economic growth. This would occur due to human workers being replaced with AI with none of the overhead costs a human has on businesses, and increased efficiency with the implementation of 24-hour workers with only the data costs and system maintenance as the overhead.

Slowing this implementation as recommended in 2040 is a noble quest. The difficulty is the market, driving implementation. As with the internet, everyone, including teenagers with a Claude code subscription, are trying to cash in on healthcare and AI. There are billions of dollars in this field up for grabs, and it is being co-opted by those outside of it, with little understanding of how the system works, and the infrastructure that supports it.

Technology is good. It has many benefits to save money and staff time. It also has the potential to cost money, a lot if left unchecked, in more ways than just processing and storage costs. It could cost people their lives. That’s not a reason to stop, it’s a reason to govern.

The very first point is to work out what problem are you trying to solve? There are a myriad of ways you can implement AI in medical practices, but without a clear goal of what you’re trying to do, you can easily end up with a patchwork of tools, duplication and inefficiency.

Once you know what you’re trying to solve, the path forward becomes easier.

Let’s look at staff training and onboarding as an example.

Having a dedicated chat interface for a new staff member to ask questions is invaluable. It’s a repository of all of your processes and procedures and the staff member can ask it directly, what do I do when? The answer will come back as any model answers a query, but it is specific to your practice and your practice documents.

Implementing out of the box tools is more successful if all of your processes and procedures are available in one place for the tool to find them. If you’re adding an appointment management tool it will need to know your appointment policy, who gets an appointment and under what conditions. These directives need to be explicit for the tool to work best for your practice.

Before any AI tool can work effectively in your practice, it needs to know what your practice actually does. That sounds obvious. But most practices don’t have it written down. Processes live in the practice manager’s head, in email chains, in the memory of the staff member who’s been there longest. When that person leaves, the knowledge leaves with them.

AI in all its forms, is based on language and data. It processes that language and data back in the form of analysis or answers to direct queries, or orders of function directed by the user. AI systems need to know exactly what you want to do, so it can then go ahead and do it.

For practice integration, nothing can be assumed. Every rule, every exception, every policy needs to be written down before the AI can apply it. The large language models like Claude are already preloaded with information that it draws from. This is what you need to do for your practice, with practice-specific instructions and rules to function.

So where do you start? Pick one problem. Just one. The practice that tries to implement AI scribes, scheduling tools, billing software and a staff training assistant simultaneously will implement none of them well. Choose the single biggest pain point in your practice right now and solve that first.

Let’s look at how to set up a Claude Folder for staff training as an example.

Setting up an AI assistant for staff training

Claude — one of the leading AI models — has a feature called Projects. Think of it as a dedicated folder where you upload your practice documents and then have conversations with an AI that knows only your practice.

Here’s what the setup looks like:

Step 1 — Create a Project
Go to claude.ai and create a new Project. Give it a name — “Practice Assistant” or “Staff Training/Onboarding.

Step 2 — Upload your documents
Upload your process and procedure documents into the Project. Your practice manual. Your billing process. Your referral process. Your phone management guide. Every document you want the AI to know about goes in here.

Step 3 — Give it instructions
Add a brief set of instructions telling the AI its role. Something like — “You are a practice assistant for [Practice Name]. Answer questions using only the documents provided. If the answer isn’t in the documents, say so and suggest the staff member ask their practice manager.”

That last instruction is important. You want the AI to know the limits of its knowledge.

Step 4 — Test it
Ask it questions a new staff member would ask. “What do I do when a patient calls to cancel?” “How do I handle a billing exception?” “What is the DNA process?” If the answers are wrong or incomplete — update your documents.

Step 5 — Share it with your team
Claude allows you to share Projects with team members. Your entire practice can access the same assistant from the same documents.

The result:

A new staff member on day one can ask questions and get consistent answers based on your actual practice procedures, not someone’s memory of them. Your practice manager isn’t interrupted every five minutes. And when staff leave, the knowledge stays.

The only maintenance required will be to update processes and procedures as they occur to keep the model accurate. Staff should be encouraged to follow documented processes rather than working around them. When workarounds occur they become habits and habits that aren’t documented can’t be taught to AI or to new staff.

These are the practical manual steps that lay the foundation for your AI to be effective. The core issue is again, what problem are you solving, how do you want to solve it, how will you measure it?

Implementing AI is not an out of the box proposition. It is very specific to your practice. The more specific, the more helpful and efficient AI can be. The arduous task is to create the documents AI needs to function at its full potential within your practice. If you need something to get started, take a look at my practice procedure and process document package.

Scribes

AI scribes have been rolled out since 2024. Scribes are used to record patient interactions with their medical practitioner, so they can focus more directly on the patient, rather than doing the paperwork at the same time. The consultation notes become part of the patient record after the practitioner signs off and confirms the record is accurate, at the end of the consultation. As at November 2025, approximately 40% of GPs are using this software. That’s a rapid uptake of a new technology, however a survey in 2025 found that more than 1,400 clinicians felt burnout had dropped by roughly a fifth within 3 months of adoption.

Note accuracy is anecdotally said to improve in accuracy with less mistakes, the software is accessible for rural and regional practices, it can reduce after-hours administrative burden and aid the efficiency of the practice.

There are a few different companies offering them but they all face the same considerations.

Patients must consent to their voice being recorded. Intimate details may be exchanged in a consultation and not everyone will be comfortable with this concept. Vernacular and abbreviations need to be considered and checked in the consult notes after the recording for accuracy. What happens to the audio afterward also needs to be explained to the patient so they are fully informed.

Clinical scribes have not been approved by the TGA under medical devices, so there is no mandatory accuracy standards, no post-market surveillance requirement and no regulatory accountability if something goes wrong.

Each consultation, with more extensive notes are still subject to the same data breach risks from a security perspective. Note take gives more lengthy detail than what some practitioners have used as their own short hand and abbreviations which are known in the healthcare arena, but not so much if you were just reading a cold paragraph of notes.

UPDATE 4 AUGUST 2026 - The TGA has announced compliance action on AI scribes citing scope creep, inadequate monitoring and lack of transparency. This validates the governance concerns raised in this piece.

