Why AI Models Struggle with Spreadsheets
A manager uploads a sales workbook and asks an AI assistant which region performed best last quarter. The assistant returns a clear answer, complete with percentages and an explanation. During review, someone notices that it included a subtotal row alongside the transactions used to calculate that subtotal. The math may be correct for the rows it selected, but the answer is wrong.
Spreadsheets create opportunities for this kind of mistake because their meaning depends on how information is arranged, calculated, and maintained. A workbook can contain several tables, assumptions scattered across tabs, and formatting conventions understood only by the people who use it.
Language models can help analyze these files. Their reliability depends partly on whether the surrounding software gives them enough information to interpret the workbook correctly.
Layout Carries Meaning
People read spreadsheets spatially. A heading spanning several columns may identify a reporting period. Blank rows may separate departments. A shaded section may contain estimates, while the section below contains actual results.
When software converts a worksheet into plain text, some of those relationships can disappear. A model might receive the values without enough information to determine which heading applies to which cells. Screenshots preserve visual details, but they introduce another challenge: reading and locating individual values accurately.
Microsoft’s SpreadsheetLLM research identifies spreadsheet layout, formatting, and the limits on how much information a model can receive as obstacles to reliable interpretation. Its results show that the way spreadsheet information is represented can materially affect performance.
Consider a worksheet containing monthly sales figures followed by quarterly totals. Before calculating annual revenue, the system must distinguish the monthly figures from the quarterly summaries so that the same sales are not counted twice. A request as simple as “add up the sales” leaves room for an expensive misunderstanding.
A Displayed Number Has a History
A cell displaying $125,000 might contain a manually entered figure, a formula referencing another worksheet, or a calculation dependent on an external workbook. Those differences matter when someone asks why the number changed.
Reading the displayed value provides limited evidence about how it was produced. Reading the formula introduces further work: identifying its inputs, following references, and checking the assumptions behind them.
The tools used to open the file also matter. For example, the Python library openpyxl can read and write formulas, but it does not evaluate them. It can retrieve a previously stored calculation result when one is available. That stored result does not establish that the workbook has been recalculated after subsequent changes.
An AI application therefore needs to establish what it actually inspected. Formula text, saved results, and freshly recalculated outputs provide different kinds of evidence. An explanation based on an outdated result can sound entirely reasonable.
Business Rules Are Often Missing
Getting every value out of a workbook correctly does not guarantee that the analysis will answer the right question. The spreadsheet may never explain what its labels and conventions mean.
Take a column called “Revenue.” One company might use it for booked contracts, while another uses it for issued invoices or recognized revenue. Blank and negative cells can be just as ambiguous. Employees may know from experience that a blank means an entry is awaiting approval or that a negative number represents a refund, but the file itself may not say so.
This becomes a problem when someone asks which customers are most profitable. The workbook may contain sales and shipping expenses while leaving out support costs or discounts recorded elsewhere. A model could still produce a convincing ranking, but it would be based on an incomplete definition of profitability.
In that situation, the application should pause and ask what costs belong in the calculation. One short clarification can prevent the system from answering a different question from the one the user intended.
Reliable Analysis Needs Checks
A useful spreadsheet workflow begins with inspecting the workbook: identifying worksheets, table boundaries, headers, formulas, and relevant notes. The system then needs to select the records that belong in the analysis and establish how ambiguous entries should be handled.
Calculations should run through appropriate tools, with checks that address the task. A sales analysis might reconcile its total against an existing summary after checking that both use the same records, period, and definitions. Agreement is a useful check, but the summary itself may also need verification. A percentage calculation should identify its denominator. A comparison between periods should confirm that both periods use equivalent definitions.
Tool use still requires scrutiny. A script can calculate the wrong range with complete precision. Reviewers need enough detail to trace the answer back to the selected worksheets, rows, and assumptions.
For companies introducing AI into spreadsheet work, a practical starting point is a bounded task with a clear definition of success. Identify overdue invoices using an agreed cutoff date. Compare expenses using named categories. Explain a formula while preserving its references.
These tasks give teams something concrete to verify. As the work becomes more complex, the same discipline matters: preserve context, define the question, execute calculations, and check the result against the workbook.