For routine, repeat transactions (the same software subscription, the same utility, the same fuel station) AI categorizes about as well as a careful bookkeeper, as long as it is given your client's own chart of accounts and rules. For the ambiguous ones, such as a big-box store run that mixes supplies and personal items, owner transfers, loan payments or a purchase that might be an asset, it does not, so the job becomes reviewing its exceptions; you can build a categorizer that saves every line for review in AI Expense Categorizer Agent.
The difference comes down to context. A bookkeeper knows that this client's "Home Depot" charges are job materials, that the 3rd-of-the-month transfer is the owner's draw, and that the new laptop goes on the balance sheet. AI only knows what is in the transaction line and what you tell it. Give it more context and it gets better. Leave it guessing and it picks the likely answer, confidently.
Which transactions can AI handle, and which need a bookkeeper?
Usually fine for AI, with spot checks:
- Recurring vendors with one obvious account: software, phone, internet, rent, insurance
- Bank and card fees, interest income
- Fuel, parking and tolls for a business with vehicles
- Meals and travel when the client's rules are clear
Needs a person every time:
- Mixed purchases from stores like Amazon or Costco, where one charge covers several accounts
- Transfers between the client's own accounts, owner draws and contributions
- Loan payments that must be split into principal and interest
- Purchases that might be fixed assets instead of expenses
- Sales tax, payroll liabilities and anything that hits the balance sheet
- Refunds, chargebacks and anything that looks like a duplicate
- A vendor the client has never paid before
How do I set AI up to categorize well?
- Give it the client's chart of accounts, with a one-line description of what goes in each account. A generic list of categories gives generic results.
- Add the rules you already know: "Shell is always Vehicle Fuel", "transfers to savings are not expenses", "anything over your capitalization threshold goes to review".
- Ask for a short reason with each category, so a reviewer can see why.
- Tell it to mark low-confidence lines "Needs review" instead of choosing.
- Review the exceptions and every new vendor, then add the answer as a new rule.
- Reconcile afterward. Categorizing is not reconciling. The bank rec still has to tie.
What does this look like in a real lesson?
In the AI Expense Categorizer Agent lesson you build an agent in n8n for a made-up firm, Maple Street Bookkeeping. You add a chat trigger, an AI Agent with a system prompt that holds the firm's categories and how to note each expense, an OpenAI chat model and simple memory. Then you create a Google Sheet called Expense Log with five columns (Date, Vendor, Amount, Category and Note) and connect it as a tool with Append Row. In the test you paste a batch of expenses, the agent categorizes them with a note for each, and on your next message it writes the rows to the sheet, where they wait for review before anything is posted.
AI Receipt Extractor does the same from a photo: the agent reads a Corner Mart receipt, files it as Groceries and adds the date, vendor and total to a sheet. Whether "Groceries" is right for that client is exactly the call a bookkeeper still makes. And Month-End Bank Reconciliation shows the step that catches what categorizing misses, such as a bill recorded twice.
Which tools do bookkeepers use for this?
- QuickBooks Online, whose Intuit Assist AI categorizes transactions, and Xero, with bank rules and its JAX AI assistant.
- Booke AI, built to categorize and reconcile transactions for accounting firms.
- Dext, which pulls data from receipts and invoices and posts it to your accounting software.
- n8n if you want to build your own categorizer that writes to a review sheet first.
What mistakes should I avoid?
- No chart of accounts in the prompt. Without it the AI invents its own categories.
- Auto-posting. Keep a review step between the AI and the ledger, at least until you trust it on that client.
- Never reviewing the "obvious" lines. Spot-check a few each month, since one wrong vendor rule repeats every month.
- Ignoring the client's tax context. Personal versus business, and meal and vehicle rules, change the answer.
- Pasting full bank statements with account numbers into a personal chatbot. Use approved tools and strip identifiers.
For the basics of the account list itself, see What Is a Chart of Accounts?, and for more automations, AI automation for bookkeepers. For the bigger career question, see Will AI replace bookkeepers?
Key takeaways
- AI handles repeat, single-account vendors well when it has the chart of accounts.
- Mixed purchases, transfers, loans, assets and new vendors need a bookkeeper.
- Ask for a reason and a "Needs review" flag on every uncertain line.
- Review exceptions, add rules, then reconcile.
Frequently asked questions
- Is AI categorization accurate enough to skip review?
Not on a new client, and not for balance sheet items. As rules build up, you can review fewer of the routine lines, but keep reviewing exceptions and new vendors.
- Can ChatGPT categorize a bank CSV?
Yes, if you upload the client's chart of accounts with it and ask for a reason and a review flag on each line. Use a business plan, remove account numbers, and enter the results in your accounting software yourself.
- Does AI get better on one client over time?
Tools that learn from your corrections do, and so does a prompt you keep updating with new rules. A one-off chat does not remember last month unless you give it the rules again.