Appointment Management

Practice management software includes appointment AI. In Australia HotDoc and Health Engine have been on market and widely used in Australian by thousands of medical practices. Other software integrate within existing practice software to manage automated recalls, reminders and patient communication.

There are already international services operating in Australia handling patient data. International receptionists, which are AI, answer calls, book routine appointments and will escalate clinically urgent or difficult matters to a human.

For practices this can fill the gap of a reliable receptionist, AI never has a bad day and can handle multiple calls at once. Anecdotally no shows and cancellations have decreased utilizing these systems, with the follow up ensuring patient attendance more efficiently.

These tools need read/write access to patient records, appointment systems and billing systems. They also need clear instructions as to policy of the practice. Many international scheduling tools process data on overseas servers which is a data sovereignty issue for Australian healthcare data. Integration of these types of system sounds great in the marketing but in practice can be problematic and may infringe on existing licences of the existing practice software.

Referral Management for Specialist Practices

Referral management is built into some practice software but it still requires human intervention in triage. This may be due to not enough clear instruction able to be given to the AI to determine when an appointment is necessary from the referral alone.

Some referrals are not just paper. Some are follow up with emails from the referring practitioner with patient specific details, concerns or triage information. This all needs to be linked to the referral and patient record for the AI to make a determination regarding urgency and scope of practice issues.

Billing Systems

The Australian billing software market is growing but critically AI billing tools can’t replace MBS expertise. Even when a clinic uses AI support for billing optimisation, staff still need real knowledge of MBS item descriptors. GPs are over seven times more likely to underbill than overbill. Australian specific expertise is needed for these tools.

Coding systems are slightly different with many international tools deployed to Australia. Coding is based on international coding requirements so it is easier to implement against clinical data.

AI systems can detect underbilling in retrospect but have not been able to capture it at time of appointment thus far. It’s not far away, but it will need full notes completed by the practitioner for the AI to read and determine length and complexity of the consultation to determine the correct item for general practitioners and medical specialists.

Every major AI billing tool in the world is built for the US Current Procedural Terminology coding system. The MBS has approximately 5,700 items with fund-specific interpretation rules that vary between Bupa, Medibank, HCF and others. No commercially available AI billing tool is built specifically for the Australian MBS with Australian fund-specific rules. The tools being marketed to Australian practices are either generic or US-built. AI billing tools need a practice specific structured workflow to function correctly. Without it they bill from incomplete information and generate more exceptions not fewer.

Many business are eager to enter the healthcare industry with AI assisted tools. There are some really great ideas to improve practice efficiencies, but not all of the ideas can be easily implemented onto existing infrastructure and rules of medical practice under Medicare and Private Health Fund rules and processes.

AI can help a practice in other ways, specifically staff training and onboarding, process and procedures for the practice, and administrative workflows. AI needs to understand what to do and how to do it, before you tell it to do anything. AI needs to be trained to then train the staff and other software.

Before any of these tools can work safely in an Australian specialist practice, the practice needs to know, and have documented, what it actually does. How referrals are triaged. How billing exceptions are managed. How patient communication is handled. How aged debtors are followed up. Without that foundation, AI amplifies chaos rather than creating efficiency

It’s easier than you think to implement.

Practice Management Document Pack — AI-ready foundations for your practice. Available at virtuallyapractice.com

Fine — examined the security of the data. It isn’t.

Great — examined uses and misuses. Both are significant.

Free — examined the cost and who takes the profit. Not the same people.

Legal — examined liability. Nobody has it.

Australian Healthcare AI is not in a bubble it’s in a giant pickle jar. Each of those areas are their own pickles which all need to be dealt with as we move forward in applying AI to healthcare in Australia.

The data is not fine. Security still remains an issue with the Medibank breach in 2022 given as an example of what could happen to patient records. Not only was the company held to ransom, but patients could have been exploited as well with threats to reveal sensitive health status.

The data is great but not for the patient or citizens of Australia. Collected from presentations for treatment as a byproduct of that treatment, then used by insurers and Government, some for good, but a lot for bad.

The data is not free. It costs companies and hospitals money to process and store, and those costs are passed on to the consumer in increased premiums and reduced rates of payments.

The use of the data is not defined as legal or illegal at this point in time. That’s the problem. Collection, analysis and application goes on regardless.

The answers are not simple. Government and private enterprise share the system but not the same incentives. The patient funds it all and controls none of it.

Private enterprise incentive is to make a profit. This is fine. It helps support the economy and provides services to Australians who need them.

The Government incentive is to spend Australian taxpayer funds wisely to supply essential services to the population they govern, and to make sure private enterprise operates within the laws of Australia.

The tension is between the money private enterprise stands to make from health data and the Government’s obligation to ensure that profit doesn’t come at the expense of the patients who generated it.

At the moment we are in no-man’s land. Private enterprise is profiting from the data and the Government has not taken timely action. It’s not just healthcare that is affected by this.

The confidence that Australians have in their Government is being eroded. The faith they have in the health system to be there for them, also eroding. Trust in anything is diminishing. Hope has left the conversation, driven away by demonstrated inaction.

The question now turns to Australians themselves.

Australians need to decide and act upon how they want their healthcare data handled, who can use it and in which capacity, how much data is collected, and whether or not they can financially benefit, or prevent others from doing so. This requires the Government to intervene in a timely manner and not allow markets to roar away with the profits in a deregulated environment.

The pickle jar is sealed and the pickles are marinating in brine, bubbling away, maturing, gaining flavour and becoming more valuable as they age. The label reads fine.

It isn’t.

The Australian Government’s report Safe and Responsible Artificial Intelligence in Health Care -Legislation and Regulation Review Final Report, was published in March 2025.

“Our current regulatory system is not fit for purpose.”

At the time, there was no Australian body dedicated to overseeing AI in healthcare. The Office of AI announced in July 2026 should fix this, except there was no mention of healthcare in the announcement and doesn’t kick in until 2027. So that’s a two year gap in action from the recognised problem to a possible fix.

It also found there was a lack of evidence to support potential benefits of AI in healthcare. Let that sit for a minute as we move on.

The Government acknowledged there was a need to clarify who owns patient data, but hasn’t done anything about it yet.

Finally, there was no incentive for industry to deliver quality AI technologies for Australian needs.

