ChatGPT can open an Excel file, clean it, compute metrics and draw charts in a few minutes. Getting numbers you can put in front of a client or a board takes more discipline: preparing the Excel file before upload, asking in the right order and checking every figure before it leaves the conversation.
Key Takeaways
- A clean Excel file, with one table per sheet and explicit headers, gives reliable analysis from the very first request.
- ChatGPT writes and runs Python code for each calculation, and asking to see that code is the fastest way to audit its logic.
- No number leaves the chat without a spot check in Excel on two or three values.
Have an AI Sum Up This Article
ChatGPTPrepare the Excel file before ChatGPT ever sees it
Most analysis errors start before the first question. A workbook designed for human eyes, with merged cells, three rows of decorative titles and subtotals sprinkled through the data, forces ChatGPT to guess where the actual table begins, and a wrong guess at that stage contaminates everything that follows.
The ground rule fits on one line: one table per sheet, one header row, one row per record. Delete subtotal rows, unmerge cells, and give every column a name that makes sense without context, such as “revenue_net_usd” instead of “Amount”.
A short legend sheet helps too. A few lines explaining what each column means, where the export came from and when it was pulled give ChatGPT the context a new hire would ask for, and cut down on creative interpretation when a question turns ambiguous.
Regional formats deserve a sentence in your first request. Decimal commas, day-first dates and thousands separated by spaces are common in exports from European subsidiaries, and a tool that assumes US conventions can misread them without complaint. Stating the format up front removes the ambiguity.
Size is the next constraint. OpenAI puts the ceiling for a spreadsheet or CSV at roughly 50 MB, depending on how heavy each row is, well below the general 512 MB limit for other files. An export that is too large can be split by period or business unit, or trimmed to the columns your question actually needs.
Then comes sensitive data. Before uploading an Excel file with customer names, salaries or contact details, replace those columns with neutral IDs. A margin or seasonality analysis almost never needs to know who the customer is, only that two rows belong to the same one.
Your account settings matter as well. On an individual plan, the “Improve the model for everyone” toggle under Data Controls decides whether your conversations are used for training. Temporary Chat stays out of your history, does not use or create memories and is not used for training, although OpenAI keeps it for 30 days for safety reasons before deleting it.
On business plans the default is different. OpenAI says it does not train on inputs or outputs from ChatGPT Business, Enterprise or the API by default, a point we covered when looking at the new Premium seat for ChatGPT Business. For teams handling client data every day, that default is often what settles the question of which plan to use.
Finally, think about where the file comes from. An Excel file can be dropped straight into the conversation or added from Google Drive or Microsoft OneDrive, which avoids scattering local copies of a document that already lives in your team’s shared space.
Whatever the route, keep an untouched copy of the original Excel file next to the one you upload. Every check in the last part of this guide compares ChatGPT’s output with that original, so it needs to stay exactly as it came out of your system.
Four requests that make an Excel file talk
Once the Excel file is uploaded, ChatGPT works on it through code. It writes and runs Python, mostly with the pandas library, in a notebook-style environment that remembers earlier steps in the conversation. That mechanism explains both its power and its traps.
The first request checks that the file was read correctly. Ask for a description of the structure: number of rows and columns, the type detected for each column, missing values and possible duplicates. If the row count does not match what you see in Excel, stop there and fix the import.
This inventory tends to surface the format problems mentioned earlier. A revenue column detected as text, dates read as plain numbers, a sheet that was skipped entirely: all of it shows up here, before a wrong calculation gets built on top of it.
The second request handles cleaning, and it has to be explicit. Spell out what should happen to duplicates, empty rows and outliers, then ask ChatGPT to list everything it removed or changed. Silent cleaning is the hardest source of discrepancy to trace after the fact, because nothing in the final table tells you it happened.
The third request produces the metrics. Phrase it the way you would brief an analyst: “net revenue by month and by region, in dollars, over the last twelve months, with the change versus the same month a year earlier.” The tighter the definition, the less the tool has to interpret.
Results come back as interactive tables you can scroll directly in the conversation. For a cross-analysis, ask for an explicit pivot table and name the rows, the columns and the aggregate you want, whether that is a sum, an average or a count.
The fourth request covers visuals. ChatGPT produces interactive bar, line, pie and scatter charts that you can hover over, customize and recolor without leaving the chat. Ask for one chart per message you want to deliver, with a title that states the finding rather than the metric.
