ChatGPT for accountants and bookkeepers in New Zealand

Where the hours go in a small practice

Most of the effort in a bookkeeping or accounting practice is not the calculation. It is the checking, chasing and explaining around it. Someone matches a bank export to coded transactions. Someone writes the note that explains why wages jumped in March. Someone drafts the email asking a client, for the third time, for the invoice behind a large purchase. Those are language and pattern tasks, and that is where ChatGPT earns its place.

The calendar sets the pace. A GST return is due by the 28th of the month after the taxable period ends, with two exceptions: the period ending 31 March is due by 7 May, and the period ending 30 November is due by 15 January¹. Inland Revenue also says a GST return cannot be extended the way an income tax return can¹. Employers who file electronically must send employment information within 2 working days of each payday⁵. A tool that helps the team prepare faster is welcome. A tool that adds an error under that pressure is not, so the routine below puts the human check first.

Choose the right account before anyone uploads a ledger

Client files contain personal information. A sole trader's drawings, an employee's pay, a trust beneficiary's distribution: each is information about an identifiable person, and the Privacy Act 2020 applies to it. The Office of the Privacy Commissioner expects an organisation to approve a generative AI tool at senior level, to run a privacy impact assessment, to have a person review outputs before the business acts on them, and to make sure personal or confidential information is not retained or disclosed by the tool⁸.

That rules out staff using personal Free or Plus accounts on client work. On ChatGPT Business and Enterprise, OpenAI states that it does not train its models on business data by default, and that the customer owns its inputs and outputs where the law allows⁹. Business also gives the practice an admin console, so a partner can see who has a seat and switch features on or off. The National Cyber Security Centre warns that generative tools make phishing emails more convincing and easier to tailor to a target¹⁴, so set up the workspace with single sign-on or strong multi-factor authentication from day one.

• Buy ChatGPT Business seats for everyone who will touch client data. Do not reimburse personal subscriptions.
• Write a one-page rule: which client data may go in, which may not (for example, IRD numbers and bank account numbers), and who checks the output.
• Name the privacy officer who approves new uses, as MBIE's voluntary guidance for businesses recommends treating privacy and security as part of normal governance¹³.
• Use Temporary Chat for a one-off question that should not sit in history. OpenAI keeps a copy of a temporary chat for up to 30 days for safety reasons and does not use it to improve models¹⁵.

GST and coding checks with data analysis

ChatGPT's data analysis feature reads .xlsx and .csv files, writes and runs Python against them, and returns tables and charts¹⁰. OpenAI's help centre advises structured files with clear column names and one record per row¹⁰. A transaction export from your ledger software already looks like that, which makes it the best first use for a practice.

Treat the output as a list of questions, not a list of answers. ChatGPT does not know the client's GST registration details, the nature of each supplier, or the private-use apportionment you agreed last year. It can still find patterns that a tired person misses on the 27th.

• Export the GST period's transactions with date, payee, account code, GST code, net, GST and gross columns. Remove IRD numbers and bank account numbers before upload.
• Ask: list every line on a standard-rated code where GST is not 15% of the net amount⁴. Allow a few cents for rounding.
• Ask for every payee coded to more than one account in the period, and every zero-rated or exempt line above a value you set.
• Ask for purchases over the threshold where full taxable supply information is needed, and check that the file holds the supplier's GST number, date, description and amounts that Inland Revenue lists³.
• Open each flagged line in the ledger yourself. Fix the code in the ledger, not in the chat.

Month-end packs and variance commentary

Variance commentary is where a junior spends an afternoon and a manager rewrites half of it. A ChatGPT Project helps here. A Project keeps chats, files and instructions together, and ChatGPT uses that shared context across every chat inside it¹¹. Shared projects use project-only memory, so they cannot draw on a member's memories from outside the project¹¹. Set up one Project per client or per client group, never one for the whole practice.

