Mac Sweeny
Quick answer
ChatGPT is a drafting, explaining and summarising tool for an accounting practice, not a source of tax answers. It earns its keep on first-draft client emails and letters, plain-English explanations clients will actually read, summaries of long documents and meetings, and internal SOPs. Anything it says about tax law, rates or thresholds has to be checked against the source before it leaves your office, because it will be wrong occasionally and confident every time. Paid individual plans sit at around US$20 per person a month, and team plans cost more per seat while adding admin controls and business data terms.
Keep the accountant in the review loop
A possible document workflow, with professional judgement kept with the accountant.
- 01
Approved inputs
Choose the source documents and confirm what may be shared.
- 02
AI first pass
Extract or summarise the material and identify items to investigate.
- 03
Accountant checks
Verify against the source and make the final professional judgement.
Do not treat generated text or extracted figures as verified accounting records.
What it is actually being used for
The same explanation, forty times
Every practice answers the same handful of client questions all year, rewritten for a different reader each time. The thinking was done years ago. The typing was not.
Letters and emails from a blank page
Engagement updates, fee discussions, chasers, an explanation of why the tax bill does not match the profit. Not hard, just constant, and always at the end of the day.
Documents nobody has time to read
Trust deeds, contracts, long correspondence, sixty-page management reports where four paragraphs actually matter to the client sitting in front of you.
Seven jobs it does well
These are the uses that hold up in a real practice, in rough order of how quickly they pay off. Every one of them produces a draft that a qualified person then reads, corrects and owns. That review step is not a formality, it is the job.
- First-draft client emails. Give it the situation, the outcome you want and the tone, and it will produce something you can edit in two minutes instead of writing in ten. The gain compounds because your reluctance to start the email disappears.
- Plain-English explanations. Take a technical position you have already settled and ask for it in language a business owner can follow, with an analogy. You are not asking it what the answer is, you are asking it to translate an answer you have verified.
- Document summaries. Paste in a long report, contract or piece of correspondence and ask what the client needs to know and what obligations it creates. Then ask it to quote the passage behind each point so you can check it in ten seconds.
- Meeting notes into file notes. Rough notes or a transcript in, structured file note with actions and owners out. This is the single most under-used one in practices I talk to.
- Internal SOPs and checklists. Describe how your team does a job today and have it written up as a procedure, then correct the parts it guessed. Getting a first draft out of a partner's head is the hard bit, and this removes it.
- Spreadsheet and data wrangling. Formula help, cleaning up an inconsistent list of entity names, reshaping an export into the format your software wants. Check the output against the source; do not check the explanation of the output.
- Meeting preparation. Feed it the client's numbers and ask what a sceptical adviser would ask about them. It is a decent devil's advocate and a very cheap one.
Prompt patterns worth copying
The difference between a useless answer and a usable one is almost always context. Tell it who it is writing for, what you already know, what the constraints are, and what good looks like. These four patterns cover most of what a practice needs.
| Job | Prompt pattern | What to check |
|---|---|---|
| Client email | You are drafting on behalf of an Australian accounting firm. Client: [context]. Situation: [facts]. I need them to [action] by [date]. Tone: warm, direct, no jargon. Under 150 words. | Facts, dates and figures. It will happily smooth over a detail you left vague. |
| Plain-English explanation | Explain [concept, as I have described it] to a business owner with no accounting background. Use one analogy. Do not add caveats I have not given you. About 120 words. | That it explained your position rather than substituting a generic one it prefers. |
| Document summary | Summarise this document for [client type]. List the obligations, the dates and anything unusual. Quote the exact wording behind each point you make. | The quotes. If it cannot point to the text, treat the point as unverified. |
| Procedure or SOP | Turn these rough notes into a step-by-step procedure for a first-year accountant. Flag any step where my notes are ambiguous instead of guessing. | The flagged gaps, which is where the real knowledge transfer happens. |
The hard limits
ChatGPT does not know the law, it knows how text about the law is usually written. Those look identical until the moment they are not. Rates, thresholds, dates and treatment questions are exactly where it fails, and it fails fluently, in the register of a confident senior colleague. Verify everything substantive against the primary source. If that sounds like it cancels out the time saving, it does not: drafting and explaining is most of the work, and checking a specific claim is fast.
Three other limits matter in a practice. It does not know your client unless you tell it, so generic output usually means a thin prompt. It has no audit trail worth the name, so the reasoning behind a position still has to be recorded the way it always was. And responsibility does not move: the practitioner who signs the letter owns every word of it, no matter what drafted the first version.
Which plan a practice actually needs
Free accounts are the wrong place for client work, not because of the model but because of the account. Individual paid plans, at around US$20 per person a month, are fine for one person testing the water. Once more than a couple of people are using it for client work, move to a team plan: you get a shared workspace, control over who has access when someone leaves, and business-grade data terms rather than consumer ones.
Enterprise plans exist for organisations that need single sign-on, formal security review and central policy control. For a practice under about thirty people that is usually money spent on governance you do not have the headcount to need yet. Whatever tier you choose, read the current terms yourself before client data goes in, because they change and vendors do not send you a memo.
Common questions
Can accountants use ChatGPT for tax advice?
Not as a source. It will produce confident, well-written answers about rates, thresholds and treatments that are sometimes out of date and occasionally invented, and it has no way of telling you which is which. Use it to draft the explanation and to structure your thinking, then verify every substantive point against the legislation, the ATO guidance or your own research before it goes anywhere near a client. The accountant signs the advice and the accountant carries it.
Is it safe to put client information into ChatGPT?
That is a decision for the partners, and it depends on the plan you are on. Business plans behave differently from free personal accounts on data handling, so read the current terms yourself rather than trusting a summary, including this one. Practical middle ground: strip names and identifiers out of anything you paste, keep the firm on sanctioned accounts rather than personal logins, and put the rules on one page that staff have actually read. This is general information, not legal advice.
Which ChatGPT plan does a small accounting firm need?
Most small firms are best served by a paid team plan rather than individual subscriptions. Individual paid plans sit at around US$20 per person a month; team plans cost more per seat and add a shared workspace, admin control over who has access, and business-grade data terms. Enterprise is worth looking at when you need single sign-on, formal security review or organisation-wide controls, which for a practice under about 30 people is usually not yet.
Official resources
Explore the tools and check current details with the people who make them.
