Ask most Australian business owners what "AI" is and they'll describe the chatbot on their phone. That's the shape of the misunderstanding, and there's a pair of numbers that shows what it's costing.
The gap, in numbers
Two-thirds of Australian SMBs say they use AI. Five per cent are set up to get real value from it.
The source is Deloitte's AI Maturity Index — a survey of more than a thousand Australian small and medium businesses. The gap between the two numbers is the article in miniature. When "we use AI" can mean anything from a browser tab to a production system, most people mean the browser tab. Adoption on the strict count, AI doing something in the business, is about one in twenty.
Thirteen to one. We've turned up to the technology and almost entirely failed to build with it.
The question we ask every client. Is the AI you're using getting you a result? Has anything measurably changed? Time saved, output up, quality better, mistakes down. Or is it a fun tool that makes the day a bit easier?
Most people can't answer honestly. Not because they're not paying attention. Because nobody put a number on it going in, so there's no number to check against. That's the gap between the 66% and the 5% in one sentence. Using it isn't the same as getting anything from it.
Worth noting where it is moving: financial and insurance services sat at one per cent in 2021–22 and sits at twenty-four per cent now. Twenty-four times, in four years. When a sector turns, it turns fast.
Why the gap exists
It isn't laziness, and it isn't budget. It's structural.
The people who could authorise a build don't do the admin. The people buried in the admin can't authorise a build. Leadership treats AI as a strategy question, something to have a view on at the offsite. The team treats it as another browser tab. Nobody in the middle is holding a list that says these four tasks could be gone in six weeks.
So it stays a conversation. Nearly every business we speak to has had the AI conversation. Very few have had the AI build.
What people believe that keeps them stuck
Under most stalled AI projects sits a belief. Usually one of these four.
1. "AI is a chatbot." It isn't. A chatbot is one thin slice of the technology. The front-end demo, the free tier, the thing your kids use for homework and half your staff already have open in a tab. The build is different. AI reads the inbox and tags the ones that need action. Parses the insurer PDF into the accounting system. Watches the weekly report your team hand-builds every Monday and drafts it before they open the file. Chat is where AI shows off. Workflow is where AI does the work.
2. "It has to be transformational to matter." The build that ships is small. The one that stays on the whiteboard is transformational. Save one person three hours a week, prove it works, then compound. Sexy first builds are graveyards.
3. "We need to clean up our data first." You don't. Start on the messy data. The build almost always improves the data as a byproduct, because now something depends on it being right. Waiting for clean data means waiting.
4. "We should wait until the tech matures." The tech will always be shifting. If you wait for it to settle you'll wait forever, and the businesses that started three years ago will have compounded three years of advantage. The workflow problem in front of you is stable. The tools to fix it are ready enough now.
What a build actually buys you
Three things.
Free up. Hours back from a task nobody was hired to do. The same team absorbing more work without the next hire. Not usually cash. Capacity.
Grow. Manual process scales linearly: twice the volume, twice the people. A build breaks that link. Doubling the work stops being a hiring decision.
Retain. Clients don't leave over the work. They leave over the mistake nobody caught. A person doing the same job for the four hundredth time makes mistakes. Not through carelessness, through repetition. A build does it identically every time. The error you never make is worth more than the hour you save, and it's the benefit nobody puts in the business case.
The four fair questions
Past the beliefs sit four questions that come up in every first meeting. None of them are stupid. Three of them are right.
"It's too expensive." Compared to a subscription, yes. Compared to the task, usually not. A subscription is around thirty dollars per person a month, and your team still has to do all the work - the software just gives them somewhere to do it. A build is a real project with a real invoice, and it removes the work. The number that decides it isn't our fee, it's what the task costs you now. Work that out first — there's a calculator that does it in two boxes. If it's under about ten hours a week across the roles doing it, we'll tell you not to build anything.
"It's not safe." This one isn't a misconception, it's a requirement. Many businesses that haven't adopted AI say some version of it, and they're right to. The answer is architectural, not reassuring noises: your data stays in Australian infrastructure, it isn't used to train anything, and you own the system at the end. If a vendor can't tell you where your client data physically sits, that's your answer.
"It's going to take our jobs." We won't pretend this never happens. What we see more often is roles changing shape. The reconciliation goes, the exception handling stays, and the person stops being a data-entry clerk. But if your honest plan is to cut headcount, say so at the start. It changes what gets built, and it changes how the team treats it.
"It's not accurate." Left unsupervised, correct. These systems will produce a confident wrong answer. That's a design problem, and design solves it: narrow the task, check the output against something that already exists, build it to flag what it can't handle instead of guessing. Anything with money or compliance attached keeps a human signature on it. A build you can't audit isn't finished.
What this looks like in practice
The first thing we built was for the practice one of us works in. Commission statements arrive from every insurer in a different format. Different columns, different labels, formats changing without warning. Someone was reconciling them by hand against what had been banked.
The build reads each statement, pulls the numbers out, and matches them to the bank. Every insurer has its own recipe, and the system notices when a format changes rather than quietly producing a wrong answer.
That last part is the whole argument about accuracy. Not a promise that it won't get things wrong. A design that tells you when it might have.
Two weeks to build. Three days a week off someone's plate. $12,000 a year off the software bill.
Where we sit
whatcould.ai builds custom tools for Australian businesses whose admin has outgrown the software running it.
Sometimes the right answer is a tool you can buy, or a workflow change, and we'll say so. But when the task is specific to how your business runs, there usually isn't a product for it. That's the gap we work in.
If nothing else
If your business has had the AI conversation more than twice and never had the build, that's not a technology problem. It's a decision problem.
Pick the most repetitive task in the building and work out what it costs you. If it's more than ten hours a week across the roles doing it, it's worth doing something about.
Not using this technology at all, in 2026, is starting to look the way not using a computer looked in 2003. The businesses that got serious about software early compounded a twenty-year advantage over the ones that waited to be sure. The same shape is playing out now, faster.
Sources
- Deloitte, AI Maturity Index, Australian SMB survey, 2026
- Australian Bureau of Statistics, Characteristics of Australian Business, 2024–25
- Roy Morgan, AI usage, March quarter 2026