Complete Tutorial on Using ChatGPT for Data Analysis: 6 Steps to Understand Your Excel Data from Scratch in 2026

Complete Tutorial on Using ChatGPT for Data Analysis: 6 Steps to Understand Your Excel Data from Scratch in 2026

Many people actually have a bunch of Excel sheets sitting in their hands, including sales records, accounting details, questionnaire results, and inventory lists. But when they really want to get some clues out of them, they get stuck. If you can't remember all the functions, opening a pivot table is a headache, let alone drawing charts for reports. The good news is, now you don't have to wrestle with formulas yourself. Hand the spreadsheet to ChatGPT and ask questions in plain language. It can help you clean, run statistics, draw charts, and even give conclusions. This tutorial explains the entire process from scratch. Even if you have never learned data analysis, you can follow along.

Why ordinary people can use AI for data analysis

In the past, when doing data analysis, the threshold was stuck at the tools. You needed to be able to write formulas, understand statistical concepts, and be able to use Python or BI software. ChatGPT repackages these links in natural language. You just need to upload the data and describe what you want to know, like chatting with a colleague, and it will help you complete most of the technical work behind the scenes.

According to public explanations, ChatGPT's advanced data analysis capability can directly read uploaded files and run Python code in the background to calculate and draw charts. This means that you don't just see the answer it "guessed," but that it actually calculates your numbers. For people who have no programming foundation, this is equivalent to having an assistant beside you who can write code. You are responsible for asking questions and making judgments, and it is responsible for doing the calculations. Of course, it will make mistakes, so manual verification is always necessary, which will be discussed specifically later.

Before step one: prepare the data cleanly

Spending a few minutes sorting out the data before uploading will make everything much smoother later. The original table usually has some common problems, such as confusing headers, mixed Chinese and English in the same column, merged cells, and blank rows in the middle. AI also gets confused when faced with a messy table, so cleaning it up can significantly improve accuracy.

The basic principle of organizing is to make the table into one row per record and one column per field. Delete the redundant title rows and make sure the first row is the field names. The field names should be as short and clear as possible, such as using words like "date," "amount," and "region." If there are merged cells, unmerge them first and fill them in. The date format should be unified into a single style as much as possible, and the amount column should contain only numbers without unit symbols.

Another very important thing is anonymization. If the table contains sensitive information such as customer names, mobile phone numbers, ID numbers, and home addresses, it is best to delete or replace it with a code name before uploading. Data analysis cares about trends and numbers, and in most cases does not require real personally identifiable information. Develop the habit of anonymization to protect others as well as yourself.

Step 2: Upload the data and confirm that the AI read it correctly

When you're ready, just drag the Excel or CSV file directly into the ChatGPT dialog box. After uploading, don't rush to ask questions; let it confirm the data first. You can ask like this: "Please read this table and tell me how many rows and columns there are, what each column means, and whether the data types are correct."

This step may seem redundant, but it's actually crucial. AI sometimes treats numeric columns as text, or misrecognizes dates, causing problems with subsequent calculations. Let it repeat the data back first, and you'll be able to detect misunderstandings immediately. If it says that a certain column is recognized as text, you can directly ask it to convert it to a numeric value before continuing. After confirming that it is correct, you'll have confidence, and the subsequent analysis will hold up.

Step 3: Describe your analysis goals in plain language

Next comes the real analysis. The core of this step is to turn the vague question in your mind into a goal that the AI can understand. You don't need to understand any terminology; just speak plainly.

For example, if you have a sales report, you can say: "I want to know which month sells the best, which product is the most profitable, and whether there are obvious off-peak and peak seasons." Another example is an accounting sheet, where you can ask: "Which categories do I spend the most on each month? Has there been a trend of increasing spending over the past six months?" Stating clearly what you really care about is much more important than worrying about which method to use. The more specific the goal, the more useful the AI's results will be. If the question is too complicated to ask all at once, you can break it into several smaller questions and ask them separately, which makes things clearer.

Step 4: Let AI give you the analysis approach before it takes action

Of course you can ask for the answer directly, but a safer approach is to let the AI explain its approach first. You can add: "Before you start calculating, please tell me how you plan to analyze it, how many steps it will take, and what indicators you will use."

This has two benefits. First, you can understand its logic. If a certain step is unreasonable, you can correct it on the spot, instead of only realizing the direction was wrong after the calculation is done. Second, it makes the AI more rigorous, because explaining the approach first effectively forces it to think the problem through. When you feel the approach is fine, just say "Let's start with this approach," and it will execute. This method of aligning first and then starting work can avoid a lot of back-and-forth rework.

Step 5: Generate charts to make numbers intuitive

Numbers sitting on the page are often hard to feel; only charts can make the trend clear at a glance. ChatGPT can draw charts directly based on your data. You just need to specify what kind of chart you want to see.

If you want to see trends, have it draw a line chart; if you want to compare sizes, use a bar chart; and if you want to see proportions, use a pie chart. You can say: "Please draw monthly sales as a line chart, with the horizontal axis being the month and the vertical axis being the amount." If you are not satisfied with the chart, for example, the colors aren't good-looking, the title is missing, or the axes are confusing, just keep asking for changes. The generated charts can usually be downloaded as images and placed directly into reports or documents. It should be noted that the chart only visualizes the numbers it calculated, so whether the chart looks good is secondary; whether the underlying numbers are correct is the key.

