5 ChatGPT Prompts Every Accountant Should Be Using This Year

5/7/2026

Most articles offering accountants AI prompts are written by people who have never had a client, and they suggest asking the tool to explain depreciation. Here is a more useful framing.

A language model is a fast, articulate, endlessly patient assistant with no access to authority and no accountability. That combination makes it excellent at some things practitioners spend real time on, and dangerous at exactly the things practitioners are paid for.

Before any of the prompts, though, there is a rule that matters more than all of them.

Do Not Paste Client Data Into a Consumer AI Tool

This is the part most coverage of this topic skips, and it is the part that can cost a license.

Professional confidentiality obligations apply to client information regardless of the technology involved. Entering identifiable client information into a third-party service is a disclosure, and it needs to be authorized.

For tax practitioners the rule is stricter and specific. The statutory restriction on the disclosure and use of tax return information requires client consent in a prescribed form and manner before that information may be disclosed to or used by another party, with penalties for violation. A practitioner who pastes a client's return data into a public chatbot to help draft an explanation has, on a plain reading, made a disclosure without the required consent. This is not a theoretical concern and it is not addressed by good intentions.

The rules governing practice before the IRS impose their own diligence and confidentiality expectations on practitioners.

The practical controls that let a firm use these tools responsibly:

Use an enterprise or business tier with contractual data protections, including terms addressing whether inputs are used for training and where data is processed. Consumer tiers frequently do not offer this, and the distinction is the whole issue.

De-identify before you paste. Replace names, identifying numbers, and specific figures with placeholders. Most of the useful applications below work perfectly well on a de-identified fact pattern, because the tool is helping with structure and language rather than with the client's data.

Prohibit consumer tools for client information in a written firm policy, and say what is permitted instead. Staff will use these tools; the only question is whether they do it inside a policy or around one.

Address consent where client information genuinely needs to be processed, using the required form for tax return information.

Everything below assumes de-identified input or an appropriately contracted environment.

1. The Client Translator

What it is for: converting a technically correct conclusion into something a client will actually read and understand — the task practitioners most consistently underestimate and most consistently do badly.

The pattern:

"Rewrite the following explanation for a small business owner with no accounting background. Keep it under 200 words. Preserve these three caveats exactly: [list them]. Do not add any new advice, conclusions, or numbers. Use plain language, short sentences, and no jargon. Explain what it means for them practically."

Why it works: the instruction constrains length, audience, and — critically — forbids adding content. The most common failure when using a model for client communication is that it helpfully invents a recommendation you did not make.

How to check it: read it once for anything you did not say. That is the only real risk, and it is quick to spot.

2. The Structured First Draft

What it is for: engagement letters, memos, procedure documentation, and client communications where you have the substance and the blank page is the obstacle.

The pattern:

"I am writing an internal memo documenting a conclusion. Here are my bullet points: [substance]. Produce a structured memo with sections for background, the question, the analysis, the conclusion, and open items. Use only the facts and reasoning I provided. Mark anything that appears to be missing with [GAP] rather than filling it in."

Why it works: you supply the substance and the judgment; the tool supplies organization and prose. The [GAP] instruction is the valuable part — it turns the model into something that surfaces holes in your outline rather than papering over them.

How to check it: confirm no reasoning appeared that you did not provide, and address every gap marker.

3. The Adversarial Reviewer

The best professional use of these tools, and the least used.

What it is for: stress-testing your own conclusion before a client, a reviewer, or an examiner does it for you.

The pattern:

"Here is my conclusion and my reasoning: [conclusion and reasoning]. Argue against it as forcefully as you can. Identify the weakest link in the reasoning, the facts that would change the answer, the alternative treatment someone might assert, and what additional evidence would make my position stronger. Do not agree with me."

Why it works: the model has no stake in your conclusion and no reluctance to disagree, which is exactly what a busy colleague lacks. It reliably surfaces the objection you had not considered — and because you are asking for arguments rather than authority, its weakness at citation does not matter.

How to check it: most of what it produces will be weak, and one item usually will not be. That one is the point. Verify any authority it references independently, and treat the arguments as prompts for your own analysis rather than as findings.

4. The Procedure Builder

What it is for: generating a first-pass step list for an engagement, a review, or an internal process, which you then edit down to what actually applies.

The pattern:

"I need to design procedures for [objective], for an entity with these characteristics: [de-identified characteristics]. Produce a numbered list of procedures. For each, state the objective, what evidence it would produce, and what could go wrong with it. Flag any step that depends on a specific standard or rule so I can verify the requirement myself."

Why it works: the model produces breadth quickly. Practitioners are good at judging which steps matter and bad at remembering all the candidates, so this plays to the complementary strengths.

How to check it: treat the list as a brainstorm, not a program. Delete aggressively, and verify every flagged item against the actual standard — because the model does not know which standard applies to your engagement and will not tell you it is guessing.

5. The Spreadsheet Assistant

What it is for: three specific tasks it does genuinely well — constructing a formula from a described intent, explaining a formula someone else wrote, and describing how to restructure data.

The pattern:

"In Excel, I have [describe columns and structure]. I need to [describe the result]. Give me the formula, explain what each part does, and tell me what it will do with blanks, text values, and errors in the source data."

