A review checklist for AI-drafted tax work
A playbook for partners and tax managers who sign off tax work that AI helped draft. It sets out risk tiers, the four places AI tax errors cluster, a 12-point review checklist to copy, what to record on the file and the professional indemnity questions to put to your broker.
Trent McLaren · Publisher, The Firm
9 November 2026 · 9 min read
Learning check · before you start
What do you already have in place?
A few quick questions, so you can skip the steps you have already done. Each one has a line explaining what it means. Not sure, or never heard of it? Choose “Not yet / not sure”: that is where most firms start, and the playbook walks you through it.
This playbook is for the partner or tax manager who signs off tax work in a firm where some of it is now drafted with Claude, ChatGPT, Microsoft Copilot or a tax-specific AI tool. By the end you will have a set of guardrails, a review checklist for AI-drafted tax work that you can copy into your workpaper template, and a clear list of what to record on the file. Setting it up takes a morning; the checklist adds a few minutes to each review.
The Firm has already covered why AI gets tax wrong (missing facts, blended rules and hidden assumptions), why accountants should be the accountable layer behind AI tax answers, and why liability is the moat AI cannot cross. This piece is the procedure that sits underneath those arguments. It is written for Australian tax agents first, with pointers for UK and US firms.
Step 1: Decide which tax work AI may draft
Your task sort your firm's tax work into three risk levels for AI use, using real job types from your own clients, and write them into your AI policy.
You said you already have this covered. Skim this step, or skip to the next.
Not every piece of tax work carries the same risk. Sort yours into three tiers before anyone opens a prompt.
- AI may draft, normal reviewresearch summaries, first drafts of client explanations of settled rules, checklists of documents a client needs to provide.
- AI may draft, enhanced reviewadvice letters, workpaper commentary, anything that applies a rule to a client’s specific facts.
- AI may assist, not draftpositions on contested or novel areas, anything going to a tax authority, private ruling applications, amendments and voluntary disclosures.
The Tax Practitioners Board’s guidance on AI, TPB(GS) 55/2026, issued 22 July 2026, says whether a practitioner took reasonable care depends on all the circumstances, including “whether AI tools have been used for tasks which should be independently verified, particularly when outside the tax practitioner’s expertise.” Your tiers are how you show you thought about that in advance.
Done when: the three tiers are written into your AI policy, with examples from your own client base.
Learning check
What would you do? Pick one; every answer is explained.
Step 2: Make the AI show its working before anyone reviews it
Your task save a standard instruction for your AI tools so every tax draft lists the facts it used, the law it applied, the country and income year, and anything it was unsure about, and tell reviewers to send back any draft without these.
You said you already have this covered. Skim this step, or skip to the next.
A reviewer cannot check what they cannot see. Require every AI-drafted tax output to arrive with four things attached, and put them in your standard prompt so nobody has to remember:
- The facts it relied on, listed, with any it assumed marked as assumed.
- The provision, ruling or guidance it applied, by name and number.
- The jurisdiction and the income year or date the answer applies to.
- Anything it was unsure about.
This does not make the output right. It makes the errors visible.
We’ve all seen it hallucinate, forget, run out of context.
Done when: the four-part request is saved as a firm prompt or project instruction, and reviewers send back any draft that arrives without it.
Learning check
What would you do? Pick one; every answer is explained.
Step 3: Check the four places errors cluster
Your task on your next three AI-assisted tax jobs, check the four places AI errors usually appear (out-of-date law, rates and thresholds, rules from the wrong country, and invented citations) and note what you find.
You said you already have this covered. Skim this step, or skip to the next.
AI tax errors are not random. They gather in four places, so look there first.
Stale law
A model’s training stops at a date, and tax law keeps moving. It may describe a rule as it stood before an amendment, or miss a measure that has since started. Ask what date the answer is good to, and check any rule that has been in the news in the last two years against the current primary source.
Thresholds and rates
Numbers go stale faster than rules. The superannuation guarantee rate, for example, rose to 12% from 1 July 2025; a tool working from older material can still give the earlier rate with full confidence. Check every threshold, rate, cap and cut-off date against the ATO, HMRC or IRS page for the right year.
Jurisdiction mixing
General models can blend tax systems. Watch for US terms in an Australian answer, UK allowances in a New Zealand one, or a state-level rule presented as federal. If the answer would read the same in another country, be suspicious.
Confident, fabricated citations
The most dangerous error looks the most professional: a ruling number, case name or section reference that does not exist, or exists and says something else. Open every citation. If you cannot find it on the regulator’s own site or a legislation database in a minute, treat it as invented until proven otherwise.
Done when: reviewers have checked all four on their next three AI-assisted jobs and logged what they found.
Learning check
What would you do? Pick one; every answer is explained.
Step 4: Run the review checklist
Your task copy the 12-point review checklist for AI-drafted tax work into your review sheet, and have a partner use it on one live job.
You said you already have this covered. Skim this step, or skip to the next.
The TPB guidance says practitioners “should verify and review AI generated content for accuracy throughout each step of the workflow and should establish processes to understand and contest AI decisions/outputs.” The checklist below is one such process. Copy it into your review sheet.
Review checklist for AI-drafted tax work. Tick each line or write why it does not apply. A draft with an unticked line does not go to the client.