A worked example: a generic six-person Hamilton bookkeeping practice prepares monthly management accounts for a dairy-support contractor. The Project holds the client's chart of accounts, last year's commentary and a short instruction file. Each month the bookkeeper uploads the profit and loss with budget and prior-year columns and asks for draft commentary on lines that moved by more than the client's agreed tolerance. ChatGPT drafts plain-English notes. The bookkeeper replaces every cause it guessed with the cause from the source papers, then the manager reviews.

• Project instructions: write for a farm-services owner, plain English, no jargon, one short paragraph per line item, flag uncertainty instead of guessing a reason.
• Upload the trial balance or P and L as a spreadsheet, not a PDF, so data analysis can calculate the movements itself.
• Ask it to show the arithmetic for each movement. Check two at random against the ledger.
• Keep the final commentary in your practice software or document system. The chat is a draft, not the working paper.

Working inside the spreadsheet

OpenAI now ships ChatGPT for Excel and Google Sheets, an add-in you install from Excel's Add-ins menu and open from the ribbon¹². It can build, update, explain or summarise a workbook, and OpenAI tells users to review formulas and changed cells before relying on the result¹². Spreadsheet chats run separately from main chat history and do not use ChatGPT memory, and VBA and macros may not be fully supported¹².

For a practice, the add-in suits three jobs. It can explain an inherited workbook that nobody understands, write a lookup formula that reconciles two lists, and build a clean summary tab from a messy export. Work on a copy of the file. Lock the original working paper. When it changes a formula, read the formula, not just the result.

Client letters and the emails nobody wants to write

Writing is a natural fit. ChatGPT can turn bullet notes into a clear letter, adjust tone for a nervous client, and shorten a five-paragraph explanation of provisional tax into three. Good letters still come from good inputs, so give it the facts and the decision, then ask for the words.

Turn the practice's standard letters into a skill. In ChatGPT, a skill is a reusable workflow that can hold instructions, examples, supporting files and code, and a workspace can share it with a group or with everyone¹⁶. Put the practice's style guide, approved engagement-letter wording and a sample of good client emails into it. Leave out any client file. If the practice built custom GPTs in earlier years, plan to move them now: OpenAI retires custom GPTs on 11 December 2026 and points creators to plugins instead^17,19.

• Records request: list the missing documents, say why each one matters, and give a date. Inland Revenue requires business records, including GST records, to be kept for 7 years², so the letter can explain that the client must keep the originals.
• Year-end summary: give the skill the signed figures and ask for a one-page explanation for a non-accountant owner.
• Debt chase: ask for a firm, polite second reminder that keeps the relationship intact.
• Scam warning: Inland Revenue says it will never put a refund amount in an email or text, and asks you to log in to myIR directly rather than through a link⁶. Draft a short client note with those points and the [email protected] reporting address⁶.

What stays with the practitioner

ChatGPT can produce figures that are not in the file, and it can state them with confidence. The Privacy Commissioner's expectation of human review before action⁸ fits accounting work exactly: action means a return filed, a payment made, or advice sent to a client. Keep these decisions with a qualified person, every time.

If client data goes somewhere it should not, treat it as a possible privacy breach. Under the Privacy Act 2020, a breach that has caused or is likely to cause serious harm must be notified to the Commissioner and to affected people as soon as practicable⁷. The Office asks for notice ideally within 72 hours of the business becoming aware of a notifiable breach¹⁸.

• The GST position, the tax treatment of a transaction and any figure on a return.
• Advice to a client about structure, provisional tax or a dispute with Inland Revenue.
• Any change to supplier or client bank details. Confirm by phone on a known number.
• The decision on whether a breach is notifiable.

A two-week start plan for a practice

Start small and with real work. Pick one GST client and one month-end client, run the steps above in parallel with your normal process, and compare. Do not cut the old check until the new one has caught something the old one would have missed, or until two cycles show they agree.

• Days 1 to 2: buy Business seats, set up sign-in, write the one-page data rule, name the privacy officer.
• Days 3 to 5: build one Project per pilot client and one letters skill. Test the skill on last year's letters before sharing it.
• Week 2: run a GST anomaly check and a set of variance notes in parallel with the existing process. Record every error ChatGPT made and every one it caught.
• End of week 2: partners review the log, decide what becomes standard, and update the data rule.