What AI Can and Cannot Do for a Bookkeeper's Client Admin

Think of an AI writing assistant as a fast drafter with no memory of your clients. It phrases things well. It doesn't know your facts. Most of the sorting below follows from that.

How the tool works inside

The tools people call AI assistants are built on large language models. Wikipedia describes one as software trained on huge amounts of writing. It can typically produce, condense, translate and examine text.¹

Many begin training by guessing the word that comes next. Extra training then teaches them to follow instructions and act like an assistant.¹

So the tool continues text in a believable way. It isn't looking anything up in your ledger. Unless a product has been connected to your files, it knows only what you type into it.

That's why fluent and correct are two different things. Wikipedia calls a made-up answer a hallucination, meaning false or misleading information given as fact.² The Australian privacy regulator agrees these tools are known to give inaccurate or false results.³

Where it helps: the talking around the books

The useful jobs are the ones where you already know the facts and need good words.

• A polite first request for a missing statement, see chasing missing documents
• A welcome message and a list of what a new client should send, see client onboarding messages
• A reminder built from dates you've checked, see reminder and deadline messages
• A short question about a payment you can't identify, see query emails
• Rewording your own rough notes into something shorter and clearer

In each case you hold the facts. The tool holds the phrasing. Because you know what the message should say, you're well placed to spot a draft that doesn't.

Where it should stay out

Some jobs look like wording but aren't.

Advice. What counts as a deductible expense, how to treat a transaction, what a client owes, what a contract means. These are tax, accounting and legal questions. The tool will answer them in a confident voice, and it may be wrong. Send readers to the official source instead. The Australian Taxation Office says that if its information doesn't fully cover your circumstances, you can contact it or seek professional advice.⁴

Facts it can't hold. Due dates, fees, balances, tax rates, names. If a message needs one, you type it in, and you check it again after the draft comes back.

Sending. Nothing should reach a client without a person reading it. The Australian regulator talks about building human oversight into processes when a business adopts an AI product.³ For a small practice, that can be as plain as one rule: you read it, you send it.

Client details. Names, account numbers and ID numbers raise privacy questions before they raise writing questions. There's a whole page on the client data question.

A simple test for any task

Ask two questions before you hand a job to the tool.

First, do I already know the right answer? If yes, it's a wording job. If no, the tool can't supply it.

Second, would a mistake in this message cost me or the client something real? If so, slow down and read it twice, or leave it out.

What good use looks like

Treat the tool like a quick drafting helper who sometimes states guesses as facts. You'd give that person clear instructions, use placeholders instead of real details, and read everything before it went out.

Start with one small, low-risk message type. Once the drafts reliably match what you'd have written, add the next. The starter prompts page has wording you can copy for that first trial.