How to Use AI as a Tax Preparer: Uses, Risks and the Rules

Tax preparation has one property that makes it a specifically bad fit for blind trust in a language model: the rules change every year, they depend on jurisdiction and filing status, and a model has no way of knowing whether the threshold it just quoted you is this year’s, last year’s, or from a provision that has since been repealed. It will state any of the three with identical confidence. That single fact should shape every decision about where these tools belong in your workflow.

This is the practical companion to will accounting be replaced by AI, which looks at the risk to the job itself. This one is about getting real value out of the tools without creating a problem you have to explain to a client, or worse, a regulator. If your work sits closer to general bookkeeping than return preparation, how to use AI as a bookkeeper covers the adjacent version of this.

Start with the obligations, because they narrow the options

The professional bodies have been explicit that the ethical framework does not change because the technology did. Seven UK tax and accounting bodies, AAT among them, published topical guidance on the ethical use of AI in tax work, and its central instruction is blunt: outputs from AI tools should never be treated as authoritative tax or legal advice, and every output needs review by a qualified professional in the specific context of the client. The Journal of Accountancy’s 2026 piece on accounting ethics and AI makes the same point in one sentence worth remembering: professional judgment cannot be delegated to AI, and responsibility for the final work product rests with you whether the tool is in-house or third-party.

Practically, that gives you three tiers of work: things you can do with anonymised or generic inputs in any tool, things that need a tool with an enterprise agreement and no training on client data, and things you should not hand over at all. A written policy on which tasks fall into which tier — even one page — is worth having before you start, and the AICPA’s small firm generative AI policy template is a reasonable starting point if your practice has nothing written down. The International Ethics Standards Board for Accountants makes the underlying principle explicit in its own July 2026 publication on emerging technology: professional accountants, tax preparers included, remain responsible for the judgements and decisions in their work regardless of which tool produced the draft.

What it is genuinely good at

The pattern across all of these: the model produces a draft or a structure, and you supply the tax judgement. None of them involve the model being the authority on anything.

1. 01Client correspondence. Explaining why a deduction was disallowed, why a refund is smaller this year, or what a notice actually means, in plain language, without writing it from scratch for the fortieth time. This is the highest-volume, lowest-risk win in most practices.
2. 02Turning a client’s messy notes into an organised intake. Paste anonymised, unstructured records of income and expenses and ask for a categorised summary to review — you are checking a draft, not trusting a filing.
3. 03Explaining a new rule in plain English once you have already read the primary source yourself. Useful for turning IRS or HMRC guidance into a client-facing paragraph; not useful as the first place you learn what the rule says.
4. 04Formula and spreadsheet construction. Describing what you want in words and getting a formula back for a client worksheet, easy to verify by running it against a return you can compute by hand.
5. 05Rehearsing a client conversation. Ask it to raise the objections a client is likely to raise about a position you plan to take, so you walk in prepared rather than improvising.

A worked example

A client’s home-office deduction got flagged and you need to explain, in a way they will actually read, why part of it was disallowed.

The bad version: paste the client file and ask for "an explanation of the home office deduction rules". You get a generic, textbook answer that may not reflect this year’s square-footage method, this client’s specific facts, or the exact reason the disallowance happened — stated with the same fluent confidence whether it is right or wrong, because that confidence is a property of the system, not a sign it has actually checked anything.

The version that works: strip identifiers, state the specific facts you have already established yourself — which method was used, which figure exceeded the allowable percentage — and ask for a 150-word explanation for a non-specialist client that leads with the actual reason. You get a usable draft, you check the two numbers against your own working papers, and you have saved fifteen minutes without outsourcing the one judgement that mattered: what actually happened on this return.

Use it where you already know the answer and the work is expressing it clearly. That is the whole rule, and tax preparation is exactly the field where skipping it costs the most.

Where it will quietly get you into trouble

• —Citing tax code or thresholds from memory. A model will produce a section number, a dollar threshold or a filing deadline with total confidence, and it may be from last year, a different jurisdiction, or a repealed provision. Treat every figure and citation as unverified until you have checked the primary source yourself — this is the single most common and most expensive way a tax preparer gets burned by these tools.
• —Arithmetic across a full return. A language model can produce a plausible total that is simply wrong, and unlike tax software it will not flag an inconsistency — it will just be off. Compute in software built to compute; use the model for language, not the return itself.
• —Anything that depends on a fact only the client knows. It will fill a gap with a plausible assumption rather than ask you to check, unless you have explicitly told it to flag missing information instead of guessing.
• —Classification questions dressed as facts. "Is this a business expense or personal" gets an answer delivered with the tone of certainty and none of the accountability. You still own that call, and so does your engagement letter.
• —Anything client-identifying pasted into a consumer tool with no enterprise agreement. What AI is actually bad at covers the general failure pattern; the confidentiality risk here is separate from and in addition to the accuracy risk.

Getting started without a project plan

1. 01Write down your practice’s rule on client data before doing anything else, even if it is one sentence. A short written policy beats no policy, which is what most small practices currently have.
2. 02Pick one recurring, low-risk task where you can judge the output instantly — client correspondence is the usual right answer.
3. 03Do that task both ways for two weeks, your way and with a drafted version, and time both. You want evidence of the gain, not a feeling that it helped.
4. 04Build a small library of the prompts and instructions that produced usable drafts. The value compounds through reuse.
5. 05Expand to a second task only once the first is routine. Adopting five workflows at once tends to leave a practice using none of them well.
6. 06Decide, in advance, what you will tell clients if asked whether AI touched their return. Disclosure norms in this field are still forming; having a considered answer is better than improvising one.

The part that actually protects your job

The tasks these tools genuinely speed up — organising intake, drafting correspondence, structuring a first pass — were already the tasks furthest from judgement, and they were already under pressure from tax software that predates any of this. What is left is advice, exception handling, the conversation with a client about a decision they actually have to make, and being the named preparer who signs the return. That work is not automated by a system that cannot be held accountable for being wrong, which is the whole reason a signature exists. How to use AI as an accountant makes the same argument for the broader field, and how to use AI as an auditor covers the version of it shaped by evidence standards rather than filing deadlines.

The useful stance is neither refusing the tools nor deferring to them. It is being the preparer who can produce a client explanation in a fraction of the usual time and then catch the one thing in it that is wrong before it goes out. That second half is the actual skill — the same one covered from the client-facing, suitability-driven side in how to use AI as a financial advisor — and it is worth practising deliberately rather than assuming it comes for free with the tool.

The short version

Use AI for client correspondence, organising intake, plain-language explanations and formula construction, on anonymised inputs or inside a properly vetted tool. Never trust it for a tax code citation, a threshold, arithmetic across a return, or a classification judgement you would have to defend — verify every one of those against the primary source yourself. Write down a data rule before you start, pick one task, measure the time you save, and expand only once the first one is boring.

Coursium teaches exactly this layer — practical use of the tools plus the checking habit that has to go with it. Stay ahead of AI by learning them on your phone.