Payroll, tax claims, expense interpretation: the financial tasks where AI keeps getting it wrong for Australian businesses, and what to use it for safely instead.
AI tools are genuinely useful for running a small business. The problem is not the technology. It is the questions being asked of it.
New research from Dext, surveying 500 accountants and bookkeepers across Australia, found that 63 per cent are aware of businesses that have suffered direct financial losses from incorrect or misleading AI-generated advice. Forty-four per cent say they are correcting AI-driven client mistakes every single week.
The losses include overpayments, missed allowances, penalties, incorrect tax claims, and payroll errors. All avoidable. All happening because a business owner asked a general-purpose AI tool something it was not built to answer reliably. So where does the line actually sit?
Where AI actually helps
Used in the right way, AI can save small business owners real time on financial administration.
It is genuinely useful for understanding concepts. If you want to know what a term on your BAS means, how depreciation works in general, or what the difference between a contractor and an employee looks like on paper, AI can explain that clearly and quickly.
It handles drafting well. Asking AI to help write a payment terms clause, structure an invoice template, or draft a query to send to your accountant are all low-risk uses that save time without creating compliance exposure.
It is also useful for preparation. Using AI to organise your thinking before a meeting with your accountant, generate a list of questions to ask, or summarise a document you need to understand is exactly the kind of task it handles well.
Where it keeps going wrong
The errors showing up in Australian books right now are concentrated in specific areas, and they share a common thread: decisions that require knowledge of your actual business context, current Australian tax law, and individual circumstances.
The most common mistakes found in the Dext research include incorrect interpretation of business expenses, reported by 45 per cent of accountants, incorrect tax claims or charges at 43 per cent, flawed personal tax planning at 39 per cent, incorrect business tax planning at 38 per cent, and payroll errors at 29 per cent.
These are not edge cases. They are everyday business finance decisions where AI confidently produces an answer that looks right but is not right for your situation.
Paul Wittich, General Manager APAC at Dext, puts it plainly. “There’s a fundamental difference between specialist tools built for accounting and bookkeeping, and general-purpose chatbots that don’t know a business’s true financial context,” Wittich said.
A general-purpose AI tool does not know your industry, your structure, your state-based obligations, or the specific concessions you may or may not be eligible for. It produces answers based on patterns in its training data, not your circumstances.
The cost of getting it wrong
The financial hit comes from two directions. The first is direct: wrong tax claims, missed entitlements, payroll errors, and potential ATO penalties. These are the losses the Dext research captures, real money leaving businesses because an AI answer was treated as professional advice.
The second is less visible. Among accountants regularly encountering AI-related mistakes, 52 per cent spend up to three hours per month correcting them, and 38 per cent spend between four and six hours. Those correction hours get billed back to the business. The time saved by asking AI instead of your accountant can end up costing significantly more when the error needs fixing.
“The damage is no longer hypothetical. Across the country, businesses are already losing money, and accountants and bookkeepers are spending valuable time correcting avoidable mistakes,” Wittich said.
The rule of thumb
A simple test for any financial question you are considering putting to an AI tool: would the answer change if your business structure, turnover, state, or industry were different?
If yes, it is not a question for a general-purpose AI. It is a question for your accountant or bookkeeper.
AI is a useful starting point for understanding. It is not a reliable endpoint for decisions. Using it to prepare smarter questions for your professional adviser is good practice. Using it to replace that conversation is where the losses are happening.
With the ATO running an estimated two million compliance checks annually and actively using data matching to flag inconsistencies, this is not the financial year to find out the hard way where that line sits.
The quantitative research was commissioned by Dext and conducted by Censuswide.
Keep up to date with our stories on LinkedIn, Twitter, Facebook and Instagram.
