ChatGPT Data Analysis: How to Analyse a Spreadsheet Without Excel Skills

Inside a ChatGPT Data Analysis Spreadsheet Session

When you attach a spreadsheet, ChatGPT does not read it the way it reads a paragraph. It opens a private computing session, writes Python, and runs that code against your file. The numbers come from actual calculation rather than from a language model doing arithmetic in its head — which matters more than almost anything else on this page.

According to OpenAI’s data analysis documentation, that environment is a stateful notebook session — it remembers what it worked out a moment ago, so each question can build on the last. It also carries one hard boundary worth knowing on day one: it cannot reach the internet. If your analysis needs last year’s industry benchmark, you have to upload that too.

What comes back falls into four shapes:

• An interactive table you can scroll, expand and click into
• Charts — static images by default, interactive for bar, line, pie and scatter
• A written summary in whatever format you specified
• Downloadable files, though the table download button gives you a CSV rather than an Excel workbook

Spreadsheets, PDFs and plain text formats are all supported. If your organisation has connectors switched on, you can attach files directly from Google Drive, OneDrive or SharePoint instead of downloading them first.

One capability worth knowing on day one, because it removes a job most people still do by hand: a ChatGPT data analysis spreadsheet session can hold more than one file at a time and join them for you. Upload a customer list and a purchases export that share a customer ID, ask a question that spans both, and it will merge them itself rather than making you build a lookup first. That alone retires a good deal of the formula work this article is meant to replace.

The Sheet-to-Signal Method

The Sheet-to-Signal Method is a three-pass routine for turning a raw spreadsheet into a finding you can defend: Shape the sheet, Interrogate it in layers, then Verify the maths. Each pass exists because something specific goes wrong when you skip it.

Skip Shape and the model misreads your columns, which is how most ChatGPT data analysis spreadsheet sessions quietly go wrong. Skip Interrogate and you get a bland recap instead of an answer. Skip Verify and you end up presenting a figure you cannot source when someone asks. The Sheet-to-Signal Method we teach at PromptPeakAI is deliberately unglamorous in the middle — most of the value sits in the first and last pass, which are precisely the two everybody rushes.

Shape the Sheet Before Upload

Most bad output traces back to the file rather than the prompt. OpenAI’s own guidance is unusually specific here: descriptive column headers in row one, one record per row, plain-language column names. Everything below is in service of those three things.

Delete the title banner sitting above your headers. Remove blank rows and columns that split the data in half. Move any second table onto its own sheet. If a value you need analysed only exists inside a pasted image, retype it as data — images get discarded.

Col C tells the model nothing. actual_spend_gbp tells it the meaning, the unit and the currency in one go. If your file is genuinely chaotic rather than just untidy, our guide to cleaning messy data with AI goes deeper than this section needs to.

Before the file goes anywhere, strip whatever identifies a person or a client. Employee names become E001. Client names become Client A. You lose nothing analytically, because the pattern you are hunting lives in the numbers, not the labels.

You are a data analyst reviewing a file before any analysis begins.

Describe the structure of the attached spreadsheet: how many rows and columns, what each column appears to contain, and what data type each one seems to be.

Then list every structural problem you can see — headers that are not in row 1, blank rows or columns, numbers stored as text, inconsistent date formats, duplicate records.

Do not analyse the data yet. Do not fix anything yet. Just tell me what you found.

Interrogate Your Data in Layers

The single biggest upgrade over typing “analyse this” is asking in sequence. Find out what is broken, then ask the business question, then ask for the shape you need to present. Three prompts, in that order, beat one long one every time.

ChatGPT is good at spotting missing values, outliers, duplicate rows and columns where the data type is wrong — the four faults that quietly distort totals. Ask it to find them and propose options, but hold it back from fixing anything until you have chosen.

Check this data for common issues: missing values, outlier values, duplicate rows, and incorrect data types.

For each issue you find, tell me:
- which column it affects and roughly how many rows
- what my options are for handling it
- which option you would choose, and why

Wait for me to choose before you change anything.

Name the decision the analysis is feeding. “Summarise this data” produces a summary nobody asked for; “which categories went over budget by more than ten percent” produces an answer. The same Role, Goal, Context, Format structure from our Logic System guide applies here, with one addition — an explicit instruction to calculate rather than estimate.

You are a [finance analyst] helping me prepare for [a quarterly budget review with my directors].

Using the attached file, answer this: [which expense categories went over budget by more than 10%, and by how much].

Rules:
- Use Python to calculate every number. Do not estimate anything.
- Use only the [budget_gbp] and [actual_gbp] columns.
- If a value is stored as text, clean it first and tell me that you did.
- Give me a table of results, then three sentences on what this means for [next quarter].

If anything here is ambiguous, ask me before you calculate.

