A practical guide to preparing individual tax returns with AI

Preparing an individual return is not one task. It is roughly a dozen small ones stitched together, and only some of them benefit from automation. Understanding which is which is most of the work.

Walk through what a preparer actually does: chase the client for documents, sort what arrives, read each document, transcribe the relevant figures into a workpaper, reconcile against pre-fill data, chase the gaps, apply the judgement calls, and hand it up for review. The transcription and reconciliation steps are mechanical and consume a disproportionate share of the time. The judgement calls are where a preparer earns their fee. A well-designed automation takes the first category and leaves the second entirely alone.

The six stages of the pipeline

1. Intake and classification

Documents arrive by every route imaginable — email attachments, a client portal, photographs of paper, a shared drive that somebody remembered to update. The first job is simply working out what each file is: a PAYG summary, a private health statement, an annual tax statement from a managed fund, a rental agent's summary, a receipt, or something irrelevant that arrived by accident.

Classification is the least glamorous stage and the one that most determines whether the rest works. It is also very tractable — these documents have consistent structure even when their layout varies wildly between issuers.

2. Extraction

Once a document is classified, extraction pulls the specific fields that matter for that document type. The critical design decision here is that extraction should be typed and constrained, not open-ended. You are not asking a model to "summarise this document" — you are asking it to return a defined set of named fields, each with an expected format, or to say explicitly that a field is absent.

3. Normalisation

Extracted values are rarely usable as-is. Amounts arrive with currency symbols and thousands separators, sometimes in brackets for negatives. Dates arrive in three different formats. Entity names arrive with and without "Pty Ltd". Normalisation converts everything to a consistent internal representation before anything else touches it — and this is deterministic code, not a model. Anything that can be done with a rule should be.

4. Mapping to the workpaper

Normalised fields are then written into the firm's own workpaper template — the same one the team already uses. This matters more than it sounds. An automation that produces its own novel output format asks the whole team to change how they work; one that populates the existing template asks them to change nothing except how long it takes.

5. Reconciliation and flagging

This is where the automation earns its keep beyond typing speed. With everything structured, you can run checks that a preparer would do slowly or not at all:

• Does the sum of extracted income agree to the pre-fill data, and if not, by how much and where?
• Did a source that appeared last year not appear this year — a closed account, a sold property, a fund that has not issued yet?
• Are there deductions materially out of line with the client's own prior year, or with comparable clients?
• Did any document fail to classify, or classify with low confidence?
• Are there duplicate documents — the same statement supplied twice under different filenames?

Each of these becomes a flag on the workpaper rather than a silent correction. The automation's job is to raise the question, not to answer it.

6. Review handoff

The preparer opens a workpaper that is already populated, with every figure linked to its source and a short list of flagged items at the top. Their work becomes verification and judgement rather than transcription and judgement. That is the whole proposition.

---

The two places it reliably goes wrong

Confidence that is not calibrated

The failure mode that costs firms the most is not extraction being wrong. It is extraction being wrong confidently. A pipeline that reports high confidence uniformly gives reviewers no signal about where to look, so they either check everything (no saving) or check nothing (unacceptable risk).

A system that is right 97% of the time and knows which 3% it is unsure about is worth far more than one that is right 99% of the time and cannot tell you where.

Getting this right is mostly an engineering problem rather than a model problem: cross-checking extracted values against independent sources, comparing totals that should agree, and treating any disagreement as a flag regardless of what the model reported about its own certainty.

Silent handling of the missing case

The second failure is quieter. A document does not arrive; the pipeline runs anyway; the workpaper looks complete. Nothing is wrong on the page — the problem is what is not on it. A pipeline needs an explicit expectation of what should be present for this client, derived from the prior year and from pre-fill, and it must flag absence as loudly as it flags error.

What stays with the accountant

It is worth being explicit about the boundary, because it is what makes the workflow acceptable to a reviewing partner. The automation does not decide deductibility. It does not apportion private use. It does not determine residency, choose a depreciation method, or decide whether an expense is capital. Every one of those is a judgement with professional and regulatory consequences, and every one stays with a person.

What the automation does is ensure that when the accountant makes those calls, they are looking at a complete, structured, source-linked picture — and that they got there in fifteen minutes rather than ninety. The related principle, and how to structure the review gates so the team actually trusts the output, is covered in human in the loop.

Where to start

Do not start with your hardest clients. Start with a segment where the document set is predictable — salary income, a handful of investments, standard deductions — and where you have enough volume to see the pattern within a season. Get that segment fully working, including the flags, before extending to rental properties, trust distributions, or anything involving a foreign income statement.

The narrow version is live in weeks and tells you something true. The comprehensive version is a project plan.