AI for Bookkeeping: What It Speeds Up and the Totals It Should Never Touch
TL;DR: Use AI to sort, describe, summarize and draft. Let your accounting software and a calculator do the math. Ask for every figure with the line it came from, and check those lines yourself before anything goes in the books or out to a client.
AI for bookkeeping is useful, as long as you're clear about which jobs you hand it. An AI assistant is fast at the reading and writing around the numbers: suggesting a category for a vague transaction, explaining a line on a statement, listing what looks out of place, or drafting the note that goes with a monthly report. It's not a calculator, and it's not your ledger. Its drafts always need checking, and it never gets the final say on a total.
This lesson walks through what to hand over, what to keep, and a prompt that makes every figure traceable. You don't need to install anything beyond the AI assistant your company approves.
The one rule: AI reads and writes, software counts
AI chat assistants generate text. When they produce a number, they're writing what looks like the right number. Sometimes it's wrong, and a wrong number reads exactly as confidently as a right one. The US National Institute of Standards and Technology calls this risk "confabulation", the "production of confidently stated but erroneous or false content" (NIST AI 600-1, Generative AI Profile, July 2024). You'll also hear it called hallucination.
In most writing, a small error is a typo. In bookkeeping, a wrong total can mean a wrong tax figure, a wrong invoice or a wrong decision. So the rule for this whole lesson is simple: totals, balances, tax amounts and anything you'd file or bill come from your accounting software or a spreadsheet formula. The AI can point to where a number is. It doesn't produce the number of record.
What AI speeds up
These are the bookkeeping tasks where an AI draft saves real time and a person can check the result quickly:
- Suggesting categories. Paste a list of transaction descriptions (with account numbers removed) and ask for a suggested category for each, with a one-line reason. You approve or change each one.
- Explaining statements. Ask what a fee, code or line item on a supplier statement probably means, then confirm with the supplier or the bank.
- Spotting what looks off. Ask it to list duplicate-looking entries, round-number payments or descriptions that don't match their category. It's a second pair of eyes for you to follow up on.
- Drafting the words. Month-end notes, a plain-language summary for a business owner, a polite email asking a client for a missing receipt.
- Writing a checklist. Turn your month-end close into numbered steps so it doesn't live in one person's head.
Every one of these produces a suggestion that a person accepts or rejects. None of them changes the books by itself.
What it should never touch
Keep these with your software, your formulas and your judgment:
- Account balances, reconciled totals and anything that has to tie out
- Tax calculations, payroll figures and amounts on any filing
- Final invoice amounts and payment approvals
- Deciding how to treat a transaction for tax when the rules aren't obvious; that's a question for an accountant
If you want a summed figure in a draft, calculate it in your spreadsheet or accounting software and paste the result in. Don't ask the chat to add it up.
Before you paste anything: what stays out
Follow your company's AI policy and use only the AI tool your company has approved. Then strip the data before it goes in. Never paste:
- Full bank account, card or routing numbers
- Tax ID numbers, Social Security numbers or national ID numbers
- Online banking logins, passwords, API keys or one-time codes
- Client or employee names and addresses when the task doesn't need them
- Payroll details for named people
Replace them with placeholders such as [CLIENT A] or [ACCOUNT ending 1234]. Bank details, pay and tax records can do serious harm if they leak, so keep them out of the chat unless your company's AI policy says otherwise.
The prompt: find what doesn't match, and show the line
This prompt compares two lists, such as a bank statement export and the transactions in your books for the same month, and flags what looks unmatched. It never calculates a total. It points you to the lines so you can check them.
The Prompt:
You are helping a bookkeeper review a month of transactions. Do not calculate totals or balances.
Compare LIST A (bank statement) with LIST B (our books). For each item that looks unmatched, missing, duplicated or different in amount or date, output one row:
| Issue | Line in LIST A | Line in LIST B | What looks different | Suggested next check |
Rules:
- Quote the exact line from each list. If an item is only in one list, write "none" for the other.
- Copy amounts exactly as written. Do not round, convert or add them.
- If you're not sure two lines are the same transaction, say "possible match" and explain why.
- After the table, list anything you could not read or compare.
LIST A:
"""
[paste the statement lines, account numbers removed]
"""
LIST B:
"""
[paste the matching lines from your books]
"""
Why This Works: The first line takes arithmetic off the table, which removes the riskiest job. Quoting the exact line from each list means every flag can be checked in seconds against the source. "Possible match" gives the model a safe way to be unsure instead of forcing a yes or no. Asking for a fixed table keeps the output in the same shape from month to month. When software needs to read the result, the stricter version of this is called structured output.
What you get back: A short table of items to look at, each with the original lines quoted, and a note on anything the AI couldn't compare. You then check each row in your accounting software, where the actual reconciliation happens. The AI's table is a list of leads to check. The reconciliation in your software is the record, which is why the software stays in charge.
Trace every figure back to its source
Whenever an AI draft includes a number, it should come with where that number came from: a line on a statement, a cell in a spreadsheet, a page in a report. Add this line to any prompt that drafts a summary or a client note:
For every figure you mention, add the source in brackets: the document name and the line, row or page it came from. If a figure isn't in the material I gave you, don't include it.
Then check each bracket against the original. If a number has no source, delete it. This habit also keeps your records clean. The AI's summary is a working note. The original statements, receipts and invoices are your records, and you'll need them later. The IRS, for example, says to keep records for 3 years in most cases and employment tax records for at least 4 years (IRS, How long should I keep records?). Rules differ by country and situation, so ask your accountant what applies to you.
Where to go next
If you work with an accountant or you are one, two Applied guides go further: drafting client advisory letters with AI and prompting for CPAs after tax day. For more lessons written for your job, go back to the finance and bookkeeping track, or browse the Applied pillar.
Frequently asked questions
Can AI do bookkeeping?
AI can help with parts of it: suggesting categories, explaining statement lines, flagging entries that look wrong and drafting notes and emails. It shouldn't keep the books or calculate totals. Your accounting software and a person reviewing the work stay responsible for the numbers.
Is it safe to upload bank statements to ChatGPT or another AI tool?
Only to the tool your company has approved, and only after you remove account numbers, tax IDs, logins and personal details. Your company's AI policy decides which tools are allowed and what data can go into them. If there's no policy yet, ask before you upload financial records.
Can AI do bank reconciliations?
It can help you find items to look at, such as missing, duplicated or mismatched transactions, if you ask it to quote the lines and not calculate anything. The reconciliation itself should be done and signed off in your accounting software, where the totals are calculated from the actual records.
Why does AI get numbers wrong?
AI chat assistants generate text, and a number in their answer is generated the same way as the words around it. It can be wrong while sounding certain. NIST calls this confabulation (NIST AI 600-1). That's why totals belong in software and every figure in an AI draft needs a source you can check.
Will AI replace bookkeepers?
Nobody can promise how any one job will change. The approach in this lesson uses AI only for the routine reading and drafting, because someone still has to decide how transactions are treated, reconcile the accounts, answer for the numbers and talk to the client or the business owner. Those are the parts this lesson keeps with you.
Rather have someone set it up for you?
That's a fair call. LTFI, the agentic business operating system from Kief Studio, can set up the work these guides describe and run it with you. No website yet? Kief Studio will build you one for free.

