ChatGPT for Accountants: What It's Actually Good At (and Where It Will Bite You)

Be honest about how this started at your firm. Nobody ran a procurement process. Someone in the team opened a free ChatGPT account in 2023 to reword an awkward email, and it spread from there. Three years on, ChatGPT is still the front door to AI for most accountants. Most firms' policies, where they exist at all, still don't mention it by name. That gap between what your team is actually doing and what your firm has actually decided is the single riskiest thing about ChatGPT in practice. The tool itself is manageable. The unmanaged use of it isn't. The numbers say nearly every accountant is already using AI, and only a fraction are doing it on purpose.

So: what ChatGPT is genuinely excellent at inside a firm, the specific ways it will bite you, and what has to be true before it goes anywhere near client work.

What it's genuinely excellent at

Drafting client communications. This is the killer app, and it's not close. The email explaining to a client why their tax bill went up when their profit went down. The engagement-scope pushback that needs to be firm without torching the relationship. The fee-increase letter you've been putting off since March. ChatGPT turns "here are the four points I need to make, the client is annoyed, keep it warm" into a solid draft in ten seconds. You'll edit it (you should edit it), but you're editing instead of staring at a blank compose window. For a firm sending dozens of these a week, the time saved is real and immediate. The one thing ChatGPT can't see is your meetings, which is where a purpose-built assistant like Vinyl picks up, drafting the follow-up email from the call itself (Vinyl is a commercial partner of The Firm).

Explaining concepts at any level. "Explain Division 7A to a 26-year-old founder who's never heard of it." "Give me a one-paragraph explanation of accrual accounting for a restaurant owner." "Reframe this depreciation schedule conversation for a client who keeps asking why they can't just deduct the whole ute." It's a genuinely good translator between accountant-speak and client-speak, in both directions, and it's just as useful for training juniors as for briefing clients.

First-pass analysis of exported spreadsheets. The paid tiers let you upload files, and ChatGPT can write and run actual code against them, which matters, because code doing arithmetic is reliable in a way the model's own mental maths is not. Upload an aged receivables export and ask which clients blew out past 60 days this quarter. Upload twelve months of P&Ls and ask what moved. It's good at finding the shape of the story in the data: trends, outliers, the three things worth a partner's attention. It is a first pass, not a working paper, but as a first pass it's faster than any grad.

Research triage. Not research: triage. "What are the general eligibility considerations for the R&D tax incentive?" is a fine question if you treat the answer as a map of what to go verify, not as the verification. ChatGPT is good at telling you what the relevant concepts and likely provisions are, so you spend your billable research time in the actual legislation and rulings instead of orienting yourself. Used that way, it shortens research. Used as the research, it's a liability, for reasons the next section makes uncomfortably specific.

Where it bites

It makes up numbers. Not occasionally: structurally. A large language model is a text-prediction engine, not a calculator, and when you ask it to compute something in conversation it produces figures that look right with total confidence. Ask it the same superannuation contribution question three times and you can get three different answers, each delivered with the same certainty. The fix is knowing which mode you're in: ChatGPT running code against your uploaded file is doing arithmetic; ChatGPT answering a numerical question from memory is doing vibes. Never let a conversationally-generated figure into anything a client sees without recomputing it yourself.

Tax specifics by jurisdiction. ChatGPT's training data is a blend of US, UK, Australian and everything-else content, and it will cheerfully blur them. It will quote you a threshold that's two years stale, apply an IRS rule to an ATO question, or describe a UK relief with an American accent. It's also trained to be agreeable, so a leading question ("this is deductible, right?") tends to get a leading answer. Anything involving a current-year rate, threshold, deadline or eligibility test needs checking against the ATO, HMRC or IRS source: every time, no exceptions.

Citations. This is the failure mode with a body count. In Mata v. Avianca (2023), two New York lawyers filed a brief citing six cases that ChatGPT had invented outright, complete with plausible case names, citations and quotes. When challenged, they asked ChatGPT whether the cases were real, and it assured them they were. A federal judge sanctioned them US$5,000 and the story went global. Swap "case law" for "tax rulings" and "private binding rulings" and the mechanism is identical: ChatGPT will fabricate a convincing-looking ruling number or ATO ID with exactly the same fluency it fabricates court cases. If you can't pull the source document yourself, the citation doesn't exist.

The free-tier data problem. On ChatGPT's consumer plans (Free, Plus, and the cheaper tiers), OpenAI can use your conversations to train its models by default. There's an opt-out in settings, but defaults are what actually happen, and the default on the account your senior pasted a client's payroll summary into is: that content may go into the training pipeline. That's not a hypothetical breach of confidentiality. Depending on your engagement terms and jurisdiction, it may be an actual one. This isn't unique to ChatGPT, and it isn't a reason to ban AI; it's a reason to stop doing firm work on personal accounts. We've covered the broader boundary problems in what AI still can't do in accounting. This one is just the cheapest to fix.

