AI vs automation: what your clinic actually needs
I connected Xero to my own clinic. That was automation, and nothing in it learned a thing. It reconciled, it pulled in the ACC invoicing, and it gave me a full picture of the money so I could forecast off it. Same inputs, same result, every time.
An AI scribe writing up a consult note from a recording is the other thing. It reads something messy and makes a call about what matters.
Both get sold to you as "AI". Which one you're actually buying changes the price, the delivery timeline, and how carefully somebody has to check the output.
In brief: most of what a clinic gets sold as AI is really a rule you could write down yourself. Rules cost less and need less checking, so sort the job first and then go shopping. If every step fits on a whiteboard with no gaps, buy automation. Pay AI prices only for the step that has to read something and make a call.
Why the two words get used as if they mean the same thing
Partly it's because the category itself is genuinely loose. Moveworks says it straight: "The media frequently conflate AI and automation. While they're not the same technologies, they're not completely different either" (Moveworks). Moveworks sells AI-powered automation, so a blurry boundary suits them perfectly.
The official guidance does it too. The Australian Government's small-business AI page calls AI "computer programs that can do things which normally need human intelligence". Then it lists "Automating tasks" as one of nine things AI is used for (business.gov.au). Automation isn't even treated as its own technology there.
So when a vendor tells you their product is AI, there's no neutral definition available to hold the claim up against.
What automation actually is
Automation is, in Zapier's words, "setting up predefined workflows, 'When X happens, do Y', to complete repetitive or routine tasks without manual effort" (Zapier). It runs a fixed set of steps, and nothing in it learns.
A workflow is the path those steps run along, from the thing that kicks it off to the last action it takes. You write the path down once and the software walks it for you.
Red Hat adds the bit that matters for you. Automation is "Deterministic: If the input is the same, the output will likely be the same" (Red Hat). A recall SMS that fires six months after a patient's last visit does the same thing on a Tuesday as it does in June. Nobody has to babysit it, because all the thinking happened once, back when someone decided what the rule should be.
The catch is that same property, flipped. Zapier again: automation "falls short when an action along the path requires interpreting data and making a decision before it can proceed."
What AI actually is
AI is, in Red Hat's phrasing, "the brain power that learns and adapts to make non-rule-based decisions about complex problems", sitting next to automation as "the machine that executes tasks based on predefined rules". It reads something messy and unstructured, then works out what to do with it.
Zapier lands on the same idea. AI makes decisions "by interpreting data rather than following fixed rules."
And where automation is deterministic, AI is "Nondeterministic: AI generates results based on statistical probability and learned patterns. The output can change based on new data or an evolved model, even if the initial prompt is similar."
That last line is the entire reason an AI tool needs supervision and an automation mostly doesn't. Feed it the same consult twice and you can get two different notes, both of them sounding perfectly reasonable.
The one question that separates them
Red Hat boils it down to one test: "You can summarize the distinction between AI and automation in 1 question: Does it follow rules or does it learn and suggest rules?"
The Australian IT firm Think Technology puts it in a shape you can run on a real job: "A useful rule of thumb: if you could write out every step of the process on a whiteboard with no gaps, it's probably an automation job. If the process would have five 'it depends' notes on that whiteboard, it probably needs AI" (Think Technology).
Try it on a recall. Every step is knowable: find patients whose last visit was six months back, check they haven't rebooked, send the message. No "it depends" anywhere. Now try it on an enquiry that lands at 9pm as a paragraph of free text, where nearly every step is an "it depends".
Then run it on your own week. Pull four jobs off the administration list, write the steps out properly, and count the "it depends" notes on each one. For example, here's how four jobs every clinic does come out.
| A task from your week | Can every step be written down? | What that makes it |
|---|---|---|
| Sending appointment reminders the day before | Yes. Find tomorrow's bookings, send the message. | Automation |
| Reading a referral letter and pulling out the history that matters | No. What matters changes letter to letter. | AI, on that step only |
| Working out which marketing brought a patient in | Yes, once every enquiry records where it came from. | Automation |
| Drafting the consult note from the recording | No. The recording is messy and something has to decide what to keep. | AI, with a practitioner checking it |
Reminders are the plainest of the four, and they're the ones clinics most often go shopping for a model to do. We've written up what a reminder sequence should actually look like in our guide to conversational AI in healthcare.
Side by side, in a clinic
| Automation | AI | |
|---|---|---|
| What it does | Runs a fixed set of steps, the same way every time | Reads something unstructured, such as a message, a note or a recording, and works out what to do with it |
| How it decides | Follows the rule it was given: when X happens, do Y | Weighs the context and makes a judgement call, based on patterns in what it has seen before |
| A clinic example | A recall SMS goes out automatically six months after a patient's last visit | An AI scribe drafts a consult note from the recording, for the practitioner to check and sign |
| What it needs to work well | A clearly defined, repeatable process | Enough relevant data or examples, and someone checking its output |
| What happens with something it hasn't seen before | It does the wrong thing, or nothing, until a person changes the rule | It has a go anyway, which is why checking its output matters more, not less |
| Who checks the output, and how often | Spot-checked now and then, low risk if the rule was right to begin with | Checked every time, at least early on, because a wrong answer can sound as confident as a right one |
| Where we help | AI workflow automation | AI agents and assistants, or AI strategy and adoption |
Sorting a real list
This bit is mine, not a case study I read somewhere. The clinic list is what I built while running a multidisciplinary clinic in New Zealand. The sales list under it comes from selling technology into practices. Both are sorted by the test above.
