AI adoption in healthcare: cost, risks and where to start
I ran a multidisciplinary clinic in New Zealand. Every time I bought software the question was the same. Is this worth the money, and will I be able to tell whether it worked?
Most businesses buying AI right now can't answer the second half. MIT's NANDA project looked at $30 to $40 billion of enterprise AI spending and found "95% of organizations are getting zero return" (MIT NANDA, The GenAI Divide). The tools mostly worked fine. What got skipped was the dull stuff. Clean data, system integration, governance, training people. Buying the tool is the fun part, so that's the part that happens.
That leaves you with two questions. Which problem do I fix first, and what's it worth spending to fix it?
The short version. Pick the problem before you pick the tool. Work out what that problem costs you right now, in hours or in dollars, and spend against that number rather than against a vendor's pitch. If two ideas both look good, run the one that shows a result fastest, because that's the one that pays for the next one and gets your team on side. Budget integration, training and governance as real line items. And don't start at all if the basics of your service still need work.
What this page covers, and what it doesn't
This page is about money and risk. Which problem to pick, what drives the cost, why projects fall over, and when to wait.
The running order is a different job. Who owns each step, what you check before you buy, when consent and training happen, what breaks when you skip one. That's all in how to build an AI implementation roadmap for healthcare. Read this page first, then that one.
After the day to day picture instead, such as what AI actually does across a clinic week? Start with AI in healthcare administration.
Problem first, budget second
When I was buying software for my own clinic, and later when I was on the other side of the table selling it, the same rule held. Work out what the problem is costing you, then decide whether fixing it is worth the spend.
A small problem that costs a lot to fix isn't worth fixing. That sounds obvious. It stops being obvious the moment a good demo lands in your inbox.
Judge the spend against your margin. I always knew my clinic's profit margin, so I could see the moment overheads started eating into it. That was my line. No clear return, no spend.
Business development is the exception. Getting new patients through the door needs an outlay you accept going in, and a level of risk with it. You can still put a number on it. It just won't be a tidy one.
When two ideas both look good
Pick the one that shows a result fastest. Not the biggest one. The fastest one.
Two reasons. It pays for itself sooner, which justifies the next spend. And your team gets to watch it work, which beats any explanation you give them.
For example, say you've got automated patient recalls on the list and a reporting dashboard next to it. The recalls start filling gaps in the diary inside a fortnight. The dashboard gives you a better view of the business in three months. Do the recalls, then let them fund the dashboard.
How long do you give it
In my New Zealand clinic I built a referral programme that gave patients a discount after treatment. Our practice software couldn't run one, so I built my own. I gave it six months. At the three month mark I looked at what was coming through, changed a few things, then let it run for another three.
Six months suited a clinic. At the SaaS company the loops were far tighter, because the business needed them to be. Your window depends on how fast your own business gets feedback.
Check your regulator before you copy any of that. In New Zealand, the Osteopathic Council allows an inducement only where the terms and conditions are set out clearly and patients' best interests are considered (OCNZ advertising policy). An inducement is any offer that gives a patient something extra for booking, such as a discount, a free add-on or a prize draw. In Australia, AHPRA wants the terms and conditions of any discount or inducement stated in the advertisement itself (AHPRA advertising guidelines).
Where healthcare has actually got to
Adoption in healthcare is real, and it's narrow.
That same MIT report put healthcare among the least disrupted sectors it looked at. Its one line verdict on the sector was "Documentation/transcription pilots; clinical models unchanged" (MIT NANDA). It also found healthcare executives "express no expectation of reducing physician or clinical staff hiring."
A survey of 43 US health systems (Poon et al., JAMIA) found the same shape from the inside. Ambient notes was the only use case where every single respondent reported adoption activity, and 53% called it a high degree of success (Poon et al., JAMIA). Ambient notes is AI that listens to a consult and drafts the clinical note, which a clinician then reads, fixes and signs.
| What the US survey found | Figure |
|---|---|
| Health systems that answered | 43 of 67 invited (Poon et al., JAMIA) |
| Reporting adoption activity on ambient notes | 100% (Poon et al., JAMIA) |
| At least partial deployment in imaging and radiology | 90% (Poon et al., JAMIA) |
| Naming immature tools as a barrier | 77% (Poon et al., JAMIA) |
| Naming financial concerns as a barrier | 47% (Poon et al., JAMIA) |
Those are big US health systems, not clinics like yours. Read the shape, not the numbers. One low risk use case is doing nearly all the work, and money is the second loudest complaint.