Concerns raised about data included:

38% of Australians trust that companies using AI will protect their personal data.

54% of respondents said healthcare information should be kept in Australia.

The consequences of mishandling health data “may be long lasting and irreversible, leading to the potential for mental and physical harm and, in some cases, death.”

Some respondents said patient data is impossible to deidentify. Skin scans. Genetic data. 3D CT reconstructions that can reproduce a person’s face.

The document explicitly lists billing as one of the areas AI is already impacting Australian healthcare. It also noted the department should ensure statutory frameworks including the

Professional Services Review Scheme (PSR) and compliance with billing requirements. That’s the government acknowledging AI in billing needs governance, and then not providing it.

Clinical Liability

The question of who is liable if an AI misdiagnoses a cancer is currently unresolved in Australia. In most jurisdictions the radiologist signs the report and is legally responsible, but they didn’t build or train the AI. The software developer may carry product liability, but they didn’t make the clinical decision. The hospital deployed the system, but they relied on regulatory approval.

Traditional civil liability frameworks may not fit and legal reforms are needed to address accountability for AI-related harm.

Existing legislation that could apply are:

Privacy Act 1988 — APPs: Governs collection, use and disclosure of personal information. Applies to health data. But was written before AI existed and doesn’t address AI-specific risks like model training on patient data without consent.

TGA Medical Device Regulation: AI diagnostic tools that meet the definition of a medical device must be TGA approved. But general purpose AI, like a large language model used for clinical documentation, may not be classified as a medical device and therefore evades TGA scrutiny entirely. The I-MED/Annalise.ai situation sits in this grey zone.

Australian Consumer Law: Could apply if an AI healthcare product causes harm but the causation chain between AI output and patient harm is legally complex and untested in Australian courts.

Civil Liability Acts — state based: Each state has its own civil liability legislation. The professional standard defence, that a professional is not negligent if they acted in a manner widely accepted by peer professional opinion, may protect a radiologist who followed standard practice even if the AI was wrong. But it doesn’t resolve who pays.

Collectively, there is no answer to the question, and no-one is liable for harm.

Data from 2024 showed a 14% increase in malpractice claims involving AI tools compared to 2022 in the US. The majority stemmed from diagnostic AI used in radiology, cardiology and oncology. Missed cancer diagnoses by machine learning software have become a central focus in several high-profile lawsuits. Australian law will follow similar patterns as AI deployment accelerates here.

Planning Liability

When AI models inform government or health system decisions about resource allocation such as bed numbers, hospital closures, service distribution, workforce planning, and those decisions cause harm, the liability chain is even more diffuse than clinical AI.

The scenario: Government uses aggregated claims data fed through an AI model to recommend closing beds in a regional area. Beds close. Demand spikes. Patient presents to emergency, waits too long, suffers harm. Who is liable?

The AI vendor, provided a tool, not a decision. The government agency made the decision but relied on AI recommendation. The hospital implemented the decision but had no choice. The Medicare data was used but they didn’t make the recommendation.

Australian administrative law provides some framework for challenging government decisions but AI-assisted government decisions are largely untested. The Administrative Decisions (Judicial Review) Act 1977 allows review of federal administrative decisions but was written before AI existed. Proving that an AI recommendation caused a specific harm requires establishing causation that courts haven’t yet tested.

Administrative Liability

When a private health fund’s AI system incorrectly rejects a legitimate claim who is liable for the consequences? The fund deployed the AI but says it followed its contract terms. The AI vendor built the model but says the fund configured it. The hospital submitted the claim correctly but bears the cashflow consequence. The patient loses access to care but has no direct legal recourse against the AI system that rejected their claim.

The PSR has jurisdiction over inappropriate billing by practitioners, not over inappropriate rejection by funds. The Private Health Insurance Act 2007 governs fund conduct but doesn’t address AI adjudication specifically. The Privacy Act governs data handling but not payment decisions. The Australian Consumer Law could theoretically apply but no case has been brought and the causation chain is complex.

There is currently no legal mechanism in Australia for a hospital or provider to challenge an AI-generated claim rejection on the grounds that the AI was wrong.

Who is liable for a misdiagnosis from an AI model? Still nobody.

Who is liable when AI resource planning closes a hospital and a patient dies on a waiting list? Still nobody.

Who is liable when an AI claims adjudication system incorrectly rejects thousands of legitimate claims and contributes to a hospital’s financial collapse? Still nobody.

The government’s own March 2025 Final Report acknowledged the regulatory system is not fit for purpose. The Office of AI was announced in July 2026 with no health mandate. Standards are expected in draft in early 2027. In the meantime, AI continues to be deployed across clinical diagnosis, health service planning and payment adjudication all without a clear legal framework for accountability when it goes wrong.

Recommendations

1. Mandatory pre-market approval for AI used in clinical diagnosis extending TGA framework to cover general purpose AI used in healthcare contexts.

2. A national healthcare AI oversight body as 70%+ of consultation respondents requested and the government has not created. The Office of AI did not include healthcare in its statement.

3. Clear liability legislation for AI-assisted healthcare decisions specifying the chain of responsibility between AI vendor, deploying organisation and affected patient or provider.

4. Specific administrative AI governance. The Private Health Insurance Act must be amended to address AI claims adjudication, transparency obligations and accountability when AI generates incorrect rejections.

5. Data ownership legislation, clarifying that patients own their health data and establishing consent requirements for AI training on that data.

The law hasn’t caught up with the technology. The technology isn’t waiting for it.

What does it cost to keep the data?

Total Australian health expenditure was $295.3 billion in 2025-26. Medicare alone was $32 billion. The remaining $263.3 billion covers hospitals, aged care, community health, medications and the digital infrastructure that supports it all — an undisclosed proportion of which is data management and storage.

Private health insurance management expenses:

10.9% of contribution income in 2024-25, approximately $3.3 billion annually just in management costs. Those management expenses have grown 51% in six years while hospital benefit payouts grew only 18%.

What’s in those management expenses? IT systems, data infrastructure, claims processing systems, the very systems collecting and storing 650 million health data transactions annually. The Australian taxpayer and the premium payer are funding the collection and storage of data that private companies then use for their own commercial purposes.

