Tax return and financial reporting automation

Individual return preparation is the highest-volume compliance workflow in most Australian practices, and the one where the manual effort is least defensible. A preparer opens a PDF, reads a figure, types it into a workpaper, and repeats that forty times per return. None of that is professional judgement. All of it is billable time that clients increasingly resist paying for.

It is also the workflow where automation goes wrong most visibly, because a wrong number that looks right is worse than no number at all.

The six stages

1. Intake and classification. Documents arrive as a mess — a PDF bundle, six phone photographs, a forwarded email chain. The first job is working out what each document is: PAYG summary, dividend statement, private health statement, rental agent summary, deduction receipt.

2. Extraction. Pulling the figures out. Straightforward on a structured statement, considerably harder on a photographed receipt taken at an angle in poor light.

3. Normalisation. Getting everything into consistent units, periods and signs. This is where a lot of silent errors are introduced and where careful engineering pays for itself.

4. Mapping. Each figure into the right workpaper line, against your template rather than a generic one.

5. Reconciliation and flagging. Cross-checking against ATO pre-fill where available, prior-year comparatives, and internal consistency. Anything that does not tie gets flagged rather than quietly adjusted.

6. Review handoff. The output arrives with provenance attached. Every extracted figure carries a pointer back to the document, the page and ideally the region on the page, so a reviewer can verify a number in one click instead of re-opening the source bundle.

We walk through this in detail, with the checks that keep it defensible, in a practical guide to preparing individual tax returns with AI.

The two failure modes worth designing against

Confidence that is not calibrated. A model that is equally confident about a clearly printed PAYG figure and a smudged handwritten receipt is useless for review, because the reviewer cannot tell which numbers to trust. We build calibration in and surface it.

Silent handling of the missing case. The client sent eleven documents last year and ten this year. An automation that quietly proceeds has produced a return that is wrong in a way nobody will notice until the ATO does. Missing is a flag, never a zero.

Beyond individual returns

The same pipeline handles monthly and quarterly management accounts, where the inputs are more consistent and the payback is often faster. Populating a reporting pack from the ledger, writing the variance commentary against actual movements, and flagging the accounts that moved in ways worth explaining.

Company, trust and partnership returns are more variable and typically automate in parts rather than end to end — the workpaper population and the reconciliation, with the structural judgement staying manual.

What this does to the review step

Adoption fails when reviewers cannot see what the machine did. A reviewer handed a completed workpaper with no provenance will re-do the work, and you have added cost rather than removed it.

So the design target is not "the automation is accurate". It is "a reviewer can confirm the automation was accurate faster than they could have done the work themselves". Those are different problems, and only the second one saves money. Designing AI workflows your reviewers will trust covers the five decisions that determine which one you end up with.

Your obligations do not move

TPB(GS) 55/2026 is explicit: using AI does not reduce, share or transfer your professional responsibility. The return is still yours. What changes is what the file needs to evidence — which steps were AI-assisted, under what policy, and who reviewed the output. We build that trail as a by-product of the workflow rather than an administrative afterthought.