ChatGPT and similar assistants get expense totals wrong because they write text that looks like the right answer instead of doing arithmetic, and long lists of numbers make that worse. The fix is simple: let a spreadsheet or tracker do the adding, ask the assistant for row counts, and cross-check any total before you rely on it.
Why do language models miscount?
An assistant builds its answer one piece at a time, choosing what is likely to come next. Adding two small numbers is easy for it; adding sixty is a long chain in which one slip changes the result and nobody notices. It may also skip a row in the middle of a long paste, read a credit as a debit, or round a figure without saying so. The result is a total that looks tidy and is wrong.
What specific mistakes should you look out for?
- Skipped or repeated rows, especially in the middle of a long list.
- Indian number formats read wrongly, such as 1,25,000 treated as 1 and 25,000.
- Credits and debits mixed, so a refund is added instead of subtracted.
- Currency symbols or notes such as 'Cr' and 'Dr' confusing the parse.
- A total stated confidently with no list of the rows used.
Layout matters too. Pasting a table copied from a PDF often scrambles columns, so an amount ends up beside the wrong date, or a balance column gets added to the debits by mistake. Paste only the columns you need, and keep the amounts free of rupee signs and commas if you can. Ask for the first and last row it read, which reveals a truncated paste.
How do you get the right answer?
Use the assistant for understanding and a spreadsheet for sums. Paste your amounts into a sheet and use a formula:
- Grand total: =SUM($C$2:$C$200)
- Number of rows: =COUNT($C$2:$C$200)
- One category: =SUMIF($B$2:$B$200,"Groceries",$C$2:$C$200)
- Rows in a date range: =SUMIFS($C$2:$C$200,$A$2:$A$200,">="&$F$1,$A$2:$A$200,"<="&$F$2)
In the last formula, cell F1 holds the start date and F2 the end date. If you want the assistant's help, ask it to write the formula for your layout, then run it yourself.
Which prompts reduce the risk?
Ask for row counts and for the rows used: 'How many rows did you read? List them before you add them.' Ask it to show subtotals first and the total last. Ask it to flag any row it could not read as a number. Better still, keep each prompt short: ten or twenty rows at a time are far safer than three hundred.
A worked illustration, with invented numbers. You paste 62 rows adding up to ₹48,350 and the assistant answers ₹46,900. The difference is 48,350 − 46,900 = ₹1,450, which could be one skipped row or a misread amount. Run =COUNT($C$2:$C$63) to see whether your sheet also has 62 rows, then ask the assistant to list the rows it used. A single missing line of ₹1,450 explains the gap exactly, and you have your culprit.
Are some setups more reliable?
Some assistants can run code or read data from a connected tool, and the sums then come from calculation rather than prediction. Even so, you should see the rows it used, and compare the total with a report from your own records. When Claude or ChatGPT is connected to SpendixAI through MCP, you choose the scope, can start read-only and can disconnect at any time; if the assistant quotes a total, open the matching report and compare.
A habit that keeps you safe
Treat any total from an assistant as a draft. Before you act on it, do three things: count the rows, sum them in a sheet, and check that the category totals equal the grand total. If any of the three disagrees, the assistant's number is the one to drop.
Useful ideas start with a clearer picture of your own spending.
Explore your reports