Storing Australian health data is not free. The Australian healthcare cloud computing market reached $1 billion in 2024 which is the infrastructure cost of managing the digital health records, imaging files, claims data and transaction histories of 27 million Australians. Medical imaging alone generates petabytes of data annually. A single MRI produces up to 4 gigabytes, a PET scan up to 3 gigabytes, and I-MED performs over 6 million procedures each year.

The three major providers charge for Australian region storage:

AWS S3 Sydney region — $0.023 per GB per month for standard storage. $0.004 per GB per month for archival storage (Glacier).

Azure Blob Storage Sydney — $0.018 per GB per month for hot storage. $0.004 per GB per month for archive.

Google Cloud Sydney — approximately $0.020 per GB per month standard.

Each billing record is tiny, under 1MB. Even at 1MB each that’s 650 terabytes of administrative data. At AWS standard rates — $0.023 per GB × 650,000 GB = approximately $14,950 per month or $179,000 per year for the raw transaction data.

But this is only current year data. Medicare has been collecting data since 1984 — 40 years of records. At growing volumes the accumulated administrative health data archive is potentially petabytes. At archive storage rates ($0.004/GB) — 1 petabyte of archive storage costs $4,000 per month or $48,000 per year.

Medical imaging is where the costs explode. I-MED alone performs 6 million procedures annually. At an average of 500MB per image (conservative across x-rays, ultrasounds, CT and MRI) that’s 3 petabytes of new imaging data per year from one network alone.

At AWS standard storage, 3,000,000 GB × $0.023 = $69 million per month. That’s not what they’re paying. Large enterprise deals get 70-80% discounts from list pricing. Realistic enterprise pricing would be approximately $0.005-$0.008 per GB for large healthcare providers.

At $0.006 per GB for 3 petabytes = $18 million per year for I-MED’s annual imaging volume alone. That data must be retained for 7-10 years under Australian law.

Who pays these costs?

The patients. Medicare infrastructure is funded by the taxpayer. Private health fund storage costs are embedded in management expenses that represent 10.9% of premium revenue passed directly to members through their annual premiums. Hospital and practice systems are funded by the providers themselves. The cost of collection and storage is borne by patients and taxpayers. The commercial value flows elsewhere.

How much can you make on the data?

The data collected through 650 million annual health transactions is not just clinically valuable. It is commercially valuable. The Australian healthcare big data analytics market was valued at $1.5 billion in 2025 and is projected to reach $4 billion by 2034. The broader digital health market, built substantially on the same data, reached $8.9 billion in 2025 and is projected at $31.1 billion by 2034.

Claims and financial data represents 45.6% of the global health data monetisation market. The administrative billing data that Australian practices and hospitals generate as a byproduct of treating patients is the single largest commercial category in health data monetisation globally.

Private capital investment in Australian healthtech hit $416 million in 2025, a 101% increase in a single year. Smart money moves toward value. It is moving very fast into Australian health data.

The Australian patient funds the collection of this data through Medicare levies and insurance premiums. They receive healthcare in return. What they do not receive is any share of the commercial value their data generates, any transparency about how it is used, or any meaningful consent to its commercialisation.

Where is the data exactly?

Amazon Web Services, Azure and Google Cloud operate Australian data centres. The data is physically in Australia. But the parent companies are US corporations subject to the US CLOUD Act, which allows US authorities to compel access to data stored anywhere in the world without necessarily notifying the Australian government or the data subject. Choosing the cheapest cloud option such as the US hyperscalers, are typically 30-50% cheaper than Australian sovereign cloud providers. It may save money while quietly surrendering data sovereignty.

Healthcare data is collected as part of an admission or presentation for treatment. It is par for the course. What becomes of it, has become a source of enormous commercial value with profit in both the selling and analysis of the data. The cost to store and disseminate the data is borne by the patient as both a taxpayer and client of the funds. Potential financial benefits of utilising this data, i.e. selling it, are not passed onto the consumer in reduced premiums. Just the opposite is true as funds and the government expand IT infrastructure to store and process it all.

The commercial beneficiaries are not hard to identify. The cloud providers AWS, Azure and Google, charge billions annually to store the data patients generate. The health technology companies build products from patterns in that data. The life insurers use it to price risk. The pharmaceutical companies use it to identify patient cohorts for trials and marketing. The private equity firms investing $416 million in 2025 are buying access to the data flywheel.

The more data, the better the AI, the more valuable the platform.

The data was free, but the costs to keep it are enormous and growing. The public is paying both monetarily and with their sovereignty.

I have previously examined the volume of healthcare transactions, but as a guide for this article it’s 650 million combined transactions a year, between Medicare and private health insurers. That’s a lot of information.

Item numbers alone can reveal what type of pathology tests were performed, potential diagnosis, type of care required, and the potential long-term cost to the healthcare system, and the insurance risk.

Life insurers in Australia were legally using genetic test results in underwriting decisions until very recently. A 2019 industry moratorium, voluntary and self-regulated, was found to be largely ineffective. Some life insurers were simply ignoring it. No government oversight existed to check compliance.

On 1 April 2026 the Australian Parliament finally passed legislation banning the use of genetic test results in life insurance underwriting. The Act received Royal Assent on 8 April 2026 and takes effect 8 October 2026. So it’s still fine until then.

Genetic information derived from Medicare pathology claims item numbers in the genetics category was being used by life insurers to discriminate in underwriting. The practice was legal until three months ago. The ban only applies going forward. Existing policies written before October 2026 are not covered.

The Council of Australian Life Insurers and KPMG published “Australia’s Mental Health Check Up” in November 2024 using actual insurance claims data from 10,218 mental health claims across 2013-2022.

The data showed white collar workers are 7% more likely to be permanently disabled and 14% more likely to be temporarily disabled due to mental health conditions than blue collar workers. Men are 59% more likely to be permanently disabled due to mental health conditions than women.

This is insurance industry data being used to understand and price mental health risk. The same data that flows through Medicare and private health claims is informing life insurance and income protection pricing. Health insurer and life insurer organisations overlap in ownership.

Under the Disability Discrimination Act 1992 it is unlawful for an employer to discriminate based on disability or health condition including mental health. However, an employer crosses the line if they demand specific diagnoses, contact doctors directly or use health information to discriminate in promotion or hiring decisions.

