AI handled most of the routine bookkeeping.
That was not the problem.
In this illustrative case study, an 11-person software consultancy generates approximately $1.8 million in annual revenue. It uses AI-assisted categorisation, receipt extraction and suggested matching to reduce repetitive transaction processing in Xero.
A sample of 360 routine transactions shows 318—approximately 88%—requiring no change. The remaining 42 items are where context, evidence or accounting judgement matters.
The useful question is not whether AI can do bookkeeping work. It can. The question is how the workflow responds when the transaction is unfamiliar, material, incomplete or dependent on information outside the bank description.
What the automation handled well
The strongest results came from transactions with stable patterns and clear supporting information:
- recurring software subscriptions from known suppliers;
- bank fees and standard merchant charges;
- repeated supplier transactions with consistent treatment;
- invoice-to-payment suggestions;
- extraction of dates, values and supplier details from clear documents; and
- retrieval of routine information from the accounting file.
These are valuable improvements. An accountant should not need to recode the same unambiguous monthly subscription forever.
Across the sample, automation reduced routine handling and allowed review time to be directed toward exceptions. The issue was not the percentage processed automatically. It was whether the remaining items stayed visible until someone resolved them.
The 42 transactions that required review
Eighteen transactions were ambiguous
Bank descriptions did not provide enough context to determine the correct treatment confidently.
A retail purchase could have been staff supplies, a client event, travel-related food or a private purchase. A generic transfer description did not reveal which entity or account was involved.
AI could infer from previous patterns, but previous patterns were not evidence of the purpose of a new transaction.
The right response was to ask a focused question or obtain a document—not to turn uncertainty into a confident guess.
Eleven transactions went to the wrong account
The system repeated a pattern that had previously been accepted. Once the original treatment was wrong, consistency scaled the mistake.
The values were individually modest, but the classification affected how management viewed project-delivery costs compared with overheads.
This is why exception review should consider both confidence and management relevance. A perfectly reconciled bank account can still contain transactions coded to the wrong place.
Seven transactions involved timing or judgement
One was a $14,400 annual software subscription recorded entirely as the current month's expense. Another was a $32,000 project deposit received before an eight-week implementation began.
The bank feed showed that cash had moved. It did not establish the period in which management should recognise the economic activity.
These decisions depended on the contract, service period, reporting purpose and applicable accounting requirements. Pattern recognition alone was not enough.
Six transactions were missing evidence
The system could process the information it received. It could not reconstruct documents that had never entered the workflow.
Four supplier payments had no supporting invoice or receipt. Two entries related to a loan statement that had not been provided.
This is a fundamental limit of automation: it is much better at processing available information than identifying every piece of information that should exist outside the system.
External reconciliation and evidence checks still matter.
The issue was the review architecture
The risk was not that AI had touched the books. It was whether anyone had defined:
- which work could be processed automatically;
- what confidence level was acceptable;
- which accounts always required human review;
- what value or risk triggered escalation;
- how unusual transactions were queried;
- who checked the balance sheet;
- how corrected treatments affected future rules;
- who owned unresolved exceptions; and
- what had to happen before the month was considered closed.
Without that structure, AI bookkeeping can become unattended bookkeeping. With it, automation can remove a substantial amount of low-value repetition while making exceptions more visible.
Five tasks suited to careful automation
1. High-confidence recurring categorisation
Known supplier, consistent business purpose, established treatment and low judgement.
2. Document extraction
Capturing supplier, date, invoice number, GST and value from clear bills and receipts.
3. Suggested matching
Identifying likely relationships between payments, invoices and bills while preserving review and reversibility.
4. Anomaly detection
Surfacing values, suppliers, account codes or timing that differ from normal patterns.
5. Draft questions and explanations
Turning an unresolved item into a focused request for the missing context or evidence.
The machine reduces administration. The accountant remains responsible for deciding what information is needed and how the transaction should be treated.
Five areas that deserve explicit review
- Novel or ambiguous transactions. If the system has not seen the fact pattern, uncertainty should remain visible.
- Material transactions. A $19 subscription and a $190,000 asset purchase should not share the same review threshold.
- Tax-sensitive or judgement-heavy treatment. Private use, unusual GST, capital-versus-expense and related-party transactions require appropriate review.
- Balance-sheet accounts. Automation can assist the reconciliation; someone still needs to establish that the balance represents something real.
- Month-end sign-off. Closing the month is an assertion about the reliability of the resulting information, not merely a processing milestone.
Questions to ask about your AI bookkeeping setup
- What exactly is automated?
- Can you see what the system changed, suggested or reconciled?
- Can actions be reversed?
- How are low-confidence transactions handled?
- Who reviews new suppliers and unusual transactions?
- Are material transactions treated differently?
- Does someone reconcile bank, loan and processor accounts to external records?
- Does someone review the balance sheet each month?
- Are unresolved questions visible in one place?
- When a treatment is corrected, does the workflow reduce the chance of recurrence?
- Who is accountable for saying the month is closed?
If the answer to the final question is “the software,” that is the first part of the process to change.
So, can AI run your books?
AI can run a growing share of the work required to keep books current. That is already useful.
Good accounting is not defined by the number of transactions a human touched. It is defined by whether the final records are reliable enough for their purpose, whether material exceptions have been resolved and whether somebody can explain the numbers.
The future of bookkeeping is unlikely to involve people manually coding every coffee, subscription and bank fee.
That sounds like progress.
The control layer matters more, not less.
Sources and further reading
- TPB guidance on the use of artificial intelligence and the Code of Professional Conduct ↗
- AI accounting software for Australian business — Xero ↗
- Human + AI in Accounting: Early Evidence from the Field ↗
This article is not tax, accounting or financial advice for your circumstances. Accounting treatment depends on the facts of a transaction and the requirements applying to the business. Seek professional advice before making a material accounting or tax decision.