ChatGPT for Accountants: Setup, Custom Instructions, and Limits

ChatGPT is genuinely useful to an accounting or bookkeeping practice, but only after you set it up to know your firm. Out of the box it writes like a stranger: generic client emails, boilerplate engagement letters, explanations pitched at nobody in particular. This guide covers the part most "ChatGPT for accountants" articles skip — the actual setup, including the custom instructions to paste, the prompts that produce sendable drafts, the limits that matter in a regulated profession, and the situations where Claude is honestly the better tool. If you want the broader picture of which accounting workflows AI helps with at all, start with our guide to AI for accountants; this post assumes you know the "what" and focuses on the "how" for ChatGPT specifically.

TL;DR

ChatGPT for accountants works when you front-load context: use custom instructions (or a dedicated Project) to tell it your services, client types, software stack, and writing voice once, so every draft starts from an assistant that knows your practice. It excels at client emails, plain-English explanations, engagement letters, deadline reminders, and internal SOPs. It cannot be trusted with tax rules, arithmetic, or client PII, so every number and rule gets verified and identifying details stay out of prompts. For heavy document work and file organization, Claude is often the stronger fit, and the free Claude Cowork course teaches that side of the workflow.

Why generic ChatGPT produces generic accounting work

A fresh ChatGPT session knows nothing about your practice. It does not know that you run tax, bookkeeping oversight, and quarterly advisory. It does not know your clients are mostly owner-managed service businesses, that your busy season runs January through April, that you work in QuickBooks Online plus a tax-prep suite, or that you write to clients in plain English and never in jargon. So it guesses, and its guesses are averages of every accounting firm on the internet.

The fix is not better prompting on every single request. The fix is telling it those things once, in a place where it persists. The moment ChatGPT knows your services, your client base, and your voice, it stops writing like a stranger and starts writing like a junior who has been at your practice for two years. That is the difference between drafts you rewrite line by line and drafts you check, tweak one sentence, and send.

This principle applies to any AI tool, not just ChatGPT — we cover the general version in how to make AI remember your business.

Setting up ChatGPT for your firm: the 15-minute version

ChatGPT gives you three places to put firm context. Use them in this order.

1. Custom instructions (the foundation)

In ChatGPT's settings, custom instructions let you tell it who you are and how you want it to respond, and it applies to every new chat. This is where your firm's core facts go. Here is a template to adapt:

About me:
I run a small accounting practice: tax preparation, monthly bookkeeping
oversight, and quarterly advisory. Clients are mostly owner-managed
service businesses with under 20 employees. Busy season is January
through April. We use QuickBooks Online and [tax software]. I am a
[CPA / EA / bookkeeper], and all client-facing work goes out under
the firm's name.

How to respond:
Write client communications in plain English with no jargon. Warm but
professional, brief, and direct. Never state a tax rule, threshold, or
figure as fact without flagging that I need to verify it. When drafting
emails, sign off as the firm. Assume the reader is a business owner,
not an accountant.

Two lines in that block do disproportionate work. "Never state a tax rule as fact without flagging that I need to verify it" makes ChatGPT hedge where it should hedge instead of sounding confidently wrong. And "assume the reader is a business owner, not an accountant" fixes the single most common failure in AI-drafted client emails: explanations pitched at a peer instead of a client.

2. Projects (one per recurring context)

ChatGPT Projects let you group chats and attach files that every chat in the project can see. Create one project per recurring context: "Client communications," "Engagement letters and proposals," "Internal SOPs." Drop in your engagement letter template, your standard scope and fee language, and a short document describing your firm. Now a request like "draft the engagement letter for a new monthly bookkeeping client" starts from your template, not a generic one.

If you want to go further, a single well-structured system prompt describing your whole business is the highest-leverage document you can write. We walk through building one in how to write a system prompt for your business.

3. Memory (let it accumulate, but audit it)

ChatGPT's memory feature picks up facts about you across conversations. It is convenient, but for an accounting practice it deserves a periodic audit: open the memory settings, review what it has stored, and delete anything client-specific. Firm facts belong in memory; client facts do not.

The prompts that produce sendable drafts

Once the setup is in place, these are the requests that reliably come back usable. Notice how short they can be, because the context is already loaded.

Client email explaining a change:

Write a short, friendly email to a client explaining that their Q3
estimated payment is higher than last quarter because revenue is up.
Reassure them this is a good sign, note the September 15 due date,
and keep it under 150 words.

Plain-English explanation of a notice:

A client received an IRS CP2000 notice. Draft an email explaining in
plain terms what this notice generally means, that it is not an audit,
and that we will handle the response. Do not state specific deadlines
or amounts; I will add those.

Engagement letter from your standard scope:

Draft an engagement letter for a new client: monthly bookkeeping,
quarterly review calls, year-end tax prep. Fixed monthly fee. Use our
standard payment and termination terms from the attached template.

Deadline reminder sequence:

Write three short reminder emails asking clients to send tax documents:
one for early February, a firmer one for February 20, and a final
notice for February 28 explaining we cannot guarantee on-time filing
after that date. Warm tone throughout, list of needed documents in
each, portal upload link placeholder included.