The documented reality is that the Australian Human Rights Commission reported that 46% of all complaints in 2022-23 were under the Disability Discrimination Act. The law prohibits it. The discrimination happens anyway.

AIA Australia owns both health and life insurance products. Bupa offers both health and life insurance products. If the same parent company processes both health claims and life insurance applications the question of data separation, and whether health claims history influences life insurance decisions, is a legitimate governance question with no publicly documented answer.

There is no opt out option. The data is provided on application. The private companies have no obligation to share any of their data but use it for their own calculations and decisions. The same data that raises these concerns also holds genuine promise.

Healthcare services and planning is one of the better uses of data. Data can detect disease outbreaks and assist in early interventions, demand increases or decreases for planning of infrastructure like hospitals and clinics. The data can assist in workforce planning and training needs.

The Australian Institute of Health and Welfare explicitly uses health data for population health planning, long-term policy development, health equity monitoring and research. Their 2024 report identifies AI, machine learning and natural language processing as key future tools for automating medical coding, health planning and service delivery forecasting.

At Macquarie University’s Centre for Health Informatics, the same institution producing Australia’s national AI in healthcare roadmap, researchers analysed more than 12 million hospital episodes over nearly two decades to investigate the relationship between delirium and dementia. The finding was striking: delirium substantially increases dementia risk, with each additional episode raising it further. Around 40% of delirium cases are potentially preventable using interventions already available. That finding came not from expensive purpose-built research infrastructure but from routinely collected administrative health data. Data that already exists. Data that is already being collected. Data that, governed well, could transform how Australia prevents its greatest public health challenges.

Clinical applications are also seeing positive results from utilising health data. AI-powered diagnostic tools at the Royal Melbourne Hospital are helping radiologists detect early-stage cancers, reducing diagnostic errors and improving patient outcomes. These findings are based on collation of previous health data and images. But are we allowed to?

AI trained on millions of diagnostic images can detect cancers earlier, reduce diagnostic errors and improve outcomes for patients who might otherwise wait months for a specialist review. Australia’s largest diagnostic imaging network, I-MED Radiology, pursued exactly this goal, establishing a joint venture with Harrison.ai to develop Annalise.ai, a diagnostic AI trained on patient imaging data.

In September 2024 the OAIC initiated an inquiry into allegations that I-MED had shared patient chest x-rays to train the AI without first obtaining patient consent.

This is the tension at the heart of clinical AI. The technology works. The benefit is real. The data it needs to work is intimate, personal and in Australia belongs to a patient who may not have known their chest x-ray was being used to train a commercial AI model.

The Royal Australian and New Zealand College of Radiologists is now drafting consent guidelines specifically for AI use in radiology. They didn’t exist before.

Most Australians would support such use of data, the finding of clinical truths that could help humanity. The problem though is that it can be used just as easily for detrimental purposes as demonstrated with the life insurance application of data.

The speed at which AI is being applied to clinical software, diagnosis tools and patient data applications has to date been done without governance or oversight. The Government’s proposed Office of AI did not include these as priorities in the recent statement from the Prime Minister.

The governance conversation is lagging behind the commercial reality. And the commercial reality involves significant money. Who is making it?

In my previous article I examined a small amount of patient data that is handled annually, 475 million medical transactions by Medicare. Assuming the Medicare data is safe, let’s now look at the private patient data.

For the same period as the Medicare data, 2024/2025, private hospitals had 5.1 million episodes of care. Each episode may involve 3-5 separate claims transactions, hospital accommodation including theatre, surgeon, anaesthetist, pathology, or imaging, so approximately 15-25 million transactions annually.

Extras claims, dental, optical etc. are approximately 30-45 million claim transactions.

So the private health sector processes approximately 150-200 million health data transactions annually.

Who holds the data?

Private health funds are a mix of not-for-profit Australian held, to international companies operating in Australia. Some are listed on the ASX and could be vulnerable to foreign ownership.

When a foreign-owned fund deploys AI to adjudicate Australian health claims, whose AI model is it? Is it developed in-house by the Australian subsidiary or by the parent company overseas? If Bupa Australia deploys AI built and maintained by Bupa Group in London, the model was designed, trained and governed outside Australia. Australian regulators have no visibility over how it was built, what data it was trained on or what decisions it is optimised to make. The Australian patient whose claim is rejected by that model has no way of knowing whose system made the decision.

There is no law preventing a foreign entity from acquiring majority ownership of an Australian private health insurance fund. The Foreign Investment Review Board can review and potentially block such acquisitions on national interest grounds but it is a discretionary assessment, not a prohibition.

The Private Health Insurance Act 2007 governs how funds must operate but says nothing about who can own them or where their data must be processed.

Bupa, processing 43 million Australian health data transactions annually, operates under a UK parent company structure that entered Australia through Foreign Investment Review Board (FIRB) approval granted years before AI was deployed in claims adjudication.

Health funds listed on the ASX are potentially vulnerable to takeover. If this occurs, the largest health funds, holding most of Australian healthcare data, could be in foreign ownership and not managed, or protected on shore. ASX listed funds are theoretically vulnerable to foreign acquisition. But Australian data is safe regardless of ownership isn’t it? In October 2022 Australia received its answer.

The Medibank data breach — what happened, how and the full impact:

The scale:
9.7 million current and former Medibank, ahm and international student customers affected. One of Australia’s largest and most damaging data breaches ever. The data included names, dates of birth, addresses, phone numbers, email addresses, Medicare card numbers, passport numbers and most critically, health claims information including diagnoses, procedures and treatment details.

How it happened:

A contractor saved his Medibank username and password to his personal internet browser profile on a work computer. That computer was infected with information-stealing malware on approximately 7 August 2022. The malware stole the saved credentials. Those credentials were then used to access Medibank’s VPN, which at the time required no multi-factor authentication. Just a username and password.

The attacker authenticated into Medibank’s VPN on approximately 23 August 2022 and had access to most if not all of Medibank’s systems for nearly two months, from August to October 2022.

The security alerts that were ignored:

Medibank’s own endpoint detection and response software generated alerts on 24-25 August 2022, the day after the attacker first accessed the network. Those alerts were sent to the IT Security Operations email inbox but were not appropriately triaged or escalated. The attacker continued operating undetected for weeks.

