Why AI Still Struggles With Excel

Artificial intelligence can write code, generate images and explain difficult subjects in seconds. Then you ask it to make a small change to an Excel workbook, and suddenly a dropdown has vanished, the colours have moved and Excel is offering to “repair” the file.

Why does AI struggle with Excel when it seems capable of much harder work?

Because an Excel workbook is rarely just a table. A serious spreadsheet is closer to a small software application that happens to be wearing a grid.

Excel has more going on than you can see

A basic workbook may contain a few values and formulas. A business workbook can also contain named ranges, hidden sheets, dropdowns, tables, charts, conditional formatting, protected cells and links to other files.

These features are stored as separate but connected parts of the workbook. An AI tool may correctly change a visible cell while accidentally breaking the validation rule behind it. It may add a row but fail to extend a chart range or a formula elsewhere.

The workbook can still open normally. Nothing appears to be wrong until somebody notices that the monthly total has quietly excluded the last line since Tuesday.

Spreadsheets rarely explain themselves

Normal software tends to make its structure reasonably explicit. Functions have names, variables have types and failures often produce an error message.

Excel is more informal.

A yellow cell might mean “enter data here”. A grey cell might mean “never touch this”. The company’s entire pricing model may live on a hidden worksheet called Sheet3 (2).

A person who has used the workbook for years understands these unwritten rules. AI has to infer them from labels, formulas, formatting and examples. That becomes difficult when several people have edited the file over time and the closest thing to documentation is a red cell saying “DO NOT DELETE”.

This is especially risky with tools such as quote calculators, stock sheets and product-code generators. The AI may produce an answer that looks perfectly believable while missing one small rule that everyone in the office simply knows.

A small edit can have a large blast radius

A request such as “add another product option” sounds simple. Inside the workbook, it might involve updating the source table, extending a dropdown, changing a lookup formula, preserving the output format and checking that older selections still work.

Adding a row can affect:

• Formula and lookup ranges
• Dropdown sources and table sizes
• Charts and conditional formatting
• Named ranges and print areas

Some parts of Excel expand automatically. Others remain fixed. A few appear to make their decision based on whether you are already late for a meeting.

This hidden web of dependencies is one reason AI spreadsheet work can fail. The visible request is small, but the file may require several exact changes in different places.

AI usually is not using Microsoft Excel

Most AI systems do not control the full desktop version of Excel. They usually work through programming libraries that read and write .xlsx files.

Those libraries are useful, but they are not perfect copies of Excel. Microsoft has spent decades adding features, compatibility rules and historical oddities to the format. Third-party tools may support ordinary cells and formulas while only partly supporting advanced charts, pivot tables, macros, external connections or some conditional formatting rules.

The AI can therefore make the requested edit correctly and still damage an unrelated feature when it saves the file.

That is when Excel opens with a message saying it found a problem and repaired the workbook. This sounds reassuring, although it is roughly the software equivalent of a mechanic returning your car and saying, “I removed the unusual noise.”

Writing a formula is not the same as calculating it

Spreadsheet libraries can usually insert a formula such as =A1+B1. They cannot always calculate and store the fresh result.

Excel normally recalculates formulas when the file opens. Until then, another program may see an old cached value, a blank cell or a result from before the latest edit.

Newer Microsoft 365 functions add another layer. A formula may work in current Excel but behave differently in an older version, LibreOffice or Google Sheets.

So the formula can be correct while the value shown on screen is wrong. That is a particularly Excel-shaped problem.

Formatting can be part of the data

Formatting is often treated as decoration, but in spreadsheets it can carry meaning.

Colours may show which cells are editable. Number formats may preserve leading zeros. Conditional formatting may flag invalid combinations. Borders may separate product groups or stages in a process.

Changing 09 to 9 looks harmless if it is a number. It is not harmless if 09 is part of a product code. The same problem appears with dates, percentages, currency and long reference numbers.

Excel also likes to guess what a value means. It does this quickly, confidently and occasionally with the accuracy of someone assembling furniture without the instructions.

An AI editing a workbook has to preserve both the raw value and the way Excel interprets it.

The worst spreadsheet errors are quiet

A broken website normally produces a visible error. A broken spreadsheet may continue working and return a plausible answer.

A lookup range might stop one row early. A formula may use the wrong relative reference. A number may be stored as text. A total may exclude the final item. None of these problems necessarily causes a warning.

This is why spreadsheet errors can be more dangerous than obvious failures. A price of £418 may look reasonable even if the correct answer is £481.

AI can build or edit the workbook quickly, but checking every meaningful input combination is a separate job.

New spreadsheets are easier than old ones

AI is generally more reliable when creating a clean workbook from scratch. The structure can be consistent, the tables can be documented and the formulas can follow predictable patterns.

Editing an established workbook is harder because the AI first has to understand the original design, including the parts nobody remembers creating.

It is similar to renovating an old house. You remove one panel and discover wiring, pipework and a handwritten note from 2007 saying “probably important”.

The workbook may also contain years of workarounds that are technically odd but operationally essential. Tidying one of them can break three other things.

How to get better results from AI and Excel

AI spreadsheet tools work best when the workbook is treated like software rather than a disposable document.

Keep input and output areas clear. Document important lookup tables. Save a backup before major changes. Make one change at a time and test the file in Microsoft Excel afterwards.

For important workbooks, create a small set of known test cases. A product-code generator, for example, might have ten sample selections with the exact output each one should produce. After every update, those cases can be checked again.

This catches the classic spreadsheet problem where fixing one option quietly breaks another option nobody tested.

The problem is precision, not intelligence

Modern AI can often understand what a workbook is supposed to do. It can explain formulas, suggest a cleaner design and identify likely problems.

The difficult part is making exact changes while preserving every formula, style, rule and hidden dependency inside the file.

Excel is flexible because it combines data, calculations, automation and interface design in one place. That same flexibility makes complex workbooks fragile.

AI does not struggle with Excel because spreadsheets are too basic. It struggles because Excel is far more complicated than it looks — and because somewhere inside nearly every important workbook is a hidden tab called FINAL_v2_NEW_USE_THIS_ONE.