Step 6: Ask for verification, then draw the conclusion and archive it

Don't stop working immediately after you get the results. Ask a few questions first to nail down the conclusion. You can ask: "What data is this conclusion based on, has it been skewed by extreme values, and would the result change if the method were changed?" This kind of questioning can force the AI to expose weak links in its reasoning, and can also help you judge whether the conclusion is reliable.

After confirming that the conclusion is credible, it's time to archive the entire analysis process. There is a practical little habit here, which is to export the conversation together with the analysis conclusions into a formal document, which makes it convenient to reuse in the future and easy to send to colleagues. This can be done with the help of a browser extension; Save AI is the kind of tool worth trying. It can export conversations from multiple AI sites such as ChatGPT, Claude, Gemini, etc. into PDF, Word, Markdown or long images with one click.

Exporting the complete conversation of a data analysis into a Word or PDF report is equivalent to leaving yourself an archivable, reusable draft of this analysis, which you can refer back to next time you encounter a similar problem. Save AI takes a local-first route. The export process is completed locally, can be used offline, and the data does not go to the cloud. This is actually consistent with the data privacy we emphasized earlier. It is especially reassuring when dealing with sensitive spreadsheets. You can try out which export format is the most convenient.

How to write prompts so that they're actually useful

Many people feel the results AI gives aren't good, and the problem often lies in the way the questions are asked. A good prompt usually contains three elements, namely background, goal, and format requirements. The background is to tell it what the data is for, the goal is to clarify what you want to get, and the format requirement is to specify the form in which the conclusion is presented.

For example, instead of dryly saying "analyze this table," it is better to say: "This is the daily sales flow of a milk tea shop over half a year. I want to find the three best-selling products and the most obvious off-peak and peak seasons. Please summarize the conclusion in one paragraph and attach a line chart." You'll find the latter answer to be much more useful. Another technique is to ask questions step by step. First let it give you a general overview, and then dig deeper into the points of interest. This is much clearer than cramming ten questions in all at once.

Don't blindly trust AI: the results must be checked manually

This section is the most important reminder in the entire tutorial. AI sometimes makes mistakes in its calculations. It may misunderstand the meaning of a certain column, it may make an error in an intermediate step, or it may treat individual abnormal data points as a general rule. What's even more troublesome is that the tone it outputs is often very confident and seems reasonable, making it easy for people to let their guard down.

The verification method is not complicated. For the most critical numbers, you can use a calculator or manually check them in Excel to see if they're correct. If you ask for the total, add it up in the table and compare. If it says a certain month is the highest, just sort to see if that's true. When you encounter a conclusion that obviously conflicts with your common sense, you should stop and ask what it's based on. Remember one principle: AI is responsible for doing the work, you are responsible for making decisions, and it is always the human who is ultimately responsible for the conclusion. Use it as a smart but occasionally careless assistant that saves effort and is safe.

Privacy notice: what data should not be uploaded casually

The other side of convenience is privacy risk, and this line must be held. Before passing data to AI, ask yourself whether there is any information in it that should not leak out. Be extremely cautious when it comes to customer personal information, internal company finances, undisclosed business data, and medical- or legal-related private content.

A practical approach, as mentioned before, is to anonymize it before uploading: delete or replace fields such as name, mobile phone number, and account number with code names, and only keep the numbers and categories that are really needed for analysis. If it is company data, it is best to first confirm whether the company has relevant compliance requirements. The same goes for the export step. Choosing a local-first tool that does not upload data to the cloud to save analysis results can further reduce the possibility of sensitive information leaking. When it comes to privacy, thinking one step further never hurts.

FAQ

Does using ChatGPT for data analysis require programming skills?

No. All you need to do is describe what you want to know in plain language, and it will complete the code part in the background. However, understanding some basic concepts, such as averages, trends, and proportions, will help you ask better questions and judge the results.

How large an Excel table can ChatGPT handle?

Usually ordinary tables with thousands to tens of thousands of rows can be processed, but the specific capability will vary with the version and package. If the table is particularly large, you can first do some summarizing or splitting locally, and then give it the parts you care about for analysis, which is both fast and stable.

Are the results calculated by AI trustworthy?

Don't trust them blindly. Its calculations are generally reliable, but errors do occur, especially when it comes to understanding the meaning of the data and dealing with outliers. You must check the key numbers manually yourself, treat it as an assistant rather than an authority, and the final conclusion is up to you to vet.

How do I save the analysis process and conclusions?

You can use a browser extension to export the entire conversation as a document. For example, Save AI can export conversations from sites such as ChatGPT, Claude, Gemini, etc. into PDF, Word, Markdown, or long images, with a local-first approach and data not uploaded to the cloud, making it suitable for archiving and reusing analysis reports.

Are there privacy risks in uploading data?

Yes, so you must anonymize it first. Delete sensitive fields such as name, phone number, and account number, or replace them with code names, and only keep the content needed for analysis. Be even more cautious with data involving company or other people's privacy, and try to choose local-first tools when exporting as well.

The data has actually always been there; in the past we simply lacked a pair of eyes that could understand it. Now that pair of eyes is within reach, and the real difference may no longer be whether you can use the tools, but whether you are willing to ask one more question and double-check one more time.