ChatGPT prompts for bookkeepers: 15 for work the ledger can’t do

Bookkeepers who use ChatGPT, Claude or Copilot well use it for the work around the ledger, not the ledger itself. Bank feeds, bank rules and the reconciliation suggestions inside Xero and QuickBooks Online already do the matching. What still eats the week is the writing and the thinking that sits next to it: the email asking a client what 14 transactions were, the hunt for why a reconciliation is out, the export that will not import, the process nobody has written down.

That is where a chat assistant pays for itself. Below are 15 prompts, grouped by the jobs in a normal bookkeeping week, with what to remove from the data before you paste anything in. They work in any of the major assistants. For the wider picture of where AI fits in a bookkeeping practice, including the tools that work inside the ledger, see the bookkeepers section of our AI for accountants guide.

Three rules, in this order.

1. Use a tier your firm has approved. Free consumer accounts and business tiers handle your data differently. Our piece on whether AI is safe for client data sets out which tiers train on what you type and which do not.
2. Strip identifiers. Replace client names with “Client A”, delete tax file numbers, social security numbers, bank account numbers and addresses. Most prompts below work just as well on anonymised data.
3. Know which rules you are under. In Australia, the Tax Practitioners Board’s guidance statement TPB(GS) 55/2026, issued on 22 July 2026, applies to registered BAS agents as well as tax agents. It says permission is needed from each client before client information is entered into an AI tool, depending on how the tool is configured and used, and that AI output should be reviewed and the review documented. In the UK, AAT publishes topical guidance on AI tools alongside its Code of Professional Ethics, where confidentiality is a fundamental principle. In the US, the IRS reminds tax and accounting professionals that they are treated as financial institutions under the Gramm-Leach-Bliley Act and must keep a written information security plan, which should cover the AI tools you use.

This is a summary of published guidance, not legal advice. Check your own professional body’s rules for your situation.

Each prompt below has the same five parts: who the assistant should act as, the context it needs, the task, the format you want back, and an instruction to flag anything it is unsure of. The last part matters most. An assistant that is told to list its assumptions gives you something to check. One that is not will fill gaps with confident guesses.

1. The unknown transactions email.

You are a bookkeeper writing to a small business client. Below is a list of bank transactions I cannot code (date, amount, description). Write a short, friendly email asking what each one was. Group similar items, number the questions, and ask for a receipt where one would normally exist. Keep it under 200 words plus the list.

This is the most disliked job in bookkeeping, and it is where an assistant saves the most time per use.

2. The second chase.

The client has not replied to the email below after seven days. Write a follow-up that is shorter and firmer, explains that month-end is on hold until we hear back, and offers a 10-minute call instead of a written reply.

3. Why we need the receipt.

Explain to a sole trader, in plain English and under 150 words, why we ask for receipts for business expenses and what happens at tax time if they are missing. Do not quote specific thresholds or rules; I will add those.

4. The profit versus bank balance question.

A client asks why the business made a profit last quarter but has less cash in the bank. Using the summary figures below, write a short explanation with the three biggest reasons, in the order of their size. Flag any figure you are inferring rather than reading.

5. Finding a reconciliation difference.

My bank reconciliation is out by the amount below. Here are the unreconciled bank lines and the unreconciled ledger lines. List the most likely explanations: transposed digits, duplicates, split payments, timing differences. For each, show which lines you think are involved and why.

Ask for candidates, not an answer. You still tick the lines off yourself.

6. Duplicates in an export.

Below is a CSV of supplier bills. Identify likely duplicates: same supplier and amount within 10 days, or same invoice number. Return a table of suspected pairs with the reason for each. Do not remove anything.

7. Suggesting bank rules.

Here are six months of coded transactions (payee, description, account). Suggest bank rules for payees that are always coded to the same account. Leave out any payee coded to more than one account, and list those separately so I can review them.

The split list is the useful part: it shows where a rule would miscode.

