The ATO Is Using AI to Catch Tax Errors — Here's How

The ATO is using AI to catch tax errors — and this isn't a vague claim or a vendor scare campaign. It's documented in an independent audit of the Tax Office's own systems. If you run a small business in Australia, it's worth understanding what these models actually do, what they look for, and the one thing that keeps you on the right side of them: clean, reconciled books.

The evidence: a 40-plus model AI operation

In February 2025 the Australian National Audit Office (ANAO) tabled a performance audit titled Governance of Artificial Intelligence at the Australian Taxation Office (Auditor-General Report No. 26, 2024–25). It's an independent, public review of how the ATO builds and runs AI — not marketing, but the watchdog checking the regulator.

The headline numbers: the ATO had 43 AI models in production as of 14 May 2024, with 14 of them added in the preceding ten months, supported by a team of more than 100 data scientists. The same audit also flagged governance gaps — around 74% of those models lacked a completed data-ethics assessment — which tells you both how fast this is moving and that it's still maturing.

What the ATO's AI actually looks for in small business

Several of the named models target small business and the people who advise them directly:

• Black Economy Omitted Income Scoring — predicts the likelihood of cash-in-hand or otherwise unreported income, and prioritises which small businesses to look at.
• Small Business Stratified Sampling (part of the Tax Practitioner Risk Model) — flags understated business income or overstated deductions across an agent's client base.
• Unexplained Wealth modelling — compares lifestyle and spending against the income actually declared.
• Dedicated risk models for GST, PAYG withholding, superannuation guarantee and BAS lodgments.
• Real-Time Return Messaging — nudges taxpayers when a claim falls outside the expected range for their circumstances.

This sits on top of the ATO's long-running data-matching: it cross-references your lodged figures against information from banks, employers, share registries and payment platforms — reportedly more than a billion transactions a year, including income from gig platforms like Uber and Airbnb. The ATO is also trialling multimodal AI to read and check work-related expense receipts, and has signalled it wants to scale AI across the agency through to 2030.

Why this raises the stakes on messy books

None of this should frighten an honest operator. But it does change the margin for sloppiness. The models are tuned to spot the patterns that messy bookkeeping throws off, even when there's no intent to do anything wrong:

• GST coded incorrectly, so what you report on the BAS doesn't reconcile to the underlying transactions.
• Business and personal spend run through one account and never cleanly separated.
• Income that lands in the bank but never makes it to the ledger.
• Round-number estimates and duplicated bills that look fine in isolation but stand out across a year of data.

An AI model doesn't get tired, doesn't skim, and doesn't give you the benefit of the doubt. It compares your numbers to thousands of similar businesses and surfaces the outliers. If your books are clean and tied to source, you're an unremarkable dot in the cloud. If they're not, you're the outlier worth a closer look.

How to stay on the right side of it

• Reconcile properly and often — every figure should trace back to a real transaction, not an estimate typed into a cell.
• Validate GST and ABNs before you lodge, not after a query arrives.
• Keep business and personal spend separated cleanly, especially in owner-operator and sole-trader files.
• Catch duplicate and altered invoices before they're paid and booked.
• Keep the source documents — if the ATO's receipt-reading AI asks, you want the evidence on hand.

The thread running through all of that is the same: the defence against AI scrutiny isn't secrecy, it's accuracy. Books that reconcile to source are boring to an algorithm, and boring is exactly what you want to be.

The bottom line

The ATO using AI is no longer speculation — it's audited fact, with named models and real figures behind it. For small business owners the takeaway isn't panic, it's hygiene. The same clean, reconciled, source-tied books that make good decisions possible are the ones that survive an AI-flagged review without a second thought. That's a problem worth solving once, properly, rather than hoping you stay under the radar.