AI Spreadsheet Analysis: How to Tell Whether the Answer Is Right

When AI spreadsheet analysis reports a growth rate, ask which rows and definitions produced it. A fluent explanation is not enough to verify a figure. Keep the source data, inspect the calculation and row counts, and recompute a few results independently before relying on the report.

Duplicate records, refunds and mixed currencies are useful starting points for review. Each needs an explicit rule and a check against the source file.

Where the numbers usually go wrong

An apparently small choice about which rows to include can materially change the total. Check these choices before interpreting the percentage change.

Duplicates are the classic. An export that ran twice, or an order that was edited and saved as a new line, means the same sale appears two times. Sum the column and that order counts double. Nothing looks broken; the total is just a little too high.

Refunds and cancellations are the next one. If refunds live in a status column or show up as negative amounts, a total that ignores them reports gross sales while you think you are reading net. The difference can be most of a month's growth.

Mixed currencies are sneakier. A column called "amount" that holds euros for some rows and dollars for others will add up without complaint, because to arithmetic they are all just numbers. Unless something converts them, or at least flags the mix, the total means nothing.

And then there are figures that appear only in the prose. A summary says "revenue roughly doubled" or names a percentage that never shows up in any table or calculation. If you cannot point to where a number was computed, treat it as unverified, however reasonable it sounds.

What to ask for besides the answer

Ask for the calculation, not just the result. A short description of each step (which rows were kept, which were dropped and why, which column was summed) makes the calculation easier to review.

Ask for row counts at each stage: how many rows came in, how many were removed as duplicates, how many were excluded as refunds or test orders, how many were left. Keep these counts with the result. They can still be wrong, but they give you concrete figures to compare against the source file.

Ask for totals reconciled against the source. If the monthly figures add up to a grand total, that grand total should match what you get from the raw file with the same rules applied. A mismatch is a reason to investigate before using the result.

Finally, ask for the assumptions to be stated. "Treated negative amounts as refunds." "Assumed all amounts are in USD." An assumption you can read is one you can correct.

Phrasing questions so the definitions are explicit

Vague questions invite silent choices. "What were sales in March?" leaves open whether sales means gross or net, whether March means order date or payment date, and what to do with duplicates.

Compare that with: "Using order date, sum net sales per month, where net means the amount after refunds. Remove exact duplicate rows first and report how many you removed. Flag repeated order IDs whose other fields differ for review. Then give the month-over-month change as a percentage." It is longer, but every number in the answer now has a definition you chose. If you are not sure what the right definition is, say so and ask the analysis to list the options it sees in the data before it picks one.

A worked check of AI spreadsheet analysis on a sample

The public sample on ToCharts is a good place to practise, because the data is small enough to check by eye. It uses two synthetic CSV files: an orders table with 16 rows and a product table. One of those 16 rows is a duplicate.

So the first thing to confirm is the count. Sixteen rows in, one duplicate removed, 15 orders left. If an analysis reported 16 orders, you would need to check whether it reported raw rows or deduplicated orders, then inspect the totals.

Next, the monthly totals. The sample reports net sales and gross amount separately, which is exactly the kind of definition worth checking, and net sales go from 9,200 in the first month to 13,850 in the second. Now do the arithmetic yourself: 13,850 minus 9,200 is 4,650, and 4,650 divided by 9,200 is 0.5054, so month-over-month growth is 50.54%. The stated figure matches. If the duplicate had been left in, one of those monthly totals would be higher and the percentage would move, which is why the row count comes first. For the formula itself and its edge cases, the month-over-month growth topic goes through it in more detail.

The sample also ranks products and shows each region's share, and the same habit applies: check that the shares add up to the total and that the ranking uses the same net figure as the monthly view.

How the ToCharts workbench handles this

The AI spreadsheet workbench at /spreadsheet is built around these checks. You upload CSV, TSV or XLSX files, optionally with text PDF, DOCX, PPTX, TXT or HTML files as background, up to five files and 10 MiB in total. You ask in plain business language and can follow up, and later questions reuse the earlier files and results.

The agent inspects the files and can run calculations in Python in an isolated sandbox. The run log records steps, assumptions and validation results. Check what each validation covers: a code assertion checks a declared condition, while source reconciliation separately recomputes totals under a stated rule. A result may have structural checks without independent source reconciliation. Failed recorded checks block delivery, but passed checks do not prove that the business definition is correct. Deliverables include an Excel result workbook with native charts, CSV tables, JSON and an HTML dashboard; PPTX can be requested.

It needs you to sign in and uses EndTo credits. The sample at /tocharts/sample does not, and opening it is a quick way to see what a reconciled result and its run log look like before you upload anything of your own.

What this can't do yet

The workbench does not run OCR, so scanned PDFs and images of tables are not read, and it does not execute macros. Exported Excel files do not keep the original workbook's formulas, and it does not connect to databases, BI tools or cloud drives. Charts in the deliverables are bar, line and pie. For the free chart makers, there are no scatter, radar, bubble, gauge or heatmap chart types — only bar, horizontal bar, line, area, pie and donut. Reconciliation checks that totals match the input under the stated rules; it cannot tell you whether those rules are the right business definition, so read the assumptions.