Excel Copilot vs ChatGPT for Spreadsheet Work

By the XLsheetAI Team · Updated August 15, 2026 · 10 min read

The one-sentence answer

Give the job to Copilot when the answer depends on what is actually in your workbook, and to ChatGPT when the answer depends on knowing Excel — because Copilot runs inside the file and can read your columns, while ChatGPT only ever sees the description you type, which makes it excellent at reasoning about spreadsheets and unreliable at referring to yours.

That single test settles most of the "which one do I open?" moments. The rest of this article explains the mechanism behind it: what Copilot can and cannot see, why the licence matters more than people expect, why ChatGPT confidently invents column letters, what each tool does with your data, and whether paying for both is defensible.

The difference that causes all the others

Copilot in Excel is not a chatbot bolted onto the ribbon. It runs with a view of the data in the sheet you have open — the headers, the values, the shape of the table. When you ask for a margin column, it knows there is a Revenue column and a Cost column, knows which letters they sit in, and can insert a formula column that references them. When you ask it to summarise by region and month, it can build the PivotTable rather than describe one.

ChatGPT has no connection to your file at all. It is a very well-read colleague on the phone. It knows Excel deeply — the function library, the version differences, the edge cases, why your nested IF collapses — but everything it knows about your workbook is whatever you have told it in the last few messages. Ask for a formula without describing the layout and it will invent a reasonable-looking layout and answer for that one instead.

So the two tools are not competing at the same task. One is a pair of hands inside the file. The other is expertise that has to be relayed in and out by you.

What Copilot in Excel can actually see

Copilot's view is narrower than most people assume, and knowing the boundary saves a lot of arguing with it.

It sees the data you are working in, and it strongly prefers that data to be a real table: one header row, one row per record, no merged cells, no blank spacer rows, no second mini-table parked three columns to the right. Selecting your range and pressing Ctrl+T before you start is the single highest-return thing you can do for Copilot's accuracy. Sheets that are laid out as a printed report rather than a dataset are where it fails most often, and it usually fails by quietly working on the wrong block.

It does not see the rest of your world. Other workbooks, linked files, the Power Query steps that produced the sheet, the VBA module in the background, the note you left in a comment three months ago — none of that is in scope. If a number arrives via an external link, Copilot sees the number, not where it came from.

It has size limits, and they bite before you expect them. Very large ranges get summarised or refused rather than reasoned over row by row. Microsoft has been adding Python-backed analysis to Copilot in Excel, which pushes that ceiling higher for statistical and charting work, but the practical rule stands: the tighter and cleaner the range you point it at, the better the answer.

It usually proposes rather than imposes. Formula columns and PivotTables arrive as something you accept or reject, which is the right default. Treat every accepted suggestion as a change you are responsible for: spot-check the first three rows and the last one before you move on. Copilot writing a formula against the real headers removes the transcription errors, not the logic errors.

The licence question, which trips up more people than the AI does

A lot of "Copilot is missing from my ribbon" confusion is a billing question wearing a technical costume. Copilot in Excel is delivered through a Microsoft 365 subscription rather than through Excel as a product, so:

• Perpetual licences do not get it. Excel 2021 and Excel 2024, bought once and owned, sit outside the subscription channel entirely. No amount of updating adds Copilot.
• In a business tenant it is typically a paid per-user add-on layered on top of a qualifying Microsoft 365 plan, and an administrator has to assign it. Your colleague having it does not mean you do.
• Consumer Microsoft 365 plans include an allowance rather than unlimited use, so heavy sessions can run into a cap.
• File location matters. Copilot is most reliable on workbooks saved to OneDrive or SharePoint with AutoSave on. Support for locally stored files has been improving, but if Copilot is behaving oddly on a file sitting on your desktop, save it to the cloud and try again before blaming the model.

ChatGPT has none of these gates. It runs in a browser for anyone, including a usable free tier, which is exactly why so many spreadsheet questions end up there regardless of what the company pays for.

Why ChatGPT gets your column references wrong

This is the complaint that sends people back to Copilot, and it is worth understanding rather than treating as the model being careless.

When you ask for "a formula to sum amounts for the North region", the model has to fill in half a dozen unknowns before it can write anything: which column holds amounts, which holds region, which row the data starts on, whether the range is a named table or a plain A1 range, how far down it goes, whether your Excel has XLOOKUP, and whether your locale separates arguments with commas or semicolons. You supplied none of that. So it picks the most common arrangement and answers for that — amounts in B, regions in A, data from row 2 — and hands you a formula that is correct for a sheet you do not have.

It is not hallucinating in the usual sense. It is answering a well-posed question that happens not to be yours. Give it the missing half and the failure rate drops sharply:

Excel 365, UK locale (commas).
Data is an official table named Sales, rows 2-4,300.
Headers: A OrderDate | B Region | C Rep | D Units | E UnitPrice | F Amount

Sum Amount for Region = "North" where OrderDate is in the current year.

Three lines of context, and you get a formula that references F and B rather than an invented layout. Pasting the header row with its real column letters is the highest-value habit there is when working with any AI that cannot see your screen — and it is the step almost nobody takes.

