AI data summaries keep coming out wrong not because the model is careless, but because it is not a calculator in the first place. A large language model generates text token by token, predicting what word should come next, rather than actually carrying out arithmetic. With multi-digit addition, subtraction, multiplication and division, carries easily slip out of place. With a long table, it can skip rows, read the same row twice, or mix figures that were never defined the same way, and still deliver a total in a confident tone that simply does not add up. The fix is direct: stop making it do mental math, hand the calculation to formulas and code, and let it focus on understanding the request and spelling out the steps.
Where exactly it goes wrong
First is the arithmetic itself. The more digits and intermediate steps there are, the more likely text generation is to slip on one digit, and unlike a calculator it will not flag an error. Second is reading the table: when data is too long it may be truncated, so the second half is never fully read and the summary is missing a chunk. Third is inconsistent definitions, for example some rows including tax and others not, some already net of refunds and others not, all added together as if they were the same kind of figure. Stacked together, these errors sound convincing and are hard to spot at a glance.
Switch to these four steps so the result can be checked
- Define the scope before it starts: state clearly which rows are included, whether tax is in or out, whether refunds are deducted, and which field to group by, and only let it proceed once that is confirmed.
- Make it write a formula or code to do the math: ask for a spreadsheet formula, a table function or a short script so software actually runs the calculation, instead of letting it report a total directly.
- Require intermediate subtotals step by step: have it output subtotals by month, by category or by group, then add the subtotals up, so a mismatch can be located immediately.
- Spot-check against known anchors: verify the total row count first, hand-check one or two subtotals you can calculate yourself, and compare the result with a pivot summary of the original table; only accept it when all three agree.
Where AI can only play a supporting role
For payment settlement, payroll and expense claims, where a single wrong digit is unacceptable, AI can only help define the scope, write formulas and flag anomalies; the final figures must be rechecked in spreadsheet software before use. Clean the data before feeding it in as well: if the table has merged cells, repeated header rows or blank rows mixed in, split, deduplicate and standardise the format first, otherwise even a carefully defined scope will be built on messy input and every later step will inherit the error.