What the OAIC investigation found:

Medibank was aware of serious deficiencies in its cybersecurity framework for at least 18 months before the breach. A Datacom report in mid-2020 identified the lack of multi-factor authentication as a “critical defect.” A KPMG report in August 2021 also found MFA was not in place for privileged users. Medibank was warned. Twice. By independent auditors. And did not fix it.

The OAIC filed civil penalty proceedings against Medibank in the Federal Court alleging Medibank seriously interfered with the privacy of 9.7 million Australians by failing to take reasonable steps to protect their data.

What happened to the data:

The attackers initially demanded US$10 million. They later revised it to US$9.7 million — roughly US$1 per affected customer. Medibank refused to pay the ransom. The attacker, subsequently identified as Russian national Aleksandr Gennadievich Ermakov, sanctioned by Australia, the UK and the USA, published the stolen data on the dark web in multiple tranches beginning November 2022.

The files were labelled with devastating specificity “good-list” and “naughty-list.” The naughty list contained sensitive health information about customers who had claimed for treatment of HIV, alcohol addiction, drug addiction and mental health issues. A separate file labelled “Abortions” contained details of abortion procedures claimed by customers. The attacker ended the dump with “Case Closed”.

The human consequences:

The release of personal health information on the dark web exposed 9.7 million Australians to the likelihood of serious harm including emotional distress, identity theft, extortion and financial crime. People whose HIV status, mental health treatment, addiction history or reproductive health decisions were published on the dark web by a Russian criminal had no recourse and no warning before it happened.

The legal aftermath:

Federal Court civil penalty proceedings filed by OAIC. Consumer class action filed in Federal Court. Legislation subsequently passed increasing maximum fines for repeated privacy breaches to $50 million. APRA directed Medibank to allocate additional capital as a consequence of the breach. Medibank is defending the OAIC proceedings.

OAIC has still not determined penalties as at July 2026.

Australian healthcare data is not safe, it’s not necessarily on Australian soil, and even if it is, it uses foreign infrastructure which can be utilised and analysed by others. There is no transparency surrounding what AI systems are already in use determining healthcare claims and collection and storage of Australian health data.

But I’m sure it’s fine.

In his speech, the Prime Minister highlighted that the Government would introduce a set of standards for AI use in Australia. These are expected in draft early 2027 after the government considers the proposed announcement in August. In the meantime, AI roars ahead and there is no telling where we will be by next year.

In the statement, the Prime Minister focussed on data centre establishment and protecting copyright for Australian artists and media. Then the Australian consumer. No mention of healthcare or any industry sectors, or patient data.

Leaders in healthcare have voiced their concerns regarding governance, liability, feasibility and data protections in response to the Government announcement.

Governments and agencies are setting up multiple AI committees and sub-committees with parallel work underway in the non-government sector. This is leading to fragmentation of policy and application across a multitude of industries, not just healthcare.

AI has already been deployed without adequate monitoring of what it actually does in practice.

Listening to marketing chatter, you would believe that every practice needs AI in their clinical practice immediately. Clinical note taking – done by AI. Filling in forms – done by AI. Diagnosis of imaging – done by AI. Radiologists? Entire career trajectory pronounced DOA.

The proof of day-to-day application of these functions is missing, it’s presumed to be awesome, correct at all times, and another billion dollar opportunity.

AI is already widely implemented in Australian industry and business including healthcare. The horse has not only bolted, it has picked up a pack of brumbies and are tearing up the mountain, laying claims and building houses.

The clinical AI conversation cannot wait for standards in 2027. Neither can the administrative AI already deployed in healthcare billing and payment systems. Both are live. Both are consequential. Both are ungoverned.

Health data is among the most sensitive personal information that exists. When AI systems are deployed in healthcare, clinical or administrative, they process that data. The question of what those models do with the information, how it is stored, who has access, whether it leaves Australian jurisdiction and what happens when there is a breach is a legitimate governance question that the Office of AI has not addressed.

Every Medicare transaction contains a patient’s identity, their diagnosis, the procedure performed, the provider who treated them, their insurance status and the financial details of their care. Multiply that by 650 million transactions annually and you have the largest concentration of sensitive personal data in Australia which is processed digitally, touched by AI systems, and governed by frameworks designed before AI existed. Patients have a right to know when AI is processing their health data, what it does with that information, whether it leaves Australian jurisdiction and who is accountable if it is wrong. Currently they have no such assurance.

The AI systems processing Australian health data are not all Australian. The models being trained on patterns in Australian billing, diagnosis and treatment data may be owned, operated and housed overseas. Australia has no mandatory requirement for health administrative AI to be sovereign, transparent or subject to Australian law. The Croakey series on AI governance raised the spectre of health data colonialism, patient data collected under the banner of efficiency, processed by foreign models, generating commercial value for overseas interests. Australian health data is among the most valuable data in the world. Its governance cannot be an afterthought.

The Royal Australasian College of Physicians has already warned that regulatory gaps in clinical AI leave doctors exposed to liability for patient harm they did not cause. Nobody has asked the equivalent question about administrative AI.

When a private health fund’s AI system incorrectly rejects a legitimate claim who is accountable? When AI trained on ambiguous MBS rules generates systematic errors across thousands of claims simultaneously, who answers for the consequences?

Governance of AI in healthcare cannot begin and end at the clinical layer. The financial infrastructure that makes clinical care possible deserves the same scrutiny, the same transparency obligations and the same accountability frameworks. Without it, 650 million health data transactions will continue to flow through ungoverned systems.

The Office of AI’s mandate must include healthcare, not as an afterthought but as a priority. Within healthcare, it must look beyond the clinical layer to the financial infrastructure that makes clinical care possible and the safety of patient data.

AI in healthcare is not a hot topic for future consideration. It is a cold reality operating right now.

Building From Scratch

As with building a new house, land is always the first issue. For a commercial private hospital this is problematic. It needs to be close to an established community or be planned and built as part of the infrastructure. To build a new private hospital in established urban areas requires acquisition of a large parcel of land, or property that could be repurposed.

Let’s put a pin in that but allow for a conservative $5 million investment.

Now come the regulations. Each state has its own rules in this regard, so this is just an overview of the labyrinth.