The tier that matters

Here's the part that changes the conversation with your risk-averse partner. ChatGPT's business tiers, Business (formerly Team) and Enterprise, flip the data default: OpenAI states it does not train on business customers' data. The business plans are independently audited to SOC 2 Type 2, encrypt data in transit and at rest, and add SSO, an admin console and centralised billing, with Enterprise layering on longer retention controls and data residency options. Pricing moves around, but Business has sat in the range of roughly US$25-30 per seat per month. Check OpenAI's current pricing page before you budget.

Read that pricing against the risk. The gap between "our team uses personal ChatGPT accounts we can't see" and "our team uses firm accounts where client data isn't training anyone's model and an admin can see usage" costs about as much per person as a coffee run. If ChatGPT is going to touch client work at all (and at most firms it already is, sanctioned or not), the business tier isn't an upgrade. It's the entry requirement.

ChatGPT vs the alternatives, in one paragraph

Short version: ChatGPT is the best-known and most feature-broad; Claude tends to be preferred for long-document work, careful drafting and its connections into accounting tools like Xero; Copilot's case is that it lives inside the Microsoft 365 stack your firm already licenses. They're all capable, they all hallucinate, and the data-handling questions in this article apply to every one of them. The full comparison (including which one fits which kind of firm) is in our three-way breakdown of Claude vs ChatGPT vs Copilot for accounting firms, and the wider landscape lives in our AI for accounting firms guide.

Five rules before your team uses it on client work

1. Business tier or nothing for client data. Personal accounts are fine for generic questions and blank-page drafting. The moment client-identifiable information enters a prompt, it happens on a firm-controlled Business or Enterprise account, or it doesn't happen.
2. Every number gets recomputed. Any figure ChatGPT produces in conversation is a draft of a number, not a number. If it goes into a working paper, a return or a client email, a human or a spreadsheet recalculates it first.
3. Every citation gets pulled. No ruling, section reference, case or ATO/HMRC/IRS position goes into anything until someone has opened the actual source document. If you can't find it, it isn't real.
4. Jurisdiction goes in the prompt, verification goes against the source. Always state the country and year in your prompt, and still check current rates and thresholds against the regulator's own site before relying on them.
5. A named human signs off. Every AI-assisted output that leaves the firm has a person's name attached to its review, same as work from a junior. Because no AI is ever going to jail for you: the registered agent on the engagement is still you.

Write those five down, put them in the policy, and you've done more deliberate AI governance than most firms have managed in three years.

---

Frequently asked questions

Is ChatGPT safe for client data?

On a personal Free or Plus account, no. Consumer plans can train on your conversations by default, which sits badly against most confidentiality obligations. On ChatGPT Business or Enterprise, OpenAI states it doesn't train on your data, and the plans carry SOC 2 Type 2 auditing, encryption and admin controls. Safe enough is a judgment for your firm and your engagement terms, but the business tier is the minimum credible starting point. The full vendor-by-vendor breakdown of training, retention and review terms is in Is AI safe for client data?.

Can ChatGPT do tax returns?

No. It can't lodge anything, it isn't connected to the ATO, HMRC or IRS, and, more importantly, it fabricates figures and misstates jurisdiction-specific rules often enough that an unreviewed return would be professional negligence. It can help with the work around a return: summarising source documents, drafting the client query list, explaining a position in plain English. The return itself remains the job of a professional using actual tax software, with their name on it.

Is the free version of ChatGPT enough for an accounting firm?

For a sole practitioner drafting generic emails with no client data in the prompts, arguably. For anything beyond that, no. The free tier has tighter limits, misses the file-upload and data-analysis features that make ChatGPT useful on real work, and carries the train-on-your-data default. If ChatGPT is saving your team even an hour a month, the paid business tier already pays for itself several times over.

What are the best uses of ChatGPT in accounting?

The words-and-structure work: drafting client emails and letters, explaining technical concepts to clients or juniors, first-pass analysis of exported reports (on a paid tier, where it can run code against your file), summarising long documents, and orienting yourself before proper technical research. The pattern across all of them: ChatGPT drafts, a human verifies, and nothing numerical or citable ships without a check against the source.

Will ChatGPT replace bookkeepers?

Not on current evidence. ChatGPT doesn't connect to bank feeds, doesn't lodge, doesn't take responsibility, and gets details wrong in ways that only a human who knows the client would catch. What it does compress is the mechanical middle of the work (categorising descriptions, drafting queries, summarising activity), which shifts the bookkeeper's job toward review, exceptions and client relationships. The bookkeepers at risk aren't the ones AI replaces; they're the ones who ignore it while their competitors do the same work in half the time.