28 AI prompts for accountants, from tax change letters to GST

Search “AI prompts for accountants” and you’ll find the same list rewritten fifty times: “Act as an expert accountant and explain depreciation.” Prompts nobody at a real firm has ever needed, written by people who have never closed a month-end or chased a client for a bank statement in their lives.

The problem with those lists isn’t that the prompts are wrong. It’s that they’re empty. They give the model no context about your firm, no format for the output, and no constraints on what it should and shouldn’t do, so you get back the same generic slurry anyone else would get. And they never mention the step that actually matters in an accounting firm: a human reviews everything before it goes anywhere near a client.

These 28 are organised by the actual jobs in a firm, written the way working prompts are written, with the technique behind each group explained so you can adapt them rather than copy them. They work in ChatGPT, Claude, Copilot or whatever your firm has standardised on: the pattern matters more than the tool. If your firm is still deciding on the tool, start with our ChatGPT for accountants guide or the Claude setup guide for firms.

One rule before any of them: never paste identifiable client data into a consumer AI tool. Use placeholders ([CLIENT], [PERIOD], [AMOUNT]) and swap the real details back in yourself, or use a business-tier tool where your firm controls how data is handled. More on that in the FAQ.

This is the prompt people search for most, so it goes first. When a rule changes, every firm needs the same thing by Friday: a plain-English letter that tells affected clients what changed, whether it touches them, and what happens next. The model is good at the plain English. It is bad at knowing the change, because it may have been trained before the change existed. So you give it the source, and you forbid it from adding anything else.

You are drafting a client letter for an accounting firm. Context: [DESCRIBE THE CHANGE IN ONE LINE], effective [DATE]. Below is the official summary from [SOURCE, e.g. the tax authority’s announcement or guidance page]: [PASTE THE OFFICIAL TEXT]. The letter goes to [CLIENT SEGMENT, e.g. sole traders, small companies with employees], who are not accountants and will read it on a phone. Task: draft the letter. Format: a subject line under ten words, then five short sections: what changed, who it affects, what it means for you, what you need to do and by when, what we will do for you. Under 350 words. Constraints: use only facts from the text I pasted. Do not add rates, thresholds or dates that are not in it. Where the source is unclear or you would need to assume something, write [CHECK] in the text instead of guessing. Do not tell any client what their own tax position is. End by inviting them to book a call if they are unsure whether it applies to them.

How to check the output before it goes anywhere

• Check every fact against the primary source. Read the draft next to the official guidance from the tax authority (the ATO, IRS, HMRC, CRA or IRD, depending on where your clients are). Every date, rate and threshold in the letter should appear in that source, word for word.
• Clear every [CHECK]. Those markers are the model telling you where it ran out of facts. A letter with one still in it does not leave the building.
• Keep client identifiers out of the prompt. Draft one letter per segment with placeholders, then merge names and details in your practice management or email system. No client names, tax file numbers, Social Security numbers or account details go into a consumer AI tool.
• Never send unreviewed. A person who understands the change reads the final version and puts their name on it. The letter is general information; anything about a specific client’s position is advice, and advice comes from you, not the tool.

Segmenting matters more than wording. One letter per client type (“you employ staff”, “you are a sole trader”) gets read. One letter for everyone with “this may or may not apply to you” gets archived.

Every prompt that consistently works has the same four parts. This isn’t a secret (it’s the documented guidance from every major AI lab), but almost no published prompt list applies it:

• Context: who you are, who the client is, what stage the work is at. The model knows nothing about your firm unless you tell it.
• Task: one specific job, not “help me with month-end.”
• Format: exactly what the output should look like: a table, a bulleted email, three options, 150 words.
• Constraints: what to exclude, what tone to use, what it must not invent.

Here is the difference in practice. The LinkedIn version: “Write an email chasing a client for documents.” The working version:

You are drafting an email for an accounting firm. Context: [CLIENT] is a long-standing small business client, generally responsive but two weeks late sending bank statements for [PERIOD]. This is the second follow-up. Task: draft a chase email. Format: under 120 words, friendly but direct, one clear ask with a specific deadline. Constraints: no guilt-tripping, no exclamation marks, do not threaten consequences: we want the documents, not a fight.

Same job. Completely different output. Every prompt below follows this shape. Steal the shape, not just the words.