Internal SOP from a rough description:

I will describe how we run month-end close. Turn it into a numbered
SOP a new hire could follow, with a checklist at the end. Ask me
clarifying questions about anything ambiguous before writing it.

That last pattern, asking the AI to interview you before writing, is the quiet win for bookkeepers and firm owners whose processes live entirely in their heads.

The limits that actually matter in an accounting practice

This section is not a disclaimer. It is the operating manual.

It fabricates tax rules with total confidence. ChatGPT will state a filing threshold, a deduction limit, or a deadline that is wrong, outdated, or invented, in the same confident tone it uses when it is right. Tax law changes faster than any model's training data. Every rule, figure, and citation gets verified against the actual code, current IRS guidance, or your research platform before it reaches a client. No exceptions, ever.

It does not do reliable arithmetic. It is a language model, not a spreadsheet. Your accounting software does the math. ChatGPT drafts the words around the math.

It has no live data. It cannot see this year's mileage rate, the current contribution limits, or last week's guidance unless you paste them in. Treat any "current" figure it offers as stale until proven otherwise.

Client data is a professional-obligations question, not just a privacy preference. Do not paste Social Security numbers, full account numbers, or identifiable client details into any general-purpose AI tool without understanding where the data goes and whether your engagement terms and professional standards permit it. The practical habit: redact identifiers, use only the business facts the task needs, and keep sensitive specifics out of prompts entirely. "Draft an email to a client whose estimated payment went up 30 percent" needs no name attached.

Your license is the product. ChatGPT cannot sign a return, stand behind an opinion, or absorb liability. Everything it produces is a first draft that your judgment turns into work product. The hours it saves are blank-page hours, not review hours.

Where Claude is the better fit

Covering ChatGPT honestly means saying where it is not the strongest choice. For an accounting practice, three situations favor Claude.

Long-document work. Reviewing a lengthy lease, a stack of statements, or a dense contract against your notes is where Claude's handling of large amounts of text tends to shine. If a meaningful share of your week is reading long documents and extracting what matters, that difference compounds.

File and folder work. The most tedious jobs in a practice are not chat jobs at all: renaming and sorting a season's worth of client uploads, cross-checking received documents against a checklist, assembling workpaper folders. Claude Cowork works directly with files and folders on your machine, which turns "organize the client dropbox" from a copy-paste marathon into a delegated task. Chat-based tools, ChatGPT included, are simply not built for that.

Careful, hedge-where-it-should writing. Firms that draft a lot of judgment-adjacent client communication often find Claude's default writing style needs less taming. That is a tendency rather than a rule, and your mileage will vary, so run the same real task through both before standardizing.

The honest bottom line: many firms end up using both, ChatGPT for quick drafting and Claude for document-heavy and file-heavy work. For a fuller side-by-side, see Claude vs ChatGPT for business.

Frequently asked questions

How do accountants use ChatGPT?

Mostly for the writing and explaining work around the accounting: drafting client emails, translating tax topics into plain English, writing engagement letters and proposals from a standard scope, producing deadline-reminder sequences, and turning informal process knowledge into written SOPs. It is not used for calculations, filings, or as a source of tax rules, all of which stay with professional software and professional judgment.

What should I put in ChatGPT custom instructions as an accountant?

Four things: your services and client types, your software stack, your credential and sign-off conventions, and your writing voice. Add two behavioral rules: never state a tax rule or figure as verified fact, and always write client communications for a business owner rather than a fellow accountant. This one-time setup is what separates sendable drafts from generic ones.

Is ChatGPT safe for client data in an accounting firm?

Not by default. Do not paste Social Security numbers, account numbers, or identifiable client details into a general-purpose AI tool without confirming where the data goes and whether your engagement terms and professional obligations allow it. The safe habit is to redact identifiers and prompt with only the business facts the task needs, which covers almost every drafting use case.

Can ChatGPT answer tax questions accurately?

Not reliably enough to act on. It can explain concepts well, but it fabricates thresholds, deadlines, and rules with full confidence, and its training data lags current law. Use it to draft the explanation of a tax position you have already verified, never as the source of the position itself.

Is ChatGPT or Claude better for accountants?

For quick conversational drafting, both are strong and ChatGPT is a fine choice. Claude tends to be the better fit for long-document review and, through Claude Cowork, for file-level work like sorting client uploads and checking documents against a checklist, which chat tools do not handle. Many firms use both; test each on one real task from your own practice before standardizing.

Set it up once, then make it yours

The whole playbook here is one afternoon of work: write your custom instructions, create a project or two with your templates, and run this week's real client emails through it. The setup is what most firms skip, and it is the entire difference between a novelty and a tool.

If you want the file-and-document side of the workflow, the side ChatGPT does not cover, the free Claude Cowork course walks through it with real prompts, written for firm owners and operators rather than developers. And if you would rather have your firm's context, instructions, and top workflows configured for you in one pass, Get Set Up on Claude does exactly that.