Why it works: formula syntax is exactly the kind of pattern-heavy knowledge these models hold reliably, and the edge-case question is what makes the output safe to use.

How to check it:test it on data where you already know the answer. Always. A formula that looks right and mishandles blanks will produce a wrong number silently, which is the worst failure mode in spreadsheet work. Pair this with real spreadsheet skill from Essential Excel Skills or the Excel training for accountants catalog — the tool is a supplement to competence, not a substitute.

A Sixth Worth Adding: The Question Generator

Underrated, and it costs nothing:

"I am meeting a client in [industry] about [de-identified topic]. Generate 20 questions I should ask, including several that a less experienced advisor would forget. Group them by theme."

Most of the list will be obvious. Two or three will be things you would have thought of after the meeting, which is when thinking of them is useless.

What These Tools Are Bad At

Worth being specific, because the failures are predictable and expensive.

Authority and citations. Language models invent statutory sections, standard references, case names, and revenue procedure numbers, and they do it fluently and confidently. Never cite anything you have not read in the source. This is the single most important limitation.

Current rules. Anything that changed recently may be absent or wrong, and the model will not flag its own staleness. Every tax season this catches practitioners.

Arithmetic on real data. Do not use it to compute a client's numbers.

Judgment. Materiality, reasonableness, whether an estimate is supportable, whether a position has substantial authority — these are professional judgments and the model has no basis for them beyond pattern imitation.

Knowing what it does not know. It does not say "I am unsure." That absence is the core risk, and it is why the verification discipline matters more than the prompt.

The useful mental model: a bright intern who has read a great deal, has no access to the authoritative literature, cannot tell you when they are guessing, and will never admit uncertainty. You would use that person for drafting, structuring, and argument-testing. You would not file anything based on their citation.

Structured coverage is available through the AI for Accountants Certificate Program, AI Essentials for Accountants, AI Applications for Accountants, the AI for Accountants Strategy and Research Specialist series, the AI courses for accountants and CPAs catalog, and — for the professional obligations — ethics training and professional conduct and tax practitioner regulations, penalties, and security.

Firm Policy Elements

If a firm is going to permit this — and staff are using it whether or not the firm permits it — the policy needs to say:

• Which tools are approved, and that consumer tiers are not approved for client information
• What may never be entered: identifying client information, tax return information absent required consent, credentials, and unreleased financial data
• The de-identification standard expected
• That output must be verified before use, with authority read in the source
• Who owns the work product, which remains the practitioner regardless of what drafted it
• Whether use must be documented in an engagement file, which is worth deciding deliberately rather than by default
• Training, so the policy is understood rather than merely issued

Where Practitioners Get This Wrong

• Pasting client data into a consumer tool, which is a disclosure and, for tax return information, a specific violation
• Citing authority the model produced without reading it in the source
• Asking it for a conclusion rather than for drafting, structure, or counterarguments
• Using it for current-year rules without verification
• Letting it add recommendations to a client communication
• Trusting a formula without testing it on known data
• Treating breadth as accuracy — a long, confident answer is not a checked one
• No firm policy, so staff improvise

The honest summary: these tools save real time on translation, structure, and self-critique, which together consume more of a practitioner's week than anyone admits. They save no time at all on the things clients pay for — judgment, authority, and accountability — and a practitioner who confuses the two categories will eventually file something that was never true.

Frequently Asked Questions

Can accountants put client information into an AI chatbot?

Not into a consumer tool. Professional confidentiality obligations apply to client information regardless of technology, and for tax practitioners the separate statutory restriction on disclosure and use of tax return information requires client consent in a prescribed form, with penalties for violation. The workable approach is an enterprise tier with contractual data protections, de-identification before input, and a written firm policy.

What is the best professional use of these tools?

Adversarial review — giving the model your conclusion and reasoning and asking it to argue against it as forcefully as possible, identify the weakest link, and state what facts would change the answer. It has no stake in your conclusion and no reluctance to disagree, which is what a busy colleague lacks, and asking for arguments rather than authority avoids its worst weakness.

Why can't AI output be cited as authority?

Because language models invent statutory sections, standard references, case names, and procedure numbers, and they do so fluently and confidently. They also cannot signal their own uncertainty. Nothing should be cited that has not been read in the authoritative source.

How should AI-generated Excel formulas be checked?

By testing them on data where the correct answer is already known, every time, and by asking the model explicitly what the formula does with blanks, text values, and errors. A formula that looks right and mishandles blanks produces a wrong number silently, which is the most dangerous failure in spreadsheet work.

What should a firm's AI policy cover?

Which tools are approved and that consumer tiers are not approved for client information; what may never be entered, including tax return information absent required consent; the de-identification standard; a requirement that output be verified against source authority; that the practitioner owns the work product regardless of what drafted it; whether use must be documented in engagement files; and training.

What is a realistic way to think about these tools?

As a bright intern who has read widely, has no access to the authoritative literature, cannot tell you when they are guessing, and will never admit uncertainty. Useful for drafting, structuring, and testing arguments; never a source for a citation or a professional judgment.