Here are the checks I run when reviewing a draft P&L before it goes to the partner: [PASTE YOUR CHECKLIST]. Turn this into a structured review checklist with a pass/fail column and a notes column, grouped by section of the P&L, ordered by how often each check catches something.

I am writing a workpaper note explaining a variance. Context: [ACCOUNT] moved from [AMOUNT] in [PRIOR PERIOD] to [AMOUNT] in [PERIOD]. The driver was [REASON]. Draft a two-to-three sentence workpaper narrative in the neutral, factual style of an audit file: no speculation, state only what I have told you.

Here is a list of reconciling items from a bank rec, with descriptions and amounts: [PASTE LIST]. Group them into likely categories (timing differences, duplicates, potential errors, needs investigation), and for each “needs investigation” item, list the most likely explanations to check first. Do not guess amounts or invent items.

Draft a month-end close checklist for a small business client on [ACCOUNTING SOFTWARE] with payroll, GST/VAT/sales tax obligations and one loan account. Format: a table with task, owner (bookkeeper/accountant/client), and dependency. Ask me three questions about the client first before producing it.

I need to document a process so a new team member can run it. I will describe the steps conversationally; turn my description into numbered procedure steps with the software name, screen and action for each step, plus a “common mistakes” note at the end. Here is the process: [DESCRIBE IT].

Why this group works: every prompt feeds the model your raw material (your checklist, your rec items, your process) and asks it to structure, not to know. The model is doing formatting and pattern-grouping, which it’s genuinely good at, instead of inventing accounting facts, which it isn’t.

Here are two emails I have written to clients that sound like me: [PASTE TWO REAL EMAILS, DETAILS REDACTED]. Match this tone exactly. Now draft an email to [CLIENT] explaining that their [PERIOD] accounts are ready to review, with two things to flag: [POINT 1] and [POINT 2]. Under 150 words.

Rewrite this technical explanation for a client who has no accounting background: [PASTE YOUR DRAFT]. Keep it accurate but use plain language, one analogy maximum, and no jargon without a one-line explanation. The client is busy: get to the point in the first sentence.

Draft a fee increase letter. Context: [CLIENT] has been on the same fee for three years, scope has grown ([DESCRIBE WHAT GREW]), and the increase is [X]%. Tone: confident and matter-of-fact, not apologetic. Lead with the value delivered, state the new fee plainly, give an effective date. No “unfortunately”, no “due to rising costs” boilerplate.

A client has asked a question I need to push back on diplomatically: [DESCRIBE THE SITUATION]. Draft three versions of the reply (one soft, one direct, one in between) so I can pick the register. Each under 100 words.

Turn these bullet-point notes from a client meeting into a follow-up email confirming what was discussed and who owes what by when: [PASTE NOTES]. Format: short intro line, “What we agreed” bullets, “Your actions” bullets, “Our actions” bullets, one-line close. Do not add anything that is not in the notes.

Why this group works: the first prompt is the whole trick: pasting real examples of your own writing beats any tone description you could type. “Professional but friendly” means nothing to a model; two of your actual emails mean everything. And if that last meeting-to-follow-up prompt becomes a daily habit, a purpose-built tool like Vinyl drafts the email from the meeting itself.

Here are the last four quarters of summarised P&L data for a client (details anonymised): [PASTE FIGURES]. Identify the five trends most worth discussing in an advisory meeting, ordered by financial impact. For each: what the numbers show, one plausible question to ask the client about it. Flag anything where you would need more data before commenting. Do not fill gaps with assumptions.

I am preparing for an advisory meeting with a [INDUSTRY] client doing roughly [REVENUE BAND] in revenue. Generate a list of the ten benchmarks and ratios most worth calculating for this industry, why each matters, and what a concerning number would look like. I will calculate them myself from the actual accounts.

Play the sceptical client. I am going to present a recommendation: [DESCRIBE IT]. Push back the way a cost-conscious business owner would: ask me the hard questions, challenge my assumptions, and do not let me off easily. After the exchange, summarise which of my answers were weakest.

Turn this analysis into a one-page client summary: [PASTE YOUR ANALYSIS]. Format: three headline findings in plain language, one short table of the key numbers, three recommended next steps. Write for a business owner reading it on a phone between meetings. Nothing in the summary that is not in my analysis.

Why this group works: the model never touches a judgement call. It structures your analysis, rehearses your meeting and suggests what to look at. You compute the numbers and own the advice. The sceptical-client prompt is the sleeper hit: role-play is one of the few things these tools do better than most colleagues have time to.

