You will be able to get an assistant to label each transaction with a category without asking it to add anything up.
Priya pasted her cleaned March file into an assistant and typed "categorise my spending and tell me where my money went." She got a neat summary with food at about S$740, transport at about S$210 and shopping at about S$380, plus a suggestion that she cut back on eating out. It had the look of a finished budget, yet she had no way to tell which rows went into which number, whether every row had been counted, or whether the totals were added correctly.
That one request asked the assistant to do two jobs. One it does well, the other it does badly. This lesson splits them.
Start by writing your own category list, before you open the assistant. If you let it invent categories, you'll get different ones every month, and a budget is only useful if this month can be compared with the last.
Keep the list short. Something like food, transport, housing, bills, insurance, subscriptions, shopping, family, health and other is enough for most people. You can split food into groceries and eating out if that's where your attention goes. Add one category called transfer for money moving between your own accounts, which lesson 2.3, Check the labels and fix the rules, explains.
Give each category a one-line meaning if there's any doubt. "Family: money sent to parents or siblings, gifts and red packets." "Bills: phone, internet, utilities, town council." The model follows definitions far better than it guesses what you meant.
The request has three parts in this order: the category list with its definitions, then the rows, then the instruction. Priya's version went roughly like this:
"Below are my card transactions for one month, with date, description and amount. Add a fourth column called category and use only the categories in my list. Return every row in the same order and format, as CSV, with nothing added or removed and no totals. If you are not reasonably sure of a category, write CHECK instead of guessing."
Each part of that is doing a job. Using only your categories keeps the labels consistent. Same order and same format means you can paste the result straight back into your spreadsheet, next to the original rows. Asking for CSV gives you something a spreadsheet can read. Nothing added or removed makes it easier to spot when a row has gone missing, which can happen on long lists.
When the result comes back, compare the row count with your cleaned file. If you sent 87 rows and got 85, find the missing two before you do anything else.
The CHECK instruction matters more than it looks. A model asked to label every row will label every row, even when it has no idea what one of them is. A transfer to someone's initials might become shopping. A one-word merchant name might become food because food is common. Those guesses look exactly like the confident labels.
Giving it an honest way out changes the result. You'll typically get a handful of CHECK rows: PayNow transfers, unfamiliar merchants, payments that could be two things. You'd have needed to look at those rows anyway, and now they're flagged instead of buried among the confident labels. You know what that S$60 to "J LIM" was, so sorting them is quick for you and impossible for the assistant.
Back to Priya's first attempt. The trouble with asking for totals in prose is the one lesson 1.1, What an assistant is actually doing when you ask about money, described. Unless the assistant is visibly running code, it writes digits the way it writes words, by predicting them. Adding forty or eighty amounts in a paragraph is exactly the kind of multi-step sum where those predictions drift. A total can be off by a few dollars or by a few hundred, and it reads the same either way.
There's a second problem: you can't trace a number in a paragraph. If the food total looks high, there's nothing to click on to see which rows it came from, whereas a spreadsheet total can be traced in seconds. That's why the totals in this module come from SUMIF or a pivot table in lesson 2.4, Build a month's budget from your own data, where every figure points back to rows you can see.
Some assistants will run code if you ask them to, and show it. That's far better than a figure in prose. But you still need the labelled rows in your own sheet for next month, so it's simpler to get the labels and do the adding where you can check it.
If your month has a few hundred rows, send it in batches of fifty or so, with the same category list and instruction each time. Long inputs are where rows get dropped or labels drift halfway through. Batches also make the row-count check quicker.
Your next step is to send your own cleaned month through with your fixed list and the CHECK rule, then paste what comes back into a new tab beside the original.
Run your cleaned month through an assistant with a fixed category list and paste the labelled rows into a new spreadsheet tab.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).