In NSW the process requires an Approval in Principle under the Private Health Facilities Act 2007 before a single brick is laid. That approval requires architectural plans assessed against the Australasian Health Facility Guidelines, fit and proper person assessments of all directors, public consultation with third party submissions, Local Health District consultation and compliance with 19 prescribed classes of health services, each with separate licensing requirements.

The Approval in Principle is only valid for twelve months. Extensions can be sought but the clock is always running.

Estimates suggest a new private hospital bed costs approximately $800,000 to $1.2 million to build and equip. A modest 50 bed hospital represents $40-60 million in construction alone before equipment, staffing, systems and working capital. A hospital the size of Canberra Private with 38 beds and 4 theatres would cost $50-80 million to replace from scratch today.

Even after building the hospital, you need contracts with private health funds before patients can use their insurance. Those negotiations can take 12-24 months and funds are under no obligation to contract at rates that make the hospital viable. The same dynamic that contributed to Healthscope’s collapse would face any new entrant.

Surgeons admit to hospitals where they have existing privileges, equipment they know and teams they trust. Building a new hospital doesn’t automatically attract surgical lists. Credentialling takes months. Building a surgical community around a new facility takes years.

Buying Established Hospitals

The Healthscope collapse saw only 5 facilities being bought by other hospital operators. The remaining 31 are now being transitioned to not-for-profit models. Buying established facilities saves a prospective operator up to $80 million if they were to build, but no buyers were procured for 31 of the facilities. It wasn’t because they were unprofitable, it was because the lender could not accept the offer.

The remaining hospitals in the portfolio shared $1.6 billion debt. The sale-leaseback structure left 23 of the 31 hospitals not owning their buildings and carrying up to 40% of revenue going to rental payments. The lease obligations were part of the acquisition and could not be renegotiated.

Health fund contracting uncertainty also made potential buyers nervous having witnessed what could happen to their profitability as part of the Healthscope collapse.

When a private hospital closes the community doesn’t just lose beds. It loses something that took decades to build — surgical communities, specialist relationships, trained nursing teams, established fund contracts, trust from the community who utilise the facility.

Replacing that is not a matter of building a new building. It is a generational undertaking measured in years and hundreds of millions of dollars.

Sector Impact

The 70 hospitals that have closed in five years are not being replaced. The services that disappeared from Darwin, Hobart and Tasmania are not coming back quickly. When Canberra Private closes in September 2026 the 38 beds and 4 theatres don’t simply reappear somewhere else.

Health funds could lose approximately $43-45 million in annual claims payments from that facility, based on modelling. That sounds like a saving. It isn't. Those patients don't disappear. They present at other private hospitals, or they join public hospital waiting lists, undermining the value proposition of private health insurance for the ACT community entirely.

Policy is focused on private health fund rebates and the implementation of AI in healthcare. Both matter. But neither addresses the fundamental question of who will build and sustain the physical infrastructure of private healthcare when the financial model is broken?

Seventy hospitals have closed in five years. The market found buyers for five of Healthscope’s thirty-seven. The rest required charity conversion or government intervention to stay open.

The answer to ‘what happens if neither building nor buying is possible’ is already visible. Services disappear. Communities lose access. The public system absorbs what it cannot afford to absorb. And no amount of AI efficiency or rebate policy reversal rebuilds a surgical community that took thirty years to grow.

You can’t just build another one.

The focus of the roadmap, AI deployment outpacing governance in Australian Healthcare, is not unknown and is a global problem with AI being applied widely with technical abandon. While the recommendations in this report are considered, the underlying driver of healthcare viability is not mentioned.

The priority areas in the roadmap are Leadership, Governance and safety, Sovereign capabilities, Adoption, Workforce, Consumers, Industry, Research and Development.

The main concerns under these areas are patient data privacy, ethical guidelines, accreditation standards for healthcare, Australia specific AI that does not rely on international infrastructure or models with the aid of the National Government, utilising AI to identify and evaluate services, workforce training requirements and training programs within the clinical and digital health training programs, a National Chief AI Officer to be created, educate consumers and ensure specific groups are not disadvantaged, the establishment of Centres of Research Excellence in Healthcare AI and lastly funding for academia to progress it all.

The critical piece still missing in the discussion is the impact AI will have, and is already having, on medical billing and receipt of payments. While the Roadmap’s authors were consulting with 96 organisations and 20 peak bodies, private health funds were quietly deploying AI to adjudicate claims against the MBS. Nobody asked the billers.

The healthcare system runs on the flow of money through it, like any business. AI is already deployed across the payments system and the result is an increase in rejected claims to providers. While this strategy works for the health funds, it can have a devastating impact on providers and private hospitals.

I’ve examined this impact in prior pieces but for context, accepted claims via ECLIPSE are paid within 15-45 days. Rejected claims requiring manual resubmission take 45-90 days to process plus a further 7-10 days for payment, a total of up to 100 days from the original claim date. As AI-generated rejections increase, this cashflow delay compounds across hundreds of claims simultaneously, placing enormous pressure on private hospitals already operating on margins that Healthscope demonstrated can evaporate entirely.

AI could be deployed to more rapidly identify providers whose billing patterns exceed 99% of their peers, the key metric that the Professional Services Review uses.

The underlying issue remains, no matter what type of AI is deployed.

The current IT infrastructure needs to be upgraded, and the MBS itself reviewed for language. Humans cannot even agree on its application, tested in the courts for fifty years without resolution. AI is being tasked with overriding humans in this regard which will, and has, led to delays in payments to providers already, and it is only partially applied to the system as a whole.

The Roadmap raises important concerns about patient data safety, workforce training and ethical guidelines, and those concerns are valid. But healthcare administration has been left out of the conversation again.

There will always be sick people. But without careful attention to how AI is affecting the flow of money through the private healthcare system, the claims, the rejections, the delays, the margins, there may not be private hospitals left to treat them. The Healthscope collapse was not a clinical failure. It was a financial one. AI accelerating claim rejection into an already stressed payment system is not a governance gap the Roadmap’s authors can afford to ignore.

A consideration for future iterations of this Roadmap could be considered: AI deployed in medical billing, claims processing and payment adjudication should be subject to the same governance frameworks, transparency obligations and post-market surveillance being proposed for clinical AI. The people processing the bills need a seat at the table.