Here is our current job template for [SERVICE TYPE], exported as a task list: [PASTE IT]. Identify tasks that are vague (“review file”), duplicated, or missing an obvious owner, and propose a tightened version. Keep our task-naming style. Ask me about anything ambiguous before rewriting.

Draft an onboarding email sequence for a new business client: three emails. Welcome and what happens next, documents we need (I will insert the list), and how to reach us and what to expect at month one. Match the tone of this example: [PASTE ONE REAL EMAIL]. Each email under 150 words.

I am writing a job ad for a [ROLE] at a [SIZE]-person firm. Here is what the role actually involves day to day: [DESCRIBE HONESTLY]. Draft an ad that leads with the real work and the real conditions: no “rockstar”, no “fast-paced environment”, no listing ten years’ experience for a five-year role. Include three screening questions that would reveal whether someone has genuinely done this work.

Here are anonymised notes on how a client engagement went wrong: [DESCRIBE]. Write a short internal post-mortem: what happened, contributing factors, and three process changes that would catch this earlier next time. Blameless tone: systems, not people.

Why this group works: ops prompts fail when they ask the model to imagine your firm. These all hand over the real artefact (your template, your emails, your engagement notes) and ask for critique or restructuring. If you find yourself running the same ops prompts weekly, that is the point where you graduate from prompts to agent templates that run the whole job.

A caveat that applies to every prompt in this section: tax rules are jurisdiction-specific and change constantly. These prompts use AI to organise and draft around tax work, never to determine a tax position. Check local rules, always.

Here is a list of open queries on a tax return I am preparing: [PASTE LIST]. Sort them into: client must answer, I can resolve from documents I already have, and needs technical research. Draft one consolidated email to the client covering only their items, grouped so they can answer everything in one sitting.

I am researching a tax question: [DESCRIBE THE ISSUE, NO CLIENT DETAILS]. Do not give me an answer. Instead, list the concepts, terms and types of official guidance I should be searching for in my jurisdiction, and the questions a reviewer would expect my file note to address. I will do the research against primary sources.

Draft a client-facing explanation of why we need [DOCUMENT TYPE] for their return, for a client who is annoyed about being asked. Two short paragraphs: why it is required, what happens if we lodge without support. Firm but courteous. Jurisdiction-neutral: I will add the specific rule reference myself.

Here is my raw file note from a call about a client’s tax situation: [PASTE NOTE, ANONYMISED]. Restructure it into a proper file note: facts as stated, issue, what was discussed, what was NOT concluded, and open items. Flag anywhere my note is ambiguous about who said what.

Prompt: explain GST obligations to a client

Same principle as the tax change letter: the model writes the explanation, you supply every number. GST rules differ between Australia, New Zealand, Canada and everywhere else that uses the name, so the prompt keeps the figures as placeholders you fill from the tax authority’s own pages.

Explain GST obligations to [CLIENT TYPE, e.g. a newly registered cafe owner] in [JURISDICTION]. Context: they have [JUST REGISTERED / ARE APPROACHING THE REGISTRATION THRESHOLD / ARE LATE WITH A RETURN], and their books are in [ACCOUNTING SOFTWARE]. Task: a plain-English explanation of what they now have to do. Format: three short sections: what GST means for their prices and invoices, what records to keep, and how reporting and payment work. Under 250 words, no jargon without a one-line definition. Constraints: do not state any rate, threshold or due date. Write [RATE], [THRESHOLD] and [DUE DATES] and I will fill them in from official guidance. Finish with the three things they should do this week.

Prompt: draft tax-time FAQs for clients

The best source for tax-time FAQs is the questions your team actually answered last season, not the questions a model imagines. Feed it those.

Here are the questions our clients asked most during last tax season, collected from our inbox and phone notes, with client details removed: [PASTE LIST]. Group duplicates, pick the ten that came up most, and draft a client FAQ for our newsletter or website. Format: question in the client’s own words, answer in under 60 words, plain language. Constraints: where an answer depends on a date, rate or rule, write [CHECK: what I need to confirm] instead of answering. Add one final question: “When should I just call you?”, with an answer that covers the situations in the list where a general FAQ is not enough.

Why this group works: the second prompt is the model of safe tax use: “do not give me an answer” turns the tool into a research planner instead of an unreliable oracle. Everything else here is triage, drafting and structuring: the parts of tax season that eat hours without requiring judgement.