Rules, all of it. Xero connected for the reconciling and for pulling in the ACC invoicing. The forecasting I could do on top of that, once the numbers were complete enough to forecast from. A referral-source system, so I could actually see which marketing was working. And a mostly automated way of working out what I owed the contractors each month. Not one of those needed a model. They needed the steps written down properly and something to run them.
Outside the clinic, in sales work, it gets more interesting. (I've worked with both Splose and HubSpot, so treat those as disclosed relationships rather than neutral picks.) Here's what I set up there:
- A meeting and note taker feeding the CRM. Listening and summarising is judgement, while pushing the result into the right record is a rule.
- CRM properties updated off the back of calls and meetings. Reading the call to find the answer is judgement, while writing it into a field is a rule.
- Critical information pulled out and highlighted. Judgement the whole way. Deciding what counts as critical is the thing a rule can't do.
- A scheduled review of open deals that flags gaps and risks. The schedule is a rule, and what it calls a risk is judgement.
- Automated outreach over LinkedIn and email. Rules.
- Lead activity ranked by how long it had been since I last did anything against a medium or high rated account. Rules, and a good example of something that feels like AI purely because it hands you a prioritised list.
Most useful things in a clinic are a chain like that, not one or the other. The scribe is the judgement step in a note-taking workflow that's otherwise all rules.
If a patient's personal information goes through any part of that chain, privacy law applies to it, whichever half of the chain it sits in. Our guide to AI adoption in healthcare covers where that bites.
The tools that do the connecting
An integration is the connection that lets two bits of software pass information to each other, like a practice management system handing a new booking to an accounting tool. Most of the automation you want is really a set of these. Naming actual tools is more useful than staying abstract, so here are the ones that regularly come up. This is an overview, not a recommendation. We sell no software and take no commission on anything we suggest.
For wiring existing systems together, the general-purpose options are Microsoft Power Automate, Zapier and n8n. All three do the "when X happens, do Y" job across tools that don't otherwise talk to each other, and none of them is built for healthcare.
For clinics specifically, Propagator is worth watching. It connects a practice management system to the other software a clinic runs, and it's hosted in Australia and New Zealand. It's early, so treat it as one to look at rather than a settled default.
On the AI side, the tool you're most likely to meet first is a scribe. We've written separately on choosing an AI scribe and rolling one out. The next one most clinics look at is an AI receptionist, which is the same judgement step pointed at a phone call instead of a consult.
The mistake that costs the most
Buying AI for a job a rule would do. Think Technology's example:
"A Brisbane accounting firm we work with spent three months setting up an AI tool to send invoice reminders. A simple Power Automate rule would have done the same job in an afternoon, for a fraction of the cost. The AI added nothing the rule couldn't do. It just felt more exciting."
My read on that firm, and it's a read rather than a fact about them: either nobody properly laid out the options and the opportunity cost of spending three months there, or the excitement of a new tool got in front of the return it was actually going to deliver.
The failure underneath it is usually scoping, not technology. Across the work I've done at HubSpot and with Splose, I keep seeing the same pattern. A project that isn't scoped properly at the start throws up surprises, and surprises take longer than anyone planned for. The timeline stretches, motivation drops because the project is dragging on longer than expected, and eventually the whole thing falls over.
The fix is dull, and it's basic project management. Scope it properly before anyone builds, and check that whoever is building it has the resources to do it. Get a timeline with genuine milestones and deliverables on it. That's what keeps a project inside its scope, and it applies whether the thing being built learns anything or not.
How fast this is moving in Australia
Two numbers, from two different surveys, worth reading next to each other.
The Australian Bureau of Statistics reported that "around 12 per cent of Australian businesses reported the use of AI in their workplace in 2024–25", and that "uptake among small and micro businesses was lower, at around 11 per cent". That's Tom Lay, ABS head of business statistics (ABS).
Separately, RACGP's newsGP poll found AI scribe use among GPs went from 22% in August 2024 to 40% by November 2025 (RACGP's newsGP).
Different populations, different methods, different questions. You can't reasonably combine them into a single figure, and we're not going to pretend otherwise. What they do show is that the one AI tool with a ready-made clinical use case has moved fast. The rules half of the list, the recalls and reminders and intake forms that would be a same-day win, gets talked about far less.
A test to run before you buy either one
Every decision framework published on this topic is written by someone selling one side of the answer, whether that's an automation platform, an IT consultancy's hours, or an AI agent product. Think Technology's four questions are the best of the lot, and this is our version, built on the same shape and aimed at what a clinic actually buys.
- Can you write every step down with no gaps? If yes, you want automation, and you shouldn't pay AI prices for it.
- Does something in the middle have to be read and interpreted? A recording, a free-text enquiry, a referral letter. That step, and only that step, is the AI.
- What happens if it gets one wrong? A recall message sent to the wrong patient is an apology. A consult note with a symptom in it that never happened is a different category of problem, and that difference sets how much checking you need.
- Is the job worth the setup? Three months of build for something that costs your front desk twenty minutes a week is the Brisbane accounting firm all over again.
- Ask the vendor which one they're selling. Then ask what happens when the tool meets an input it hasn't seen. A rule-based product will tell you it stops, and an AI product will tell you it has a go. Both answers are fine. A vendor who can't answer the question is the flag.
Where Opal fits
Start with the list, not the tool. Write out the administration your practice gets through in a week, then run each line through question one. You'll usually discover the rules half is considerably longer than you expected and cheaper than you assumed.
That list is where we start too. AI strategy and adoption begins with an audit of which tasks are actually costing you hours, then sorts them into what a rule fixes this month and what genuinely needs judgement. For the wider view of what AI is already doing in AU and NZ practices, read AI in healthcare administration.