When AI Makes Mistakes: Human Oversight in 2025

AI makes mistakes in accounting because it cannot reliably apply Australian tax law, evidence standards, and professional judgement to messy real-world facts; therefore, human oversight is essential to ensure GST, BAS, PAYG, Division 7A, and income tax positions are supportable under ATO guidance and the legislation. In an Australian accounting practice, the risk is not merely “wrong coding” — it is incorrect tax outcomes, unsupported workpapers, and audit exposure, where responsibility remains with the practitioner under professional standards even if AI performed the initial processing.

• Misclassification errors: Coding private expenses as deductible; confusing capital vs revenue; incorrect GST treatment.
• Context errors: Missing that an entity is a trust, SMSF, NFP, or has special rules (e.g., GST-free supplies, financial supplies, margin scheme).
• Evidence errors: Treating a bank narration as sufficient proof without invoices, tax invoices, contracts, logbooks, or apportionment support.
• Temporal errors: Applying wrong tax-year thresholds or rates; misunderstanding effective dates and transitional provisions.
• Reasoning/hallucination errors: Producing plausible-sounding but incorrect explanations, “ATO references,” or assumptions about missing facts.

Australian practice reality: the agent and the taxpayer bear the risk, not the software.

Why is human oversight essential under Australian tax law and ATO expectations?

• Self-assessment relies on correct application of law: The Income Tax Assessment Act 1997 and 1936 require applying rules to facts; AI does not “know” your client’s facts unless verified.
• ATO expects contemporaneous evidence: For GST credits and income tax deductions, evidence standards (e.g., valid tax invoices and substantiation) must be satisfied, not inferred.
• Anti-avoidance and integrity rules require judgement: For example, Part IVA considerations cannot be “automated” safely from transaction lists.
• Division 7A requires precise classification and timing: Misidentifying shareholder loans, repayments, and benchmark interest has downstream consequences.

• ATO guidance on record keeping (business and tax records; substantiation expectations).
• ATO guidance on GST tax invoices and creditable acquisitions (requirements for claiming GST credits).
• Income Tax Assessment Act 1997 (general deduction rules and specific regimes like depreciation).
• Income Tax Assessment Act 1936, Division 7A (private company loans and deemed dividends).
• ATO Practical Compliance Guidelines (PCGs) and Tax Rulings (TRs) relevant to common issues (e.g., home office, PSI, residency, trusts, Division 7A administration).

Disclaimer-worthy point (but operationally critical): AI output is not “advice.” It is a draft that must be reviewed against law and evidence.

Where does AI most commonly go wrong in BAS, GST, and PAYG workflows?

• Claiming GST credits without valid tax invoices: AI may assume GST applies because a supplier is known, but the ATO requires specific invoice elements.
• Input-taxed and mixed supplies: Financial supplies, residential rent, and mixed-use expenses require apportionment and correct labels.
• GST on imports and reverse charge issues: Bank transactions rarely carry the data needed to determine correct GST treatment.
• PAYG withholding vs contractor payments: AI may code wages/contractor expenses without confirming withholding obligations, STP alignment, or PSI indicators.

• A bank feed shows “UBER TRIP” and AI codes it as “Travel – GST claimable.”
• Human review identifies:
• Outcome: BAS labels and GST claims change, and the file becomes defensible.

How does AI mis-handle income tax deductions and substantiation?

• Private vs business apportionment: Phone, internet, motor vehicle, home office, and mixed-use occupancy costs require evidence-based apportionment.
• Capital vs revenue classification: AI often treats asset-like purchases as immediate deductions; tax depreciation rules and temporary measures require careful application.
• Repairs vs improvements: AI cannot reliably distinguish repairs (potentially deductible) from capital improvements without context and invoices.
• Entertainment and FBT interactions: Meals and benefits frequently require FBT/entertainment treatment, not a simple “deductible expense” code.

• ATO guidance consistently emphasises that claims must be supported by records, and private components must be excluded or apportioned.

Why is AI particularly risky for Division 7A, trusts, and year-end journals?

• Misidentifying a shareholder-related payment as a business expense rather than a loan/payment potentially caught by Division 7A.
• Missing that repayments must be correctly characterised (principal vs interest) and supported.
• Applying an incorrect benchmark interest rate or incorrect loan term assumptions.

• Division 7A is governed by Income Tax Assessment Act 1936 provisions; ATO guidance and practice statements influence administration and expectations.

• Determining who is presently entitled and when.
• Tracking resolutions, trustee minutes, and UPEs.
• Reconciling beneficiary loan accounts and Division 7A interactions (where corporate beneficiaries exist).

