How can accountants and bookkeepers use AI?

Much of bookkeeping is careful, repetitive reading: matching a cryptic bank description to the right expense account, explaining a change in a client's numbers in plain words, building the same spreadsheet formula again. Large language models, the technology behind ChatGPT, are good at exactly this kind of first pass. They are not good at being right every time, and they carry no professional responsibility. A safe workflow uses the first strength and guards against the second.

In a 2023 interview published by CPA Canada, Malik Datardina, a CPA on its Audit and Assurance Technology Committee, put the division of labor simply: hand the drudgery to the tool and keep the critical thinking and decisions for yourself. He also warned against putting confidential client or employer data into a public chatbot. Those two ideas shape everything below.

What AI does well in an accounting practice

• First-draft categorization. Suggesting an account for each line of a bank or card export, with a reason, so you review instead of type.
• Plain-language writing. Client emails, engagement reminders, and short memos that explain what changed and why, written from facts you supply.
• Spreadsheet help. Writing and explaining Excel or Google Sheets formulas, such as a lookup that matches vendor names to account codes.
• Reading long documents. Summarizing a lease, a loan agreement or a software contract so you know which clauses to read closely.

Since May, free ChatGPT users with GPT-4o can upload files and use data analysis, the feature that runs small programs on a spreadsheet you provide. That makes it easy to work on a CSV export directly, which is also why data protection comes first.

Protect client data before you start

Remove or replace names, account numbers, addresses and tax identifiers before any file leaves your system. A vendor name like "Office supplies store" is usually all the model needs. Check how your tool handles data: OpenAI states that conversations in ChatGPT Team and Enterprise workspaces are not used to improve its models, while consumer accounts have their own data controls to review. Your firm's policy, your engagement letters and your professional body's guidance on confidentiality come before any convenience.

A worked example: categorizing expenses

Here is an anonymized extract a bookkeeper might paste in. The transactions and amounts are illustrative.

You are assisting a bookkeeper. Suggest an expense category for each line, using only this chart of accounts:
Office supplies, Software subscriptions, Meals, Travel, Telephone and internet, Bank fees, Uncertain.

Rules:
- Return a table with: line number, date, description, amount, category, one-sentence reason.
- If a line could fit two categories, choose Uncertain and say why.
- Do not change any date, description or amount.
- After the table, give the number of lines and the total of all amounts.

Lines:
1. 2024-07-02, STAPLES #1182, 84.37
2. 2024-07-03, ADOBE *CREATIVE CLD, 79.99
3. 2024-07-05, PETRO-CAN 0441, 61.20
4. 2024-07-08, MONTHLY ACCT FEE, 12.95
5. 2024-07-09, AMZN MKTP CA, 146.88

Notice what the prompt does. It limits the model to your chart of accounts, gives it an honest way out (Uncertain), forbids it from touching the source data, and asks for a count and total you can check. A sensible reply would mark the Amazon line as Uncertain, since a marketplace purchase could be almost anything, and might do the same for the fuel line, which could be travel or a vehicle expense depending on your chart.

The reconciliation checks

Before anything goes into your books, run these checks yourself:

1. Count. Five lines in, five lines out. A missing or duplicated line is a common error when models rewrite tables.
2. Total. Add the original amounts in your own spreadsheet (here, 385.39) and compare with the model's total. If they differ, find the line that changed.
3. Line by line. Compare every amount and date against the source export, not against the model's table.
4. Uncertain lines. Resolve each one from the receipt or invoice, not from the model's guess.

For a memo, use the same discipline. Give the model the verified figures and ask it to explain them, for example: "Using only these numbers, draft a five-sentence note to the client explaining why software costs rose from the prior quarter. Do not add figures or advice." Then check every number in the draft against your workpapers.

Which numbers must always be checked by hand?

Every total, subtotal and line count the AI reports, because language models can drop, repeat or mistype lines while sounding confident. Recompute totals in your own spreadsheet from the original data, never from the model's output.

Every figure that will appear in a client document, a financial statement, a filing or a payment: balances, amounts, rates, dates and periods. Compare them with the bank statement, invoice or ledger they came from. The same goes for any calculation the model performed in its head; if it did not show its work through data analysis, treat the result as unchecked.

Finally, account codes and categories on anything material or unusual. The model's reason can sound plausible and still be wrong for your client's situation. As the CPA Canada interview noted, reviewing AI output takes real expertise, and the professional who signs off is still the one responsible.