8. The export that will not import.

Reformat the data below to match these import columns exactly: Date (DD/MM/YYYY), Amount, Payee, Description, Reference. Negative amounts are payments. Return CSV only, and list any rows you could not map.

For bigger files, ask for a spreadsheet formula instead of pasting thousands of rows. Our guide to using AI with Excel covers why.

9. Variance notes.

Compare this month’s profit and loss with the same month last year (both below). List every line that moved by more than 15 percent, with a one-line possible reason for each, marked as a question for me to confirm, not a conclusion.

10. A client-specific month-end checklist.

Build a month-end checklist for a client with these features: [payroll, inventory, loan, two bank accounts, sales through a payment platform]. Order it so the tasks that depend on others come later. Keep each item to one line.

11. The owner summary.

Write a five-sentence summary of this month’s results for a business owner who does not read financial reports. Use the figures below only. Mention one thing to watch next month.

12. The pre-lodgment review list.

I am reviewing a GST return (or VAT return, or sales tax filing) before it is lodged. Build a checklist of the review steps a careful bookkeeper would work through, based on the business type below. Do not state rates, thresholds or treatments; list the questions I should answer.

The assistant organises the review. The treatment decisions stay with you, which is exactly where TPB(GS) 55/2026 puts them for BAS agents.

13. Plain-English rule change.

Below is the text of an official update from the tax authority. Summarise what changes, who it affects and from what date, in under 120 words, quoting the source for each point. Flag anything the text does not say clearly.

Paste the official text in. Never ask the assistant what the rule is from memory.

14. The SOP nobody wrote.

Turn these rough notes of how I process supplier bills for Client A into a numbered procedure a new team member could follow. Add a “what can go wrong” line to each step, and list any step where my notes are unclear.

15. The scope email.

A client has asked us to “also look after” payroll. Write a reply that thanks them, explains that payroll is outside our current engagement, lists the information we would need to scope it, and proposes a short call.

Keep the pricing conversation for the call. The assistant should not be deciding what the work is worth.

Three jobs to keep out of a general chat assistant.

• Coding from memory. It does not know the client’s chart of accounts or history unless you give it both, and even then the ledger’s own suggestions work from real history.
• Tax treatment decisions. Ask it to gather and organise the facts. Deciding the treatment is your professional judgement, and the TPB guidance is explicit that AI output is no substitute for your own analysis.
• Anything you cannot check. If you could not spot a wrong answer, the prompt is outside your review capacity. Use it to learn, not to produce client work.

If you are building a shared prompt library for a whole team, start from our broader list of AI prompts accountants actually use, and write the data rules into a one-page AI policy so everyone works to the same standard.

Can ChatGPT do bookkeeping?

Not on its own. A chat assistant cannot see your bank feeds or ledger unless you paste data in or connect it, and it has no record of a client’s history. It is useful for the communication, investigation and documentation around bookkeeping. The recording and matching are better done inside the ledger, where the rules and history live. Our explainer on what a bookkeeper does in 2026 covers what the software has absorbed.

Is it safe to paste bank transactions into an AI tool?

It depends on the tier and on what you paste. Anonymise first: remove names, account numbers and tax identifiers. Use an account your firm has approved, where data is not used for training. In Australia, a registered BAS agent also needs the client’s permission before client information goes into an AI tool, which is usually handled in the engagement letter.

Do these prompts work in Claude and Copilot as well as ChatGPT?

Yes. They are plain instructions with no product-specific features. You may need to adjust the output format line, for example asking for a table rather than CSV if the tool displays tables better.

How should I save prompts I use every week?

Keep them in a shared document or your practice management system’s templates, with a note on what data to remove first. Most assistants also let you save instructions in a project or custom assistant, which saves retyping the role and format each time.

Do I need to tell clients I use AI?

Rules differ by country and by professional body. Australian registered agents need permission before disclosing client information to a third party, which can include an AI tool. Elsewhere it is often good practice rather than a strict rule. Our guide to client consent for AI sets out the position market by market.