• Incorrect trust distribution journals can cascade into wrong taxable income allocation and compliance risk.

Can AI be used safely in an Australian accounting practice?

1. Define “AI allowed” vs “AI prohibited” tasks
2. Implement evidence-first workflows
3. Require review sign-offs
4. Maintain audit trails
5. Use exception reporting

How does MyLedger reduce AI risk compared to typical “black box” automation?

• Automated bank reconciliation with control: MyLedger AutoRecon delivers reconciliation in 10–15 minutes per client vs 3–4 hours (about 90% faster), while maintaining a spreadsheet-like interface that supports rapid human review.
• AI-powered reconciliation that learns patterns: Around 90% auto-categorisation can be achieved, but final accountability remains with the reviewer—supporting a practical “AI drafts, humans decide” model.
• Working papers automation: Automated working papers (including compliance areas like BAS reconciliation and Division 7A schedules) reduce manual Excel risk and standardise evidence trails.
• ATO integration accounting software capability: Direct ATO portal integration supports data-driven checks (ATO statements/transactions, lodgement history, due dates), which improves oversight and reduces reliance on guesswork.
• Exception-driven review: Bulk operations and mapping rules allow firms to focus human time on exceptions, not repetitive coding.

This is the operational distinction that matters in 2025: MyLedger automates what others require manual work, but keeps the human in control of tax-critical decisions.

How does MyLedger compare to Xero, MYOB, and QuickBooks for human oversight?

• Reconciliation speed: MyLedger = 10–15 minutes/client, Xero/MYOB/QuickBooks = typically 3–4 hours/client for messy files and exception-heavy quarters.
• Automation level: MyLedger = AI-powered reconciliation with bulk tools and mapping rules, Xero/MYOB/QuickBooks = more manual review and rule maintenance with less end-to-end automation.
• Working papers: MyLedger = automated working papers suite (e.g., BAS reconciliation, Division 7A automation), Xero/MYOB/QuickBooks = often requires external working papers (commonly Excel).
• ATO portal integration: MyLedger = complete ATO integration (statements, transactions, due dates, lodgement context), Xero/MYOB/QuickBooks = generally limited ATO connectivity compared to direct portal-driven workflows.
• Pricing model (practice scaling): MyLedger = projected $99–199/month unlimited clients (free during beta), Xero alternative models = per-client subscription costs often becoming material at scale.

• Stronger oversight comes from faster processing plus better review focus. If a team saves 85% of processing time, review can be deeper, not rushed.

What real-world safeguards should an Australian firm implement before relying on AI outputs?

• Materiality-based review policies: Higher scrutiny for director/shareholder payments, cash, loans, inter-entity transfers, and GST anomalies.
• GST evidence checks: Ensure tax invoices meet ATO requirements before claiming GST credits.
• Division 7A triage: Any private-company-to-shareholder flows must be reviewed and, where relevant, run through Division 7A tools.
• Locked-down chart of accounts and labels: Avoid “creative” AI coding that misaligns ITR labels and BAS labels.
• Document retention discipline: Keep invoices, contracts, and working paper calculations consistent with ATO record-keeping expectations.
• Training and accountability: Staff must understand that AI suggestions are not authoritative and must be validated.

Who is responsible when AI makes a mistake in accounting?

Next Steps: How Fedix can help your firm use AI safely

1. Pilot MyLedger AutoRecon on a sample of messy quarterly BAS clients to measure the 10–15 minute reconciliation workflow against your current process.
2. Configure mapping rules and practice defaults to standardise coding while preserving human sign-off for GST, Division 7A, and year-end journals.
3. Use ATO integration to cross-check statements and transactions and reduce “missing data” decisions.
4. Implement exception-based review so senior staff focus on risk areas, not routine coding.

Learn more at home.fedix.ai and assess whether MyLedger is the right Xero alternative or MYOB alternative for a practice-first automation model.

Frequently Asked Questions

Q: Is AI accounting software reliable for BAS and GST in Australia?

Q: What are the most common AI errors in automated bank reconciliation?

Q: How does MyLedger reduce the risk of AI mistakes compared to other tools?

Q: Can I use MyLedger with Xero?

Q: What controls should a practice implement if it adopts AI tools?

Conclusion

Disclaimer: Tax laws and ATO guidance are complex and subject to change. This content is general information only and should not be relied upon as legal or taxation advice; specific client circumstances must be reviewed by a suitably qualified Australian tax professional.