In Australia, documentation is where most practices meet AI too, usually as a scribe. We cover choosing an AI scribe separately.
Australia sits high on the adoption tables
CSIRO and the National AI Centre counted healthcare as an AI adopting sector at 19% of the public AI companies in their sample, second only to energy, raw materials and utilities on 20% (CSIRO / National AI Centre). The same report found 85% of private AI businesses employ fewer than 50 staff (CSIRO / National AI Centre). So this is mostly small operators, not giants.
Being counted as an adopter is a low bar though. Back in 2021 the Australian Alliance for AI in Healthcare ranked adoption itself 7th of 8 national priorities (AAAiH, original roadmap). It sat on 10.5% of stakeholder responses, behind safety, quality and ethics, which took 17.8% (AAAiH, original roadmap). Its third roadmap, launched in July 2026, now wants health services to "move beyond isolated pilots to large-scale deployment", and carries 28 recommendations across 8 priority areas (AAAiH, 3rd National Policy Roadmap). Five years, and the national ask went from help us plan to stop piloting.
Why so many AI projects fall over
Look at how MIT got to that 95%. It ran 153 interviews with senior leaders and read through 300 plus publicly disclosed AI initiatives. Just 5% of integrated pilots pulled out real value (MIT NANDA).
The funnel underneath that number is the useful bit. For enterprise grade AI tools, 60% of organisations had a look, 20% got as far as a pilot, and 5% reached production (MIT NANDA). Most of the money goes in before anything reaches the people it was bought for.
Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, because of "escalating costs, unclear business value or inadequate risk controls" (Gartner). Agentic AI is software that carries out a chain of tasks on your behalf, such as booking, chasing and filing, rather than waiting for you to click at each step.
Then there's the line worth pinning above your desk. Gartner reckons only about 130 of the thousands of agentic AI vendors out there are real. The rest are existing products wearing a new label, which Gartner calls agent washing (Gartner).
Neither of those findings is healthcare specific. Both come from enterprise research, and I'm citing them as context rather than as healthcare numbers. They describe what happens in any industry when the tool gets bought before the groundwork gets done.
What actually drives the cost
Three independent sources agree on the shape of the cost, even though none of them can give you a dollar figure for an Australian clinic.
Harvard Business School Online, writing about business generally, puts AI implementation cost in four buckets. Infrastructure. System integration. Maintenance and continuous improvement. And human capital, which means upskilling and training your people (HBS Online). A peer reviewed barriers review in Cureus (Ahmed et al., Cureus) and the US health system survey (Poon et al., JAMIA) land on the same categories from the healthcare side.
| Cost line | What it actually is | How often it bites |
|---|---|---|
| Licences and platform | The monthly per seat or usage fee | Smallest and most predictable of the five |
| Integration | Joining the tool to the systems you already run | 23 of the studies reviewed cited interoperability and integration as a barrier (Ahmed et al., Cureus) |
| Data preparation | Getting your records into a state a tool can use | 42 of the studies cited data quality, data access and data set size (Ahmed et al., Cureus) |
| Training and change | Teaching the team, plus the hours that takes | 35 of the studies cited workforce, 27 of those on education and training (Ahmed et al., Cureus) |
| Governance | Accountability, risk work, a register, monitoring | Six essential practices, all of them work someone has to do (Department of Industry) |
Interoperability is whether two systems can talk to each other and swap data in a form the other one can actually use. It's the line that catches people out. The Cureus review found systems that "acquire annotations and stor[e] them in a format which does not permit reuse in AI development" (Ahmed et al., Cureus). In plain words, your data is there. It's just in a shape the AI can't read.
I've spent years on the sales side of software deals. The pitch almost never gets into whether there's an API to connect to, or what happens when a practice is running something older. That's the question we push at first in AI workflow automation, before anyone commits money.
One more from the same review, and it's the line nobody budgets for. In a study of one decision support AI rollout, the biggest resource barrier turned out to be personnel time (Ahmed et al., Cureus). Not the licence. People's hours.