In 2024-25 Medicare expenditure increased by $2.82 billion in a single year, a 9.5% increase year on year. The government’s proposed PHI rebate saving is $482 million. The annual growth in Medicare expenditure is six times the proposed saving. In other words the government will save less in the next decade from the rebate cut than Medicare will spend in additional costs in the next two months.

The Professional Services Review (PSR) is the body which investigates practitioners billing to Medicare. In 1992-93 the Australian National Audit Office reviewed the Medical Services Committees of Inquiry, the standing committees appointed by the Minister to deal with over-servicing by medical practitioners. The ANAO’s report titled “Medifraud and Excessive Servicing: Health Insurance Commission” found the committees did not deal with over-servicing effectively. Key concerns were that the MSCI did not discourage over-servicing and that payments recovered from practitioners were smaller than total payments made to doctors who over-serviced.

In 1994 in response to that report and in consultation with the medical profession the government abolished the MSCI and replaced it with the PSR Scheme.

So it has been in operation for 32 years.

The PSR Directors Update for June 2026 recorded twelve agreements with practitioners identified as billing Medicare items in excess of 99% of their peers, and these practitioners were ordered to repay money to the government and in some cases restricted from using certain Medicare items. The full details are here https://www.psr.gov.au/case-outcomes/psr-directors-update-june-2026

The PSR resolves approximately 100 cases per year. Medicare processes 475 million services per year. That is one PSR outcome per 4.8 million services. The June 2026 Directors Update shows the trigger being practitioners billing items in excess of 99% of their peers. The mechanism exists. The scale is not proportionate.

In 2022/2023 financial year, $21.5 million was recovered. This is against a total of $29.5 billion in Medicare expenditure that year. Recoveries represent approximately 0.07% of Medicare expenditure — less than one tenth of one percent.

The government’s proposed saving of $482 million is more than 20 times the amount of PSR recoveries for a single year. Let’s examine Medicare usage more closely.

Between 2009-10 and 2024-25 Medicare expenditure grew 110%. The Australian population grew approximately 23% over the same period. Patient numbers grew 23%. Services per patient grew 26%. But benefits per patient grew 71% — from $790 to $1,353. Understanding what is driving that expenditure growth, and whether current oversight mechanisms are proportionate to a $32 billion annual spend, is a legitimate and urgent policy question.

For the 2024/25 year Medicare reported a National item usage rate of 475,211,213 items billed across 23,936,295 patients. This is 88% of the population of Australia. Services per patient billed 19.9. That’s 20 medicare items billed to 88% of the population. In practical terms, the population of Perth and Adelaide, did not use Medicare in 2024/2025 year but every other breathing person in the country was billed 20 Medicare items.

To be clear, 20 Medicare items per patient per year includes all GP visits, pathology, imaging, specialist consultations and allied health. For many Australians this is entirely routine. The question is not whether 20 items is too many, it is whether the growth from 15.8 items in 2009-10 to 19.9 items in 2024-25 is explained by population ageing and clinical need alone, or whether the proportionality of oversight has kept pace with the growth

The PSR currently resolves approximately 100 cases per year at an average recovery of approximately $215,000 per case. At current capacity the PSR recovers approximately 0.07% of Medicare expenditure annually. If AI-assisted peer benchmarking identified and prioritised cases more efficiently, and the PSR’s capacity was expanded from 100 to 300 cases annually, annual recoveries could reach $65 million. Scaled to 1,000 cases with AI-assisted investigation support, recoveries could approach $215 million close to half the government’s proposed PHI rebate saving, without cutting a single pensioner’s health insurance. The PSR has been operating for 32 years. Medicare expenditure has grown 110% in 15 years. The referral rate and investigation rates have not kept pace.

The Australian Government is proposing to save $482 million by cutting private health insurance rebates for 3.2 million Australians over 65. In the same year Medicare expenditure grew by $2.82 billion which is six times the proposed saving ,in a single twelve month period. The PSR has been operating for 32 years and recovered $21.5 million in its best documented recent year against $29.5 billion in expenditure. The government is looking for savings in the wrong place.

Recommendations for the Senate to consider before passing the rebate reduction proposal:

Recommendation 1 — Expand PSR referral capacity:
The Chief Executive Medicare should be resourced to increase referrals to the PSR from approximately 91 per year to a minimum of 500 per year. The PSR cannot act without referrals. The bottleneck is upstream identification not PSR investigation capacity. Increased referrals funded from existing Medicare integrity budgets would cost a fraction of the proposed PHI rebate saving.

Recommendation 2 — AI-assisted peer benchmarking:
The $146.8 million committed to Medicare payment integrity in the 2026-27 Budget should be directed toward AI-assisted provider pattern analysis using existing peer benchmarking data. The PSR trigger — practitioners billing in excess of 99% of their peers — already exists in legislation. AI applied systematically to Medicare data could identify referral-worthy patterns at a scale impossible through manual review.

Recommendation 3 — Publish expenditure growth analysis:
The Department of Health should commission and publish independent analysis of what is driving Medicare expenditure growth — from $15.4 billion in 2009-10 to $32.4 billion in 2024-25, beyond population ageing and service volume increases. A 110% expenditure increase against 23% population growth requires explanation and evidence-based response, not assumption.

Recommendation 4 — Defer the PHI rebate cut:
The proposed PHI rebate cut should be deferred pending the outcome of the Senate inquiry and independent modelling of its full system impact — including the downstream costs to private hospitals, public hospitals and the broader healthcare system documented in Parts 1 and 2 of this series.

The government’s $482 million saving from the PHI rebate cut will be consumed by Medicare expenditure growth in approximately two months. The question is not whether savings need to be found. The question is whether cutting pensioners’ health insurance is the right place to find them, when a $32 billion expenditure base growing at 9.5% annually, overseen by a 32-year-old review mechanism recovering 0.07% of expenditure, remains largely unexamined.

The mechanism to do better already exists. The PSR peer benchmarking trigger is in legislation. The $146.8 million payment integrity commitment is in the budget. AI tools capable of identifying anomalous billing patterns at scale are available now.

The question is not whether we could find the savings elsewhere. The question is whether we have the will to look.