The other reliable fix is a screenshot. ChatGPT reads images, so a picture of the sheet with the column letters and the formula bar visible communicates the layout faster than any description. This is genuinely something Copilot cannot do for you: if the spreadsheet is in a PDF, a support ticket, an email attachment you cannot open in Excel, or on someone else's monitor, the screenshot route is the only route.

What ChatGPT is genuinely better at

Explaining a formula someone else wrote. Copilot can describe what a formula does; a general model tends to be better at the part you actually need, which is why it was written that way, what breaks it, and what the previous author probably intended. Paste in a 200-character nested monster and ask for a line-by-line breakdown plus the three inputs that would make it return the wrong answer.

Drafting complex logic before it exists. Deeply nested conditionals, a LET that names six intermediate steps, a LAMBDA you plan to reuse, an array formula that spills. This is composition work rather than data work, and it goes better in a conversation where you can iterate on the logic — "now handle the case where the date is blank" — without touching the workbook until the logic is settled.

Translating between languages. Excel to SQL, Excel to DAX, formula to Python, a plain-English rule to a regular expression. This is where a general model is clearly ahead, because the job is knowledge of two systems rather than sight of one file. Our Excel to SQL formula guide covers the mapping in both directions if you would rather learn it than ask each time.

Deciding what to build at all. "Should this be XLOOKUP or INDEX+MATCH given my team is on Excel 2019?" is a judgement question with tradeoffs, not a data question — see XLOOKUP vs VLOOKUP for the version rule. Copilot will happily write the formula; it is less interested in telling you the design is wrong.

One caveat worth stating plainly: uploading a file to ChatGPT is not the same as Copilot's access to it. The upload is analysed as a copy in a sandbox, and what comes back is typically a new file you download, with values rather than live formulas, and formatting, PivotTables and conditional rules often lost. It is a good way to get an answer about data. It is a poor way to edit a workbook you intend to keep.

Side by side

Data privacy: what leaves the building

The privacy comparison is less about which vendor is more trustworthy and more about which direction the data moves.

Copilot does not move it. It operates on a file that already lives in your Microsoft 365 tenant, under the permissions you already have — it cannot show you a workbook you could not open yourself, and it inherits the sharing and retention rules your administrators set. Microsoft's commercial terms state that prompts and responses in the business offering are not used to train the underlying foundation models. For a regulated dataset, that boundary is usually the whole argument.

ChatGPT requires you to move it. Every paste and upload is an export from your systems into a third party's, and that is a policy decision your employer may already have made for you. Business, enterprise and API accounts are not used for model training by default; consumer accounts have a setting that governs it, and it is worth checking rather than assuming. Retention of uploaded files and chat history is a separate question from training, and also worth knowing before you paste a customer list.

The practical middle path most people land on: describe the structure to ChatGPT without the contents. Column names, data types, a couple of invented sample rows. Almost every formula, explanation or translation question can be answered from the schema alone, and the schema is rarely the sensitive part.

Cost, and whether you need both

Both are ongoing subscriptions, and the honest answer is that plenty of people are paying twice without a clear split in their head. Two things worth noting without quoting figures that change:

Copilot is rarely a personal purchase in a business setting. It is an estate decision — a per-user add-on your organisation buys on top of Microsoft 365 — so the question you can actually influence is whether you are on the assignment list, not whether it is worth the price.

ChatGPT is the opposite: an individual decision with a free tier that already handles a large share of formula explanation and drafting. If your spreadsheet work is occasional, the free tier plus good context in your prompts may be all you need.

A reasonable test if you are paying for both and wondering: for two weeks, note which tool you opened for each spreadsheet task and whether the answer required seeing your data. If nearly everything needed your data, Copilot is carrying the work. If nearly nothing did, you are paying for a licence to do something you never ask for.

A routing rule you can remember

1. Does the answer depend on what is in my file? If yes, Copilot — and clean the range into a table first.
2. Does the answer depend on knowing Excel? If yes, ChatGPT — and paste the header row with real column letters so it stops guessing.
3. Is the data sensitive, or the file somewhere I cannot upload from? Copilot if you have it; otherwise describe the structure without the contents.
4. Am I looking at a picture rather than a workbook? ChatGPT, every time. Screenshots are the one input Copilot has no answer for.

And a habit that pays off in both: check the output. A formula written against your real headers still encodes whatever logic the model inferred from an ambiguous sentence, and neither tool will tell you it misunderstood the question. Spot-check the top and bottom rows, and total the result against something you already trust.

Working on your phone, or from a screenshot? XLsheetAI generates Excel formulas from a plain-English description, explains a formula someone else wrote line by line, and converts a photo of a table into a real spreadsheet — no Microsoft 365 licence required.

Bottom line

Copilot's advantage is sight and hands: it can see your columns and change your file, which makes it the right choice for anything shaped like "do this to my sheet". ChatGPT's advantage is that it is not tied to a file at all, which makes it better at explaining, drafting, translating and reading a screenshot — as long as you feed it the layout it cannot see. Neither replaces knowing what the right answer should look like, and both will produce a confident wrong number if you ask an ambiguous question.

Related reading: How to create a PivotTable in Excel is worth knowing by hand before you let anything generate one for you, and Excel to DAX covers the translation job that general models handle best.