Why most Australian small businesses won't touch AI

Ask an owner why they haven't done anything with AI and you'll usually get an answer about time or money. Ask enough of them and a different answer comes out.

65% of Australian businesses that haven't adopted AI say the reason is trust and control — not cost, not skills. (National AI Centre, SME AI Pulse, 2026.) A separate figure from the same survey: 54% of small businesses not adopting AI say it simply isn't relevant to their business.

Those two numbers look like they're in tension. They aren't. They're the same person on different days.

"Not relevant" usually means "nobody showed me my version of it"

Nobody has shown a plumber the version where the thing drafts fourteen overdue-invoice reminders and his admin sends thirteen of them. What he's been shown is a chatbot, an image generator, and a conference talk about transformation. So of course it isn't relevant.

In construction and agriculture, fewer than 30% of Australian businesses have adopted anything at all. I don't read that as resistance. I read it as an honest response to a market that keeps demoing the wrong thing.

"Trust and control" is a specific fear, not a vague one

When an owner says they don't trust it, they're not worried the grammar will be wrong. They're worried about one email — the one that goes to the customer who's in hospital, or the one that quotes the wrong rate to the client who'll hold it against them for three years.

That fear is correct. The failure mode of a badly built AI setup really is a wrong thing sent automatically at 2am, to a real customer, with the owner's name on it.

Design for the mistake and it stops being a risk.

Which is why every system I install ends with a named person approving the output before it reaches a customer. Not a policy on a wall — written into the plan and written into the contract, with that person's name in it. The failure mode then becomes: your admin reads a bad draft and deletes it. Ten seconds.

The gap between "tried it" and "uses it"

There's a set of adoption numbers floating around that look like they contradict each other. The ABS puts Australian business AI use at 12% (Characteristics of Australian Business, 2024–25). Industry surveys say 43% (NAIC SME AI Pulse). Some say 69%.

They're all right, because they're counting different things. Roughly: 69% have someone who has opened ChatGPT. 43% have tried it on something real. 12% have AI inside a process that runs whether or not anyone remembers to open a browser tab.

The gap between 69% and 12% is where every owner-run business I talk to lives. Tried it. Saw something. Nobody wired it in.

That gap doesn't close with training. Sending the office manager on a prompting course produces a better prompter, not a system. It closes with two unglamorous things.

Wiring, and rules

Wiring means the AI is connected to the software the business already runs — the job system, Xero, the shared inbox — so a task starts itself. Not "someone remembers to paste the invoice list into a chat window". The overdue invoices are already in Xero; the drafts should appear without anyone fetching anything.

Rules means someone has written down who checks what. Most businesses have never done this, which is how you get the other failure I see constantly: an office manager quietly pasting customer details into a free tool for six months because nobody ever said not to.

That one is fixed with three lines on a single page:

• What AI may draft — the list of tasks it's allowed to touch.
• What a person must approve before it leaves the building, and who that person is by name.
• What never goes in — customer identifiers, payroll, anything covered by a confidentiality clause.

One page. It takes an afternoon and it's the cheapest risk control in the business. There's a free template at morehours.com.au/guides/ai-use-policy — copy it, no sign-up.

What this means for the first project

If trust is the barrier, then the first project's job is to earn trust, not to save the most hours. That rules out most of what owners instinctively pick first.

• High frequency, so the team sees it work over and over in the first fortnight.
• Low risk, so the worst case is an awkward email, not a Fair Work matter.
• An obvious approver — one person whose job already includes checking this kind of thing.

Overdue-invoice reminders fit all three. So do enquiry acknowledgements and job notes from a voice memo. Payroll fits none of them, which is why it's the wrong place to start even though it's the loudest headache.

Win the boring one in month one, and the team will let you near the hard one in month three.

How you'd know whether it worked

The other half of trust is measurement, and most of this industry is allergic to it. You get a case study with a number in it and no method.

The method matters more than the number. Time the task by hand first — an actual stopwatch, three runs, because the first run is always the slowest. Time it again with the automation. Multiply by how often it runs in a month. Write down both figures, including the months where the result is disappointing, because those months happen and hiding them is how a supplier loses the room.

If a project isn't returning hours after ninety days, stop doing that project. That belongs in the agreement, not just in the sales pitch.

The figure I quote is measured that way: 23 hours in a month across three automations, in my own business, timed before and after. Any number I give you will come with the method attached.

The honest version of the pitch

The product isn't "AI that does more". It's "AI that never sends anything without your name on it". That sounds like a limitation. It's the only version most Australian owners will actually run.

If you're at four people and busy, you probably don't need any of this yet — the owner is still the system, and that works. Somewhere between five and twenty-five people something changes: there's an office person now, and a set of tasks that happen every week the same way, done by hand every time. That's where a system starts paying.