AI for tax research: what it gets right, what to check
General AI tools give fluent tax answers whether they are right or wrong, and summit speakers showed how they fill in missing facts, misread provisions and cite sources that do not exist. This guide gives a check table against ATO and legislation sources, prompts that ask for sources, and a review routine juniors can follow.
Ask a general AI tool a tax question and you get a fluent answer in seconds, often with section numbers attached, and it sounds the same whether it is right or wrong. At the AI in Practice Summit in November 2026, three tax research sessions landed on one point: AI can speed up research, but the checking is still yours, and juniors need a routine that teaches them to do it. Here is how to check an answer, what to prompt, and a routine a first-year can follow.
Where general AI goes wrong on tax
Andrew Noble, founder of LodgeiT, opened his summit session, When AI gets tax wrong, with three failure modes. A general model recalls from memory, so it can paraphrase and mix provisions up. It fills gaps: leave residency out of a scenario and it will probably assume a resident without saying so. And it leaves no trail to show the ATO how it reached a figure.
His story was about his own product: a LodgeiT worksheet, checked with AI, carried forward too much tax loss. Jenny Haines, a chartered accountant at Natural Numbers Accounting, kept saying it was wrong, went back to the Act, and was right.
It will go, okay, I’ve found an answer. It must be the right answer, and it often isn’t.
Other speakers added more:
- Training dataDavid Boyar of The Access Group said in The great AI con that general models learned heavily from the open internet, Reddit included, and must filter out bad tax advice.
- Made-up referencesBillie Mcloughlin, founder of AIccountancy, cited in her client data session the Deloitte report for a federal department where an academic found a reference that did not exist, and Deloitte refunded the final instalment. Her advice: always check your sources.
- Quiet errorsDev Magoon, a CPA in public practice who reviewed LodgeiT’s rules, found a capital gains calculator giving sensible answers while relying on a repealed subsection and misreading “active asset” in section 152-40. In his words, a quiet error “only fails on the client that you didn’t think about”.
Natalie Lennon, founder of Two Sides Accounting, said in her session not to ask a general assistant technical tax or legislation questions, and to spot-check everything it produces. The regulatory side is in AI and tax errors: understanding the new regulations.
Why the source behind the answer matters
The vendors at the summit pitched one fix: limit what the AI can draw on. These are their descriptions of their own products, untested by The Firm. Grant Abbott, a director of TaxGuru, said it pulls the relevant paragraphs of cases, sections and ATO rulings from its own corpus, and added that “everything needs to be verified”. Andrew Noble said LodgeiT’s Tax Genii limits the AI to retrieved chunks of legislation, refuses questions outside its rules, and asks for missing facts such as residency. A narrower source makes an answer easier to check, not optional to check.
The status of the source matters as much as its existence. Sami Charaf, a lawyer and IPA fellow who also reviewed LodgeiT’s rules, said “The ATO worksheets are not authoritative”, and that ATO website guidance outside a binding ruling is not either. The ATO’s own page, How our advice and guidance protects you, sets out the levels as at November 2026:
- Public rulings (TR, TD, GSTR, LCR), class, product and private rulings are legally binding: rely on them and you are protected from the shortfall, interest and penalties.
- Website material, ATO interpretative decisions and practical compliance guidelines are not legally binding. Relied on in good faith, they protect against penalties and interest, not the shortfall.
- Edited versions of private rulings carry no protection at all.
An AI answer resting on an ATO web page or edited private ruling rests on something weaker than the law.
The law you check against keeps moving
A model that learned the law at one point cannot know what changed since. Some of 2026’s dates:
Key dates
- 24 March 2026PassedThe Tax Practitioners Board (TPB) released its exposure draft on AI and the Code of Professional Conduct.
- 17 June 2026PassedThe ATO published a decision impact statement on Commissioner of Taxation v Hall [2026] FCAFC 43. The Legal database page for TR 2021/1, on employees’ transport expenses, now says it is under review because of that decision.
- 22 July 2026PassedThe TPB issued TPB(GS) 55/2026, its final guidance on AI and the Code.
- 1 October 2026PassedDate of the latest Income Tax Assessment Act 1997 compilation on the Federal Register of Legislation, which lists 385 versions of the Act.
A tool working from memory will miss a notice like the one on TR 2021/1.