Nobody has a credible AU or NZ price
I went looking for one. There is no independently verified Australian or New Zealand dollar figure for what any of this costs a small or mid size health business.
Vendors publish their own numbers and some of them look plausible. None of them tells you how many businesses they asked, or shows the working behind it. Consider what that means when a quote lands on your desk. Treat the figure as a claim to check, not a fact. Ask what's included, ask what the integration work is priced at on its own, and ask what year two looks like.
Governance is a cost line, not an afterthought
Australia's guidance here is voluntary, and following it still costs money.
The Department of Industry, Science and Resources sets out six essential practices for AI adoption: decide who's accountable, understand the impacts and plan for them, work out and handle the AI specific risks, share information openly, check and monitor what the tool is doing, and keep a human in control (Department of Industry, Guidance for AI adoption). That guidance replaced the Voluntary AI Safety Standard's ten guardrails, which covered the same ground in more detail, including record keeping and supply chain transparency (Department of Industry, Voluntary AI Safety Standard).
An AI register is a simple list of every AI tool you run, what it does, what data it touches and who is responsible for it. Nothing on that list builds itself, even at three practitioner scale. Someone writes the policy and keeps the register current. Someone does the monitoring. Those are hours, and hours are a cost.
The Australian Alliance for AI in Healthcare says much the same from the sector side. Governance is "a shared responsibility... some of the governance tasks are local and health service providers need support to take on this new set of responsibilities" (AAAiH, 3rd National Policy Roadmap). Local means you.
Bringing the team along
Change management is the work of getting people to actually use a new thing, as opposed to the work of installing it. It's usually the bigger half of the job.
There's a lot of fear about AI taking jobs. This is my own view and not a sourced fact, but I don't buy it. AI mostly augments jobs rather than replacing them. Some roles will shrink or go, new ones will appear, the way it has gone with every technology shift before this one.
Your staff can have a view on how the change happens. The decision itself sits with whoever owns the business. Say it clearly from day one: what's changing, why, and what you expect from them.
Some people find technology change much harder than others, and in a clinic that's often the people who've been there longest. Give them more support, not less. Break the rollout into small pieces and record how to do each one, for instance a two minute video or a one page checklist they can keep at the desk.
When not to start yet
AI comes after the basics. There's no case for an AI receptionist while the service itself is patchy, or while patients still can't get an easy run through your practice. AI enhances what's already there. It won't create it.
Not sure where you sit? Here's a plain check. Can you say, today, what one specific problem is costing you per week in hours or in dollars? If you can't, that's the work to do first, and it's free.
Where to start, and where Opal steps in
Pick one problem you can put a number on. Size it small enough to prove itself quickly. Work out the integration and training cost before you look at the licence cost, because that's where the money actually goes. And sort out who's accountable before you buy anything (Department of Industry).
Once you've made that call, go and get the order right. How to build an AI implementation roadmap for healthcare has the sequence, step by step, with who owns each one.
We help practices work out which problem is worth solving first, and what has to be true before a tool goes anywhere near patients. That's what AI strategy and adoption covers. We don't sell software and we take no cut on what we recommend.
FAQ
How much does AI cost to implement in a clinic?
Nobody has published a verified Australian or New Zealand benchmark. What is well established, from the global literature, is the shape of the cost. Four buckets from HBS Online (HBS Online), plus data quality, the most cited barrier theme in a peer reviewed review (Ahmed et al., Cureus). Treat any vendor's dollar figure as a claim to check.
Why do AI projects fail?
Mostly because the groundwork gets skipped. MIT found 95% of organisations getting zero return on their AI investment (MIT NANDA). Gartner expects over 40% of agentic AI projects to be cancelled by 2027, on rising cost, unclear value and weak risk controls (Gartner).
How widely is AI used in Australian healthcare?
Healthcare is one of the largest AI adopting sectors CSIRO and the National AI Centre have tracked, at 19% of the public AI companies in their sample (CSIRO / National AI Centre). Adoption doesn't mean broad change though. In a survey of US health systems, ambient documentation was the only use case every respondent reported adoption activity on (Poon et al., JAMIA).
Where should a practice start with AI?
With one problem you can put a number on, sized small enough to prove itself. Pick the option that shows a result fastest, budget it against what it'll actually return, and sort out accountability and risk work before you buy anything (Department of Industry).