How to Use AI as an Accountant: What Works, What to Avoid, and the Rules

Accounting is unusual among the professions being reshaped by AI, because the constraint is not capability. It is accountability. You sign things. A model does not, and when it is wrong it is wrong in a fluent, confident, professionally formatted way. That single fact should shape every decision about where you let it near your work.

This is the practical companion to Will AI Replace Accountants?, which looks at the risk to the job. This one is about getting value out of the tools without creating a problem for yourself.

Start with the obligations, because they narrow the options

The professional bodies have been clear that the ethical framework does not change because the technology did. ICAEW's guidance on generative AI and ethics works through the standard fundamental principles — integrity, objectivity, professional competence and due care, confidentiality, professional behaviour — and applies each to AI use. Two of them do most of the constraining.

• —Confidentiality. Client information does not go into a general consumer tool. Once it is submitted you have limited visibility over how it is retained, who can access it, and whether it is used for training. The AICPA publishes a small firm generative AI policy template that is a sensible place to start if your firm has nothing written down.
• —Professional competence and due care. Undue reliance on an automated output is a breach whether or not the output happened to be right. You remain responsible for the judgement, which means you have to be able to check it — and ICAEW has warned separately that AI tools can land firms in regulatory trouble when that check is skipped.

Practically, that gives you three tiers of work. Things you can do with anonymised or generic inputs in any tool. Things that need a tool with an enterprise agreement and no training on your data. And things you should not hand over at all.

What it is genuinely good at

The pattern across all of these: the model produces a draft, you supply the judgement. None of them involve the model being the authority on anything.

1. 01Client correspondence. Explaining a variance, a deadline, or why a deduction was disallowed, in plain language, at length, without you writing it from scratch for the fortieth time. This is the highest-volume win in most practices and the lowest risk.
2. 02Turning a spreadsheet into an explanation. Paste the anonymised figures, ask for the three things that moved and a short narrative for the management pack. You already know what the numbers mean; the model saves you the writing.
3. 03Formula and query construction. Describing what you want in words and getting an Excel formula, a Power Query step or SQL back. Easy to verify — run it and check the result against a case you can compute yourself.
4. 04Structuring a first pass at documentation. Process notes, a memo skeleton, a working paper outline. The structure is generic even when the content is not.
5. 05Reading long documents against a question. Loan agreements, lease terms, a new policy document — "where does this say anything about early termination fees" is a search problem, and it does that well when the document is in front of it.
6. 06Rehearsing an argument. Ask it to challenge a position you plan to take with a client or a reviewer. It is useful as a sparring partner precisely because it does not care about your conclusion.

A worked example

You have a quarterly management pack with a gross margin down 4 points. You know why: two large jobs ran over and one supplier moved prices in month two.

The bad version: "write a commentary on this management pack" with the file attached. You get generic business-school language, possibly a confident explanation of a trend that is not there, and client data in a tool you have not vetted.

The version that works: strip identifiers, give it the two facts you already established, and ask for a 150-word commentary for a non-financial director that leads with the supplier price change and quantifies the job overruns separately. You get a draft in your register, you edit two sentences, and you have saved twenty minutes without outsourcing a single judgement. The model wrote the prose. You did the accounting.

Use it where you already know the answer and the work is expressing it. That is the whole rule, and it holds across every task on the list.

Where it will quietly get you into trouble

• —Tax and regulatory research. It will produce a section number, a threshold and a confident summary, and the details may be from a repealed provision, another jurisdiction, or nothing at all. Treat every citation as unverified until you have opened the primary source yourself. This is the single most common way accountants get burned.
• —Arithmetic across a large dataset. It is a language model. It can produce a plausible total that is wrong, and unlike a spreadsheet it will not error — it will just be off. Compute in a tool that computes.
• —Anything where the answer depends on a fact only your client knows. It will fill the gap rather than ask, unless you have explicitly told it to ask.
• —Judgement calls dressed as questions. "Is this capital or revenue" produces an answer with the tone of authority and none of the responsibility. You still own that call.
• —Reconciliations you did not check. Automating the matching is fine. Accepting the exceptions list without opening it is not. The exception-review routine that keeps automated coding safe is in how to use AI as a bookkeeper.

The connecting thread is that these systems fail confidently rather than visibly, which is a much harder failure mode to catch than an error message. What AI Is Actually Bad At covers the pattern in general, and How to Check an AI Answer When You Are Not the Expert covers what to do when you are working slightly outside your own specialism — which, in a small practice, is most weeks.

Getting started without a project plan

1. 01Write down your firm's rule on client data before you do anything else, even if the rule is one sentence. A template beats nothing and nothing is what most small firms currently have.
2. 02Pick one recurring task where you can judge the output instantly. Client emails are the usual right answer. Find the Repetitive Part is a short test for choosing it if the answer is not obvious.
3. 03Do that task both ways for two weeks — your way and with a draft — and keep the time. You want evidence, not a feeling.
4. 04Build a small library of the prompts that worked. The value compounds through reuse, not through cleverness.
5. 05Expand to a second task only once the first is boring. Firms that roll out five things at once end up using none of them.
6. 06Decide what you will tell clients. Disclosure norms are still forming and the professional press is actively debating them; having a considered position is better than being asked and improvising.

The part that actually protects your job

The tasks most exposed here are the ones furthest from judgement: data entry, categorisation, straightforward reconciliation, first-draft correspondence. Those were already being automated by software that predates any of this. What is left — advice, exception handling, the conversation with a client about a decision they have to make, and being the person who signs — is not automatable in the same way, because the value is partly that a named professional is accountable for it.

Which means the useful stance is neither refusing the tools nor deferring to them. It is being the accountant who can produce a draft in a fifth of the time and then catch the thing in it that is wrong. That second half is the skill, and it is the one worth practising deliberately.

The short version

Use AI for the writing, the structuring, the searching and the first draft, on anonymised inputs or in a properly vetted tool. Do not use it for tax authority, arithmetic at scale, or any judgement you would have to defend. Check every citation against the primary source. Write down a data rule before you start, pick one task, measure it, and expand only when the first one is dull.

Coursium teaches exactly this layer — practical use of the tools plus the checking that has to go with it. Stay ahead of AI by learning them on your phone.