ChatGPT for accountants: what to trust it with, and what to never

A note before the detail. The tools and their data terms change, so treat the specifics as a snapshot, and this is general guidance, not professional or compliance advice, your professional standards body and your firm's policy override anything here. The durable principle holds regardless, hand AI the drafting and the summarising, keep it away from the maths, the filing, and the client data.

You have watched colleagues use ChatGPT and felt the pull. You know the client emails you write fifty times a year, the regulations you translate into plain English, the bank statements you retype by hand.

Right next to that pull is the worry, because you know exactly what is at stake if it gets a number wrong or client data ends up somewhere it should not. Nobody has drawn you a clear line between the two, so you either avoid it and do everything the long way, or you use it and feel slightly exposed.

The line is clearer than it looks. ChatGPT is a strong assistant for the words and a dangerous one for the numbers, and almost everything sorts cleanly onto one side of that split.

The short answer. Trust ChatGPT with the repetitive writing, summarising and reformatting that eats your day, things like drafting client emails, turning dense regulation into plain English, and tidying data into a usable shape. Never trust it with the actual maths, the tax calculations, the filing, the final advice, or anything carrying identifiable client data on the standard version. It does not do deterministic accuracy, and it does not keep secrets unless you make it.

What to trust it with

These are the jobs where a fast, confident first pass saves real time and a small error costs little, because you are reviewing the output anyway.

• Draft the emails you write on repeat. Deadline reminders, document-request follow-ups, status updates on a return in progress. Give it the context and the tone you want and it gets you a solid draft to edit rather than a blank page.
• Translate regulation into plain English, both for yourself and for clients. It is good at taking a dense section of tax code and explaining it simply, or summarising a long document into the parts that matter, as a starting point you then verify.
• Reformat and tidy data. Turning the text of a bank statement into a clean CSV, restructuring a financial statement into a table for Excel, generating a draft chart of accounts. With the identifying detail stripped out, this is some of the safest time it will save you.
• First-pass review and anomaly spotting. It can scan a transaction log and flag what looks unusual, duplicate payments, odd weekend entries, amounts out of line with previous periods. It is a first layer that surfaces things worth a closer human look, not an audit.
• Generic research and the things you would rather not ask a colleague. An Excel formula, a definition, a general how-to. Quick, low-stakes, and nobody has to know you asked.

What to never trust it with

These are the jobs where it is unreliable, unaccountable, or a liability, and where being wrong is expensive. For the broader version of this list that applies to any profession, see the tasks you should never give ChatGPT.

• The actual maths. ChatGPT produces calculations by predicting what a plausible answer looks like, not by calculating, so tax computations, totals and anything that has to be exact can come back confidently wrong. Use proper software for the numbers, and if you ask AI at all, treat the output as something to verify, never to rely on.
• Preparing or filing returns, and final compliance work. It lacks the deterministic accuracy compliance demands, and it does not carry your professional responsibility, you do. It can help you understand and draft around a problem, but it cannot be the thing that files.
• Final advice and strategic calls. Use it to explain options or prepare for a client conversation, not to decide. Business situations are nuanced, recent legislation may not be in its training data at all, and it has no way of knowing what it does not know about a specific client's position.
• Identifiable client data, on the standard version. This is the one that can cost you a licence, not just a little embarrassment, and it has its own section below.

The client-data rule

Keep real client detail out of the consumer version, full stop. On the standard tiers, what you paste can be used to train the model and is retained for a period even when you have turned training off, and a confidentiality or privilege breach is a professional problem, not a privacy preference.

Anonymise before you paste. Strip the names, the account numbers, the identifying detail, and run the prompt on "Client A" and "Vendor 001" instead. If your firm provides an enterprise version with a proper data-protection agreement, use that for anything sensitive, it exists for exactly this reason.

The simplest rule: if you could not put it in a public document, do not paste it into a public chatbot. The same logic applies to Claude and Gemini. The consumer tiers train by default, the business tiers do not.

The line that keeps you safe

The thread through all of it is one rule. You are still responsible for the output, every time. AI drafts, you verify and you sign.

Use it to take the tedium out of the words and the data wrangling, and keep your own judgement firmly on the numbers, the advice and the filing. Never let a confident-looking answer skip your review just because it looked finished.

That is the same judgement that applies to any AI work, scaled to a profession where the cost of a wrong number going out is high. Get that line right and ChatGPT gives you back hours without ever putting your name at risk.

Clair helps non-technical professionals know when to trust their AI, when to check it, and when to skip it.