Why AI gets spreadsheet numbers wrong — and how to check an answer
The DataMimi team · Published 2026-09-12 · 8 min read
An AI tool that answers questions about a spreadsheet is right most of the time, and that is the problem: a wrong total looks exactly like a right one. When an answer is wrong, it is almost always wrong in one of three ways, and each one leaves a trace you can check in a minute.
1. The file was read wrong
Business exports are not clean tables. The header is on row 5 under a company name and a date range. Section titles are merged across the whole row. A totals row sits at the bottom, or in the middle after each region. Amounts arrive as text — "$1,250.00" — because the export kept the currency symbol.
A tool that reads the first row as the header, or sums a column that includes the totals rows, gets a confident answer that is double the real figure. Nothing about the answer tells you that happened.
- Headers that are not on row 1
- Totals and subtotal rows counted as data
- Merged cells that leave most rows of a column blank
- Numbers and dates stored as text, skipped by the sum
- Two date formats in one column, read as different days
2. The calculation ran on the wrong rows
Even with the file read correctly, the answer depends on which rows were included. "Revenue last quarter" needs a date filter; "revenue from repeat customers" needs a definition of repeat. Duplicate rows from a copy-paste count twice. Blank cells are silently excluded from an average.
These are not model failures so much as unstated choices. The answer is only checkable if the choices are stated.
3. The number was written, not computed
A language model writes text that is statistically likely. Asked to describe a table, it can produce a percentage or a total that appears in no calculation at all — a plausible figure, in the right place in the sentence.
The safeguard is structural: the figures in the written answer should be copied from a computed result, not generated alongside it. If a tool cannot show where a number in its summary came from, treat the number as unverified.
Five checks before a number goes into a report
None of these needs code. Each takes under a minute, and together they catch nearly every wrong answer described above.
- Row count: does the number of rows used match what you expect after filters? If the file has 1,240 orders and the answer used 2,480, something was counted twice.
- One total by hand: sum the main column yourself in Excel for the same filter. If it disagrees with the AI's total, stop.
- Every figure in the prose appears in the table: a percentage in the summary that you cannot find in the result table was written, not calculated.
- Filters are named: "Q2 2026, excluding refunds" is checkable; "last quarter" is not.
- Source columns are named: an answer about margin should say which cost column it used.
What we built DataMimi to do about it
DataMimi reads the file first and reports how it read it — which row is the header, which rows were totals, which columns held numbers as text — before any number is quoted. The analysis runs as code on the full file in an isolated sandbox; the model writes the code and the explanation, not the arithmetic.
Every answer carries a source line: the file and sheet, the columns used (checked against what the code actually read), any filters, and how many rows the figures were computed over. When the data needs cleaning, the changes are previewed with real before-and-after examples, and nothing changes until you choose.
Common questions
Is it safe to use AI for financial analysis?
It is safe to use it for the analysis and unsafe to skip checking the result. Use a tool that shows the rows and columns behind each figure, and reconcile at least one total by hand before a number leaves your team.
Why does AI sometimes give a different total each time?
Usually because the question left a choice open — which date range, whether refunds count — and the tool made a different choice each time. Stating the filter in the question, and checking the filter the answer reports, makes results repeatable.
What is the fastest single check?
Sum the main column yourself for the same filter and compare. It takes seconds and catches totals rows, duplicates and text-stored numbers at once.
Try it on your own file