Work inside the table, not the chat

This is the part most guides miss entirely. When your file loads, ChatGPT builds an interactive table you can actually work in — expand it to full screen using the arrows at the top right, then click a column heading to select it and ask a question about that column alone.

Hold Command on a Mac or Control on Windows to select several columns at once, or drag across a block of cells and ask for the average of just those. You can also have it add a calculated column to the table, then use the download button to pull the result out as a CSV.

If what you actually need is the formula rather than the answer, writing Excel formulas with ChatGPT is the better route — you keep working in your own workbook.

The third layer is the one people forget, and it is the cheapest. You already have the finding; now ask for it in the exact format the next person needs, rather than reformatting a wall of prose yourself at eleven at night. Most ChatGPT data analysis spreadsheet work ends here, in a paste-ready block.

What that looks like varies by who is waiting for it:

• Finance — a variance table plus a one-line narrative per overspend
• Sales — top and bottom performers with the month-on-month change
• HR — department-level totals only, with no individual identifiable
• Operations — an exceptions list, ranked, with nothing that is already fine

Reformat the findings above for [a slide I have to present to my directors on Thursday].

Format:
- One headline sentence stating the single most important finding
- A table of no more than [six] rows, sorted by [size of overspend]
- Three bullets on what I should recommend we do about it
- One sentence naming the biggest caveat in this data

Keep it under [150] words in total. Plain business English, no jargon, no filler.

Verify the Maths Before Sending

This is the pass that separates people who use this well from people who get caught out. The model produced code; the code produced numbers; you are about to attach your name to those numbers. Spend ninety seconds finding out where they came from.

The difference is not politeness — it is whether you can answer the obvious follow-up question in the meeting. Numbers stored as text are the classic trap: a column of values like “£1,200” reads as words rather than figures, so a total quietly comes back wrong without anything looking broken. Our guide to why ChatGPT invents things covers the wider pattern, but for spreadsheets specifically, this prompt catches most of it.

Before I use these numbers anywhere, walk me through your working.

For each figure in your summary:
- show me the code you ran to produce it
- tell me which columns and how many rows it used
- list any rows you excluded, and why you excluded them

Then tell me one thing about this dataset that could make your answer misleading.

What to Upload, What to Redact

Straight answer: for most internal ChatGPT data analysis spreadsheet work on a company laptop, an anonymised file is fine. The judgement call is not really about ChatGPT — it is about whether the file identifies a person, and whether your employer has said anything about it.

On a personal Free or Plus account, your conversations may be used to improve OpenAI’s models unless you say otherwise. The switch lives under Settings, then Data Controls, in a toggle called “Improve the model for everyone” — turning it off applies to new conversations from that moment. A temporary chat is the other option when you want a session that leaves nothing behind.

Business and Enterprise workspaces are different: business data is excluded from model training by default, which is why some companies buy a workspace purely to end this conversation. Our fuller take on whether ChatGPT is safe to use at work covers the policy side, including what to do when your employer has said nothing at all.

Plans, Limits and Upload Caps

You do not need a paid plan to do this. You may want one, and the reason is upload allowance rather than capability. For occasional ChatGPT data analysis spreadsheet work — a file or two a week — the free tier genuinely covers it.

The specific numbers matter here. OpenAI’s published file upload limits put spreadsheets at roughly 50MB per file and cap free accounts at three uploads a day, while paid accounts run to eighty files every three hours. Business seats are $20 per user per month billed annually, $25 monthly, with a two-seat minimum. If you are weighing the upgrade on more than this one task, our Free versus Plus comparison works through it properly.

One alternative worth knowing: if your company has blocked external uploads entirely, Copilot inside Excel does a narrower version of this without the file leaving the Microsoft environment. It is less flexible with messy data, but that is sometimes the trade you have to make.

Where This Falls Short

Every guide on this topic gets breathless about the time saved and quiet about the failure modes. ChatGPT data analysis spreadsheet work has four hard limits, and the failure modes are what actually get people into trouble.

Charts have their own limits. ChatGPT will produce a wide range of chart types, but only bar, line, pie and scatter charts become interactive; the rest come back as static images. Downloads are PNG, colours are adjustable including exact hex values, and anything more designed than that is faster to build yourself — our guide to creating charts with AI covers when to stop fighting the default.

When the Analysis Goes Wrong

When a ChatGPT data analysis spreadsheet run goes wrong it usually looks mysterious and turns out to be one of five things. Work down this table before assuming the tool is at fault.

The pattern behind four of those five rows is the same: the file was not shaped before it went up. That is why Shape is the first pass rather than an optional tidy-up, and why the two minutes it costs is the best-spent time in the whole method.

Once a ChatGPT data analysis spreadsheet run feels routine, the natural next step is the full end-to-end version — taking one genuinely messy file from raw export to a finished executive summary. Our walkthrough on analysing a spreadsheet with ChatGPT follows a single dataset the whole way through, and picks up where this one stops.