I have 25 minutes. Teach me [TOPIC] as if I am an experienced accountant who has never dealt with it: what it is, when it shows up in practice, the three mistakes practitioners make with it, and what I should read from primary sources to verify what you have told me. Be upfront about anything you are uncertain on.

Quiz me on [TOPIC]. Ask one question at a time, wait for my answer, tell me if I am right and why, then ask a harder one. Practical scenarios, not definitions. Stop after eight questions and tell me my weak spots.

I just read this article/standard update: [PASTE OR SUMMARISE]. Ask me five questions to test whether I actually understood it, then summarise what it changes for a firm like mine: [DESCRIBE FIRM]. Distinguish clearly between what the text says and what you are inferring.

Why this group works: the model switches from answer machine to sparring partner. Quizzing and teach-back are better learning than passive summaries, and the “tell me what to verify” instruction builds the review habit into the learning itself.

1. Paste real examples. The single biggest upgrade available to you costs nothing: show the model an actual email you sent, an actual workpaper note, an actual file note. One real example outperforms a paragraph of tone instructions every time. Redact the client details; keep the voice.

2. Make it ask questions first. Add “ask me three questions before you answer” to any prompt where the output disappoints. Nine times out of ten, generic output means missing context, and the model’s questions tell you exactly what context you forgot to provide. It’s the cheapest debugging step there is.

3. Never send unreviewed. Nothing generated goes to a client, a file or a regulator without a human reading every word. Not because the tools are bad, but because your name is on the work and theirs is not. The firms getting real value from AI treat it like a sharp junior: fast, tireless, occasionally confidently wrong, and never allowed to sign anything.

When a prompt earns a place in weekly use, save it somewhere shared with a named owner, the way you would a letter template. A dozen prompts your team trusts beats two hundred nobody opens, and it hands a new hire your firm’s trial and error on day one. Then write the boundary into your one-page AI policy so nobody has to guess what can be pasted where.

Prompts are the entry point, not the destination. Once your team is comfortable here, the bigger wins come from tools that hold your firm’s context permanently instead of being re-briefed in every chat. That’s the territory of our AI for accounting firms hub.

---

What’s a good AI prompt for drafting a client letter about a tax change?

One that hands the model the official source and bans it from adding anything else, like the one at the top of this piece. The failure to avoid is asking “write a letter about the new [RULE]” and trusting what comes back: the model may not know the change, and it will fill the gap with something plausible. Paste the authority’s own summary, name the client segment, ask for [CHECK] markers, and write one version per segment rather than one for everybody.

What’s a good AI prompt for explaining GST obligations?

One that describes the client’s situation (just registered, near the threshold, behind on returns) and leaves every figure as a placeholder. The situation is what makes the explanation useful; a generic “explain GST” produces a textbook page the client will not read. Keeping rates and dates out of the model’s hands also means the explainer stays reusable after the numbers change: you update the placeholders, not the prose.

Do these prompts work the same in ChatGPT, Claude and Copilot?

The pattern (context, task, format, constraints) works identically everywhere; it’s how all current models are trained to follow instructions. You will notice stylistic differences: ChatGPT tends to be strong on tightly formatted and structured output, Claude on longer drafting and tone-matching, Copilot on anything living inside Microsoft 365. But a well-built prompt transfers between tools with minor tweaks, which is exactly why learning the pattern beats memorising a library.

How do I use these prompts with client data safely?

Default rule: no identifiable client data in consumer AI tools, full stop. Use placeholders like [CLIENT] and [AMOUNT] and reinsert real details in your own systems, or anonymise before pasting. If your firm wants to work with real data, use a business or enterprise tier where inputs are not used for model training and your firm controls retention, and write that boundary into your AI policy so nobody is guessing. Your professional body and local privacy law may impose stricter requirements; check both.

What about custom GPTs, Projects and skills: do they replace prompts?

They are the natural next step. Custom GPTs, Claude Projects and skills let you save the context and constraints once (your firm’s tone, formats, standard caveats) so every chat starts pre-briefed instead of you re-pasting the same setup. If you find your team reusing the same three prompts daily, that’s the signal to promote them into a persistent setup; our Claude setup guide walks through doing exactly that for a firm. The prompt-writing skill doesn’t become obsolete. It becomes the thing you write once, properly, instead of fifty times badly.