- The facts the AI relied on match the client file, and every assumed fact has been confirmed with the client or removed.
- The jurisdiction and income year are stated and correct.
- Every provision, ruling, case and guidance reference has been opened at its source and says what the draft claims.
- Every threshold, rate, cap and date has been checked against the official source for the right year.
- Any rule changed in the last two years has been checked against current law, not the AI’s summary.
- The answer does not borrow terms or rules from another country.
- Calculations have been reworked independently, by hand or in a tool that gives the same result for the same inputs.
- The reasoning holds if one key fact is different, and the client letter says which facts the advice depends on.
- The reviewer would give the same advice without the AI draft in front of them.
- The work is in the right tier from Step 1, and enhanced-review work has had a second reviewer.
- Client information went only into a tool the firm has approved, with client permission where needed.
- The AI use, the checks made and the reviewer’s name are recorded on the file.
Line 9 matters most. If the reviewer could not have reached the advice alone, they are not reviewing it, they are trusting it. The TPB says practitioners “should not rely on AI output as a substitute for their own analysis of a client’s circumstances.”
Done when: the checklist is part of the review sheet for every AI-assisted tax job, and a partner has used it on one live job.
Learning check
What would you do? Pick one; every answer is explained.
Step 5: Record it on the file
Your task add an "AI use" section to your workpaper template with five fields: the tool and what it was used for, the prompt, the completed checklist, facts confirmed with the client, and the reviewer's name and date.
You said you already have this covered. Skim this step, or skip to the next.
In Australia, section 30 of the Code Determination requires records that correctly record the tax agent services you provided, and TPB(GS) 55/2026 says of the AI review steps that “Each of these steps should be documented.” From 1 October 2026 a failure to comply with the Code can also attract a civil penalty, which makes the file note worth more than it used to be.
For each AI-assisted tax job, record:
- Which tool was used, and for which part of the work.
- The prompt or instruction, or a reference to the firm’s standard one.
- The completed checklist, with what was corrected.
- Facts confirmed with the client after the AI draft.
- The reviewer’s name and the date.
File it against the job in your practice management system, wherever the rest of the record lives. Our piece on AI meeting notes and professional indemnity explains why a reviewed, dated record is what a claim turns on.
Done when: your workpaper template has an “AI use” section with these five fields.
Learning check
What would you do? Pick one; every answer is explained.
Step 6: Check your professional indemnity position
Your task ask your insurance broker, in writing, three questions about how your professional indemnity policy treats AI use, and keep the answers on file.
You said you already have this covered. Skim this step, or skip to the next.
Registered tax agents in Australia must hold professional indemnity cover that meets the TPB’s requirements. What we could not find, as at November 2026, is a published Australian insurer position applying a general exclusion to AI-assisted accounting or tax work. That does not mean your policy is silent. Wordings differ and change at renewal, so check your own policy and ask your broker:
- Does the policy mention AI tools, with any exclusion, condition or notification duty?
- Does the proposal form ask about AI use, and is your last answer still accurate?
- Would the insurer expect a documented review process like the one above?
Ask every provider who carries the liability.
For a general-purpose AI tool, the answer is you.
UK firms will find the matching standards in Professional Conduct in Relation to Taxation, and US firms in the due diligence rules in Circular 230. The source of truth for Australian practitioners is the TPB’s AI guidance. This is journalism and general information, not legal or insurance advice.
Done when: you have your broker’s answers in writing and on file.
Learning check
What would you do? Pick one; every answer is explained.
Free to join
Join The Firm free to read the rest
Members get every playbook in full, plus free templates, downloads and tools across the site. Join more than 35,000 accounting professionals. All it takes is your email.
Where we could be wrong
The assumption most likely to be wrong is that a checklist is enough. On a genuinely complex structure, a reviewer working from a confident AI draft may anchor on it even while ticking every line, and line 9 is the only guard against that. We also have not verified any insurer’s actual treatment of AI-related claims, because none we found has published one. “When AI Gets Tax Wrong: Guardrails for Complex Tax Reasoning” with Andrew Noble of LodgeiT Labs is on the agenda for The Firm’s AI in Practice Summit on 11 and 12 November 2026, and notes will be at the summit notes page after the event.
Frequently asked questions
Do I have to tell clients that AI drafted their tax advice?
The TPB’s guidance is about permission to disclose client information to AI tools, not about labelling each output. An engagement letter line saying AI may be used under professional review covers both.
Is a tax-specific AI tool safer than a general one?
It may cite better sources and show more of its working, but the checklist still applies. The responsibility for the advice does not move to the vendor because the tool is built for tax.
How long should the review take?
For a simple, settled question, a few extra minutes. For an enhanced-review job, budget the time you would spend reviewing a junior’s first draft. If it takes longer than doing the work yourself, move that job type down a tier.
What do I do when I find a fabricated citation?
Remove it, find the real authority or drop the point, and log it. If the same tool invents references repeatedly, restrict it to tasks that do not need citations.
Does this apply to bookkeeping and BAS work too?
Yes. BAS agents are covered by the same Code obligations, and thresholds, rates and GST treatments are exactly where stale or blended answers show up.
Your one-page summary
Your answers, choices and decisions, on one page
Sponsored
AI in AccountingTaxQuality ControlReasonable CareProfessional IndemnityTPBAI 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.