Check every answer
TPB(GS) 55/2026 says AI models may hallucinate and that practitioners should verify AI output at each step of the workflow and document it. This table is one way to do that.
| Claim type | Where to verify | Red flags |
|---|---|---|
| A section of an Act | The compilation in force for the relevant income year on legislation.gov.au | A section that does not exist, a repealed subsection, quoted words that are not in the text, the wrong Act |
| A public ruling or determination | The ATO Legal database | A number that returns nothing, a notice that it is withdrawn or under review, a quoted paragraph that is not there |
| A court or tribunal case | The judgment itself, plus any ATO decision impact statement in the Legal database | A case name with no citation, the wrong year, a conclusion the judgment never reaches |
| A rate, threshold, cap or date | The ATO page for that specific year (the ATO marks year-specific information) | No year stated, a figure from an earlier year, a suspiciously round number |
| ATO guidance | The ATO page, read against the protection table above | Web guidance or edited private advice treated as binding law |
| The client’s facts | Your file and the client | An assumption the tool never asked about: residency, ownership, dates, related parties |
| Anything foreign | Nowhere: it should not be there | IRS, HMRC, VAT, 401(k), state income tax, a tax year that does not run 1 July to 30 June |
Prompts that ask for sources
Grant Abbott’s point was that tools need the full facts. These prompts work in any tool and expose the gaps before you rely on an answer.
- Facts first. “Australian income tax, 2025-26 income year. Facts: [facts]. Before you answer, list every fact you would need that I have not given you, and every assumption you would otherwise make.”
- Sources per step. “Answer only from Commonwealth legislation and ATO public rulings. For each step, give the Act and section, or the ruling number and paragraph. Where you cannot point to one, write ‘no source’ instead of filling the gap.”
- Status check. “For each source you cited, say whether it is legislation, a binding public ruling, ATO guidance, a case or something else, and when it was issued.”
- Argue the other side. “Which facts would change your conclusion, and which provision would apply instead?”
Section numbers in reply to prompt two are a list to open, not proof. Our AI tax review checklist covers what to record once you have.
A review routine your juniors can follow
Maureen Gabriel, a director of TaxGuru, said in the tax research session that juniors learned by being challenged: who owns the assets, what happened before, have you considered trusts.
We end up confusing the ability to obtain an answer with the ability to understand and assess it, and those are two very different things.
A routine that keeps that learning in place:
- Write down the facts and income year before opening any tool. Missing facts come from the client, not the AI.
- Use the four prompts in a firm-approved tool. TaxGuru’s Michael Jeffriess warned staff may be putting client data into personal accounts.
- Open every cited source, read the paragraph relied on, and mark it checked, wrong or missing.
- Record it on the file: question, tool, sources checked and the date.
- Bring the partner the questions, not just the answer: the facts, the issues, what the tool assumed and what you are unsure of.
- The partner challenges the reasoning and signs off. Responsibility stays with the practitioner, as the TPB says.
Step five is where the learning happens. More in How junior accountants learn tax without partner guidance, and on who carries the advice in The role of accountants in AI tax advice accountability.
This is general information, not tax advice. For the law and the ATO’s view, start with the ATO Legal database and the ATO advice and guidance pages.
Where we could be wrong
The weakest assumption is that checking every citation is affordable on every job; on a simple question a senior may reasonably choose a lighter check, so scale the routine to the risk. We have not tested TaxGuru, Tax Genii or any general model; what is said about them here is what their own people said on stage. And the TPB guidance is new, so how the Board applies it will matter more than its wording.
Frequently asked questions
Can I rely on ChatGPT or Claude for Australian tax research?
As a starting point, not a finish. Andrew Noble described how general tools fill in facts they were never given. If you use one, ask for section and ruling numbers and read each yourself before anything reaches a client.
Can I skip checking with a specialist tax research tool?
No. The vendors at the summit said answers need verifying, and the TPB asks for review at each step. Cited sources make the check faster; your sign-off still rests on what you read.
What does the TPB expect when I use AI on tax work?
TPB(GS) 55/2026, issued 22 July 2026, says you stay accountable, should not substitute AI for your own analysis, and should verify and document each step. It is guidance on how your existing Tax Agent Services Act 2009 obligations apply to AI.
How do I know if an ATO ruling is still current?
Open it in the ATO Legal database. A ruling’s page can carry a notice that it is withdrawn, amended or under review, as TR 2021/1’s did in November 2026, and links related documents. A ruling number in an AI answer says nothing about its status.
How long should juniors spend checking an AI answer?
Set the depth by risk. A client under ATO review or a restructure warrants reading every source the answer depends on; a routine query may need only the key provision. Record what was checked either way.
TaxAI in AccountingTax researchATOTPBAI in Practice Summit
Get the editable AI tax work review checklist Download the AI tax work review checklist
The Word document (.docx), ready to fill in. Your profile unlocks every template on The Firm.
Or build it with AI
Open Claude or ChatGPT with the prompt filled in. It writes the finished document for your firm straight away, then asks a few questions to tailor it. No client data needed.
Works without client data. Don't paste client names, tax file numbers or ledgers into an AI tool your firm hasn't approved for them. AI data safety playbook.