The Smartest Way Accountants and Bookkeepers Are Using AI During Tax Season

The headline you keep seeing about tax season is “AI does your returns for you.” That’s the wrong frame, and the practitioners actually getting value out of AI know it. The returns themselves still go through the tax engine. Intuit ProConnect Tax, Drake, CCH ProSystem fx, Lacerte, and the country-specific equivalents have integrated AI assistants of their own, and that’s where the actual numbers belong.

What AI changes during tax season is everything that surrounds the return. The intake organisation, the deduction review, the plain-English client memo, the triage in the last week. That’s where a small firm finds its 30 to 40 hours back, not in the tax engine itself.

What AI should and shouldn’t touch in tax season

The rule of thumb that holds up under review: AI is good at structuring unstructured information, drafting written explanations, and applying a consistent diagnostic checklist. AI is bad (and dangerous) at applying specific tax-code rules to specific facts without verification, and at any task where the answer needs to be auditable to a citation.

Translated to actual workflow, that means ChatGPT and Claude earn their keep on the document organisation pass, the deduction review pass, and the client communication pass. They do not earn their keep filing a return, picking which schedule to use, or interpreting an ambiguous regulation without you reading the source.

The AICPA and CPA.com 2025 AI in Accounting Report found 85% of accountants feel excited or intrigued by AI, but only 37% of firms invest in training. The firms doing tax season well have closed that gap, and they’ve drawn a clear line about which tasks the AI gets and which it doesn’t.

The intake pass: turn 47 PDFs into a tidy schedule in 10 minutes

Most small firms still receive client tax packs as a chaotic mix of phone-photographed receipts, bank statement PDFs, employer summaries, and one annotated screenshot. The intake task isn’t just sorting it, it’s working out what’s missing.

Once your OCR layer (Dext or Hubdoc for receipts, or the document module inside your tax engine) has extracted the structured fields, paste the resulting summary into the model with this:

“You are a tax season prep assistant for an accountant working on the [TAX YEAR] return for [CLIENT NAME], an individual / sole trader / small company in [JURISDICTION]. Below is the list of documents we have received, with extracted totals. (1) Group these documents into the standard return categories for this jurisdiction (income types, deductible expense categories, capital items, etc.). (2) Identify any document that looks miscategorised based on the payee or description. (3) Produce a ‘still missing’ checklist of documents we would normally expect for a client of this profile but have not received. (4) Flag any items where the amount looks unusual compared with what’s typical for this category. Plain output, no commentary.”

The “still missing” checklist is the high-value output. It’s the part of intake where humans get tired and skip steps in mid-March. The AI applies the same checklist to every client, every time.

The deduction review pass: the second pair of eyes that doesn’t get bored

The deduction review is where AI earns the bulk of its tax-season fee. You’re not asking it to decide what’s deductible (that’s a code question). You’re asking it to find candidates worth your judgment.

“You are reviewing categorised expense data for the [TAX YEAR] return of [CLIENT NAME], a [INDUSTRY] business in [JURISDICTION]. Your job is NOT to determine deductibility. Your job is to flag items that warrant review by the accountant. Produce three lists: (1) transactions that have ambiguous descriptions and could be either personal or business, (2) recurring subscriptions or memberships where deductibility depends on use (flag for confirmation, do not assume), (3) any one-off transaction over [THRESHOLD] that should be confirmed as business-related and properly substantiated. For each flagged item, write the single question you would ask the client. Make no statements about whether items qualify for deduction under [JURISDICTION] tax law.”

The instruction “make no statements about deductibility” matters. Without it, the model will helpfully tell you a marketing subscription is “likely deductible,” and that helpfulness is the failure mode that gets practitioners in trouble.

The failure mode: AI as a tax authority

Here is the mistake that has burned more than one practitioner: asking ChatGPT or Claude “is this deductible under [tax code section]?” and acting on the answer.

LLMs hallucinate citations. They will write you a fluent paragraph attributing a position to a section number that doesn’t say what the model claims, or that has been amended. Tax law also varies enough by jurisdiction (and by year) that a model trained on a generic corpus will confidently mix US, UK, Australian, and New Zealand rules without flagging the difference.

The defensible workflow is: the AI surfaces the question, you check the source. If you need a tax position confirmed, that’s what your tax research platform (CCH, Thomson Reuters, the IRD or HMRC site, your jurisdiction’s equivalent) is for. The AI is the triage layer. The authority is the source.

The client memo: where AI saves the most time

The single highest-time-cost task in a small tax practice during the busy weeks isn’t the return. It’s the explanation of the return to a client who doesn’t speak tax. AI flips that calculation.

“Write a one-page client memo for [CLIENT NAME] explaining their [TAX YEAR] tax outcome. Inputs: [PASTE the key figures from the return: total income, total deductions, tax payable or refund, any unusual items]. Structure: (1) one-sentence summary of the outcome, (2) what changed compared with last year and why (paste prior-year summary if relevant), (3) two specific actions for the year ahead that would improve the position, (4) one risk or area we want to monitor. Plain English, written to a small business owner, no jargon, no tax-code references, warm and confident tone, maximum 350 words.”

A senior accountant could write that memo in 25 minutes. With this prompt, the draft lands in 30 seconds and you spend 5 minutes editing for accuracy and adding your voice. Across 80 returns, that’s 25 hours of senior time saved in the busiest month of the year.

The counterintuitive insight worth keeping

Most articles tell you AI saves time in tax season. The genuinely useful insight is the opposite: AI is most valuable in the parts of tax season that aren’t time-pressured. Pre-season, when you’re setting up the intake checklist for every client tier. Post-season, when you’re writing the structured year-ahead memos. Off-season, when you’re building the prompt library your team will use next year.

The firms that look calm in March are the ones who did the AI work in December and January. They aren’t typing prompts under pressure, they’re using prompts they already wrote, on a workflow they already trained the team on. That’s the real story of AI in tax season, and it’s the one the headlines miss.