Check the axis labels and units before you export anything. A chart that mixes thousands and millions, or leaves the currency unstated, will be misread in a meeting no matter how accurate the underlying numbers are.
Here is what the sequence looks like on a typical sales export. The inventory confirms the row count and flags a date column read as text. The cleaning step converts it and drops a handful of test orders. The metrics step builds monthly revenue by region, and the chart step shows one region falling behind in the spring. The readout then asks the sales team whether that dip matches a known event, which is the question the data alone cannot answer.
Close with a plain-language readout: three findings backed by numbers, one anomaly worth digging into, one question for the business team. That summary is what saves time in the next meeting, as long as it rests on numbers you have verified.
It also pays to keep the four requests in the same conversation. Because the environment remembers earlier steps, the cleaning rules from the second request apply automatically to the metrics in the third, which is exactly what you want and exactly why a mistake early on travels all the way to the chart.
For the settings that stop you from re-explaining the same context before every analysis, such as permanent instructions and choosing the right model for the task, our ChatGPT guide for professional use walks through the setup step by step.
More articles on Horizon
- OpenAI Models Tried to Hide Their Mistakes
- DeepMind Institute Tackles the Rise of Human-Level AI
- Review a Contract With Claude Without Leaking Data
Check every number before it leaves the chat
ChatGPT can get an Excel file wrong, and the costliest errors are the quiet ones: a filter applied one step too early, a join that duplicates rows, an average computed over empty cells. The output looks tidy, and that is precisely what makes it risky.
The first habit is to ask for the code. Every calculation maps to a few lines of Python that ChatGPT can display and comment. Even without a developer background, a misplaced condition, the wrong column or a filter on the wrong period is easy to spot once it is written out.
The second habit is the control total. Ask for the sum of a key column before and after each transformation, and compare it with the total Excel gives you on the original Excel file. A gap means a row was lost or duplicated somewhere, and you find it without rereading the whole analysis.
The third habit is the spot check. Pick two or three values from the final output, one month, one region, one product, and redo the calculation by hand in Excel with a filter. If all three match, the odds of a structural error drop sharply.
When your Excel file combines several sheets, also check the row count after every merge. A badly chosen shared key, for instance a product name spelled two different ways, is enough to duplicate or drop records without a single error message.
Watch out for answers that go beyond the data. Ask why sales dropped in March and ChatGPT may offer plausible explanations the Excel file simply does not contain. Require it to separate what the data shows from what is a hypothesis, and to label the second as such.
A simple way to enforce that is to ask ChatGPT to write down its assumptions before each calculation: which rows it kept, how it treated blanks, which date column it used. Written assumptions turn a black box into a list you can approve or correct in thirty seconds.
On long analyses with many back-and-forth turns, rerun the inventory from the first request every so often. If ChatGPT says it had to reload the file, that check confirms your earlier transformations are still applied before you go further.
To deliver, ask for an export instead of copying and pasting. ChatGPT can produce a new Excel file or a CSV holding the final tables, which removes transcription errors and leaves a trace a colleague can reuse.
Keep a record of the method as well. Saving the four requests, the code behind the key calculations and your control totals in a note lets you rerun the same analysis next month on a fresh export in a few minutes, with the same safeguards in place.
Used this way, ChatGPT becomes a fast analyst whose work you review. The time savings stay large, and responsibility for the final figure stays where it belongs, with the person presenting it.
Some work should stay in Excel regardless. Regulated reporting, audited figures and anything a finance team signs off on need a calculation chain that lives in the workbook itself, where formulas can be traced cell by cell. ChatGPT is a strong way to explore an Excel file and draft findings, and the official version of the number still belongs in the spreadsheet.
Test Yourself Before Your Next Excel File
What should your first request be after uploading an Excel file?
Show answer
A description of the structure: rows, columns, detected types, missing values and duplicates. If the row count does not match Excel, fix the import before running any calculation.
How do you check that no rows were lost during the analysis?
Show answer
With a control total: the sum of a key column before and after each transformation, compared with the total Excel gives on the original file.
How large can a spreadsheet be in ChatGPT?
Show answer
Roughly 50 MB for an Excel file or a CSV, depending on the size of each row. Beyond that, split the export by period or keep only the columns you need.
Follow the story on Horizon.
