Source: https://jtpck.com/blog/before-ai-analyzes-your-spreadsheet

# Before AI Analyzes Your Spreadsheet, Check What It Actually Reads

Spreadsheet AI can misread saved formulas, filtered totals, and dates. Check the workbook-to-data conversion before trusting its summary.

By JTPCK Team · Nov 19, 2026

You upload a workbook and ask an AI assistant why orders are running late. Its explanation sounds reasonable, but the total differs from the one your operations manager sees in Excel.

The discrepancy might start before the model receives anything. A workbook contains formulas, saved results, filters, formatting, and multiple sheets. An importer has to decide which of those to preserve. A clean-looking table can conceal a bad decision at that boundary.

We build [AI integrations](/services/ai-integrations) that connect models to business software. For spreadsheet workflows, we recommend agreeing on what the imported records mean before evaluating how well the assistant explains them.

## Define the table you meant to analyze

Consider a hypothetical parts distributor reviewing open orders. Its workbook has an order register, a summary tab, and an old planning sheet. The manager filters the register to one warehouse before asking which deliveries need attention.

Does the question concern all orders, the visible warehouse, or the summary tab? Those are different inputs. The importer should not choose whichever sheet happens to come first or combine every sheet into one undifferentiated block.

Write down the relevant sheet, header row, date range, and inclusion rule. Distinguish detail rows from subtotals and notes. Keep that description beside the resulting analysis so another person can tell whether the assistant answered the intended question.

This is an invented example, not a JTPCK client account. It illustrates a decision the business must make, not a choice the model should silently make for it.

## Distinguish a formula from its saved answer

A cell can hold a formula while the workbook also stores a previously calculated result. Software reading the file may expose one or the other.

For example, the Python library openpyxl documents a setting called `data_only` that selects formula text or stored values. Its formula documentation also states that it does not evaluate formulas. Loading a workbook through that library therefore does not itself produce fresh calculations.

That behavior is specific to the tool, not a claim about every AI spreadsheet product. Ask what your own importer does. Does it use saved values, recalculate in a spreadsheet engine, or translate formulas into another calculation system?

For the distributor, a proposed delivery total might depend on updated quantities elsewhere in the workbook. A saved answer needs a freshness check. A recalculated answer needs confirmation that the calculation engine supports the relevant formulas and has the required inputs.

We recommend preserving the original file and recording how results were obtained. Missing results, formula errors, or unavailable linked inputs should become review items, not zeroes or guessed values.

## Make filters part of the question

A total that describes visible rows may differ from a total over the underlying records.

Microsoft's documentation gives a concrete example: Excel's SUBTOTAL function excludes rows removed by a filter. For manually hidden rows, the function number matters. Numbers 1 through 11 include those rows; 101 through 111 exclude them.

This means “match the total on screen” needs a more precise definition. An importer that reads every detail row could calculate a valid total for the wrong scope.

For the hypothetical warehouse review, record the warehouse filter explicitly and apply that rule to the imported records. Do not rely on a screenshot or the word “visible” to carry the whole definition.

We would also compare the number of included records and a trusted control total before asking for a narrative. If the numbers disagree, identify the missing or extra records first. A model-generated explanation of an unexplained discrepancy can make the problem harder to spot.

## Preserve dates, units, and identifiers

The meaning of a cell can depend on information around it. A column labeled “lead time” needs a unit. An order code that starts with zero should retain its identity. A blank delivery date needs a defined meaning rather than an automatic replacement with today's date.

Date conversion deserves a specific test. Excel supports both 1900 and 1904 date systems. Microsoft documents a difference of 1,462 days between the serial numbers representing the same date in those systems.

We recommend checking the workbook's date system during conversion and comparing a few known dates with the imported values. Do not repair a suspicious date by adding or subtracting years until the underlying interpretation is understood.

Store the source sheet and cell or row reference with normalized values. That gives a reviewer a practical way to check whether “late order” arose from the source data, the conversion, or the business rule.

## Keep calculation separate from explanation

For the distributor, use a defined calculation to identify orders whose promised date precedes the review date and whose status is still open. Let the model explain the resulting set or group free-text delay notes. Keep the rule and the included records inspectable.

This division also makes testing easier. A useful sample workbook could contain an excluded warehouse, a manually hidden row, a formula without a usable saved result, and an ambiguous date. Decide the expected treatment of each before running the import.

A second check should compare the assistant's explanation with the verified records. Did it introduce a cause that no note states? Did it describe a partial set as every overdue order? Correct numbers still need a faithful explanation.

Our guide to [validating structured AI data](/blog/valid-json-does-not-mean-correct-ai-data) covers the checks needed before extracted values become business records. The spreadsheet-specific work starts earlier, with preserving what the workbook actually means.

If a reliable report or database query already answers the question, use it. AI may help interpret messy notes without replacing the arithmetic. If you need help connecting the two, [tell us what the workbook contains and what decision it should support](/contact). Start with a description rather than confidential files.
