Can AI do bookkeeping? What it can and can't do safely

The problem: hype on one side, fear on the other

Some headlines suggest AI will run the books on its own. Others suggest it's too unreliable to touch. Neither is useful when you're trying to decide what to do on Monday morning.

A better question is: which tasks are mostly writing, organizing and pattern-spotting — and which tasks require verified data, judgment and accountability? AI is strong at the first group. You own the second.

What AI can do well

Used through a chat tool with clear instructions, AI is useful for:

• Drafting client emails, month-end summaries and plain-English explanations of reports.
• Turning rough notes into organized checklists and procedures.
• Suggesting how to group a messy list of account names or vendor names you paste in.
• Explaining an accounting concept in simpler words for a client.
• Spotting items in pasted text that match rules you give it, like "flag anything labeled Miscellaneous."
• Helping you build simple, read-only tools that summarize an exported report.

What AI should not do on its own

These stay with you, every time:

• Posting entries or changing anything in a client's books without your review.
• Deciding how an unusual transaction should be classified.
• Reconciling accounts — the statement and your software are the source of truth.
• Giving tax, legal or financial advice to a client.
• Producing numbers that aren't already in a source document.
• Signing off. Your name goes on the work, so your review does too.

The AI solution: the draft–flag–review pattern

The safest way I've found to use AI in accounting work is a simple pattern: AI drafts, AI flags, you review. You write the rules into your prompt — including what it must not do — and ask it to list anything uncertain instead of guessing.

You are assisting a bookkeeper. You only draft and flag; you never decide.

Rules: Use only information provided. Do not invent numbers. Do not give tax, legal or financial advice. If anything is unclear, list it under "Needs bookkeeper review" instead of guessing.

Task: [describe task]

Input: [paste fictional or anonymized data]

A quick test: is this task safe to hand to AI?

Ask three questions before you start:

• If the AI gets this wrong, will I catch it in review? If not, don't use AI for it.
• Is the output a draft for me, or a final answer for someone else? Drafts are safer.
• Am I sharing anything my firm's policy or client agreement doesn't allow? If yes, stop and anonymize.

What you check every time

Keep a short log of what you asked, what came back and what you changed. It helps you improve your prompts and gives you a record of your review.

• Numbers match the source exactly.
• Nothing in the output is advice.
• Flags were listed, not quietly "resolved."
• Data you shared was allowed to be shared.
• You'd put your name on the final version.

The professional still matters

AI doesn't replace the knowledge that tells you a balance looks wrong, a vendor is miscoded or a client's question is really about something else. It makes the writing and organizing around that knowledge faster. That's the shift: your expertise becomes the instructions the AI follows.

Where to start this week

If you've never used AI in your work, start with something low-risk and repeatable. My suggestion is the monthly client summary email. You already know the numbers, you already review the email before sending, and the AI simply saves you the first draft.

Write a prompt using a role, rules, the input and the format you want. Run it on fictional numbers first. Compare the draft to what you'd normally write. Change one instruction at a time until the drafts feel like yours. Then try it on a real month — with your firm's data policy in mind — and review it like any other piece of work.

After a few weeks, pick a second task, like turning close notes into client questions. Build slowly. The goal isn't to use AI everywhere; it's to use it where it reliably helps and where your review will catch any mistake.