AI & Automation
AI bookkeeping: what it does well, what it gets wrong, and what still needs a human
An honest account of where AI genuinely helps in Australian bookkeeping — coding, reconciliation, document handling — and the judgement calls it should not be making.
We run AI agents across our own practice, including on client bookkeeping. So this is not a theoretical piece — it is a report on what actually works, written by people who have to fix it when it does not.
What it genuinely does well
Transaction coding at scale. Bank feed descriptions are short, noisy and inconsistent, and a language model reads them far better than a rules engine. "SQ *MITCHELLS HARDWA SYDNEY" is obviously hardware to a model and opaque to a keyword rule. On a file with reasonable history, coding accuracy is high enough to be genuinely useful.
Document reading. Supplier invoices, receipts and statements in every layout imaginable. Extraction of supplier, ABN, date, GST and line items is now reliable enough to be the default path, with exceptions routed to a person.
Finding what looks wrong. This is where the real value sits, and it is underrated. An agent can review a whole ledger and flag the things a human would only catch by luck:
- A supplier normally coded to one account suddenly coded elsewhere
- GST claimed on a supplier who is not registered
- A payment that looks like a duplicate of one three weeks earlier
- An expense that is 400% of its usual monthly value
- A transaction dated before the lock date
- A transfer reconciled on one side and not the other
A human reviewing 4,000 transactions gets tired by line 300. An agent applies the same attention to line 4,000, and that consistency is the actual product.
Drafting the commentary. Turning a set of variances into a written explanation for a client. The numbers come from the ledger; the agent writes the paragraph; a human checks whether the explanation is true.
What it gets wrong
Anything requiring facts not in the data. A $12,000 payment to a builder is capital improvement or repairs depending on what was actually done. The bank feed cannot tell you. Neither can the invoice, half the time. The model will guess, and it will guess plausibly, which is worse than guessing obviously.
GST status of unusual suppliers. Models will confidently apply GST to a supplier that is not registered, or treat a GST-free supply as taxable. This is checkable against the ABR, which is exactly why it should be a validation rule rather than a judgement call.
Related-party and private transactions. Whether a payment is a loan, a drawing, a wage or a genuine expense depends on intent and on Division 7A consequences. That is a structural decision with tax outcomes and it needs a human who knows the client.
Cutoff and accruals. Deciding which period a transaction belongs in requires understanding when the work was performed, which is rarely in the transaction data.
Anything where being confidently wrong is expensive. Payments, payroll, and anything lodged with a regulator.
The design that makes it safe
Every AI bookkeeping process we run follows the same three rules.
Read-only by default. Agents analyse and propose. They do not post, pay or lodge. The number of ways an autonomous agent can quietly corrupt a ledger is larger than the time it saves.
Validate everything mechanically. Do not trust the model's arithmetic — recalculate it. Does GST equal one eleventh of the total? Does the ABN exist and is it registered for GST? Does the sum of lines equal the invoice total? Is the date inside the period? These are cheap deterministic checks and they catch the overwhelming majority of errors.
Confidence thresholds with escalation. High-confidence, low-value, matching-history items flow through. Anything unusual, anything large, anything new goes to a person. The right proportion sent to a human is not zero — it is whatever keeps the error rate acceptable.
What this changes about the bookkeeper's job
It does not remove it. It moves it.
The work that goes away is the mechanical part: coding known transactions, typing invoice data, matching obvious payments, scanning for the errors you can define in advance.
The work that grows is the part that was always the valuable bit and was always being squeezed:
- Deciding the treatment of things that are genuinely ambiguous
- Talking to the client about what a transaction actually was
- Reviewing the exceptions the system escalated
- Noticing the thing that is technically fine and commercially alarming — margins slipping, a debtor stretching out, a subscription nobody uses
- Getting the file ready for tax rather than merely balanced
A bookkeeper spending 80% of their time on data entry and 20% on judgement is a worse deal for the client than one spending 20% on entry and 80% on judgement, at the same price. That inversion is what the technology actually buys.
The honest limits, stated plainly
It does not make a bad file good. If the chart of accounts is a mess, the bank feed has duplicates and the opening balances are wrong, AI will code the mess faster. The clean-up is still a human job and it still has to happen first.
It does not remove the need to understand your own numbers. A report you did not question is not more trustworthy because a machine helped produce it.
It does not accept responsibility. When something is lodged incorrectly, the registered agent is accountable, not the software. That accountability is precisely why the human review step is not a formality.
Where we would start
If you are running a small business file and want the benefit without the risk:
- Fix the foundation first. Clean chart of accounts, correct bank feeds, current reconciliation, a lock date after each period is finalised.
- Use the coding assistance built into your accounting software. It is decent, it is included, and it is the lowest-risk starting point.
- Add document capture so invoices are read once rather than typed.
- Add an exception review over the whole ledger monthly — this is the highest-value step and the one almost nobody does.
- Keep a human on treatment decisions, payments and lodgement. Permanently, not as a transitional measure.
The businesses getting real value from AI in their books are not the ones that removed the human. They are the ones that stopped paying a human to do typing.
We run this exact stack across our own client files. If you would like your bookkeeping reviewed — or an exception sweep run over a file you are not confident in — book a free consult.