AI Data Analysis: A Practical Guide to ChatGPT, Claude and Excel

What is AI data analysis and what does it change?

AI data analysis means uploading your Excel or CSV file to a tool like ChatGPT or Claude and asking questions about it in plain language. You type "which product peaked in which month and how did revenue shift by channel" and the tool runs the numbers, summarizes the findings and draws the chart. A large share of the work that used to require pivot tables, VLOOKUP and nested formulas now takes one well-formed question.

I am not writing this as a distant observer. Since my years working with world-renowned luxury brands I have run campaign reports, content performance tables and budget breakdowns exactly this way, and in my training sessions this is the topic that convinces people fastest. A participant uploads their own real file and watches a breakdown that used to eat half a day appear on screen in five minutes.

One thing to set straight from the start: AI does not make analysts obsolete. It empowers people who never got along with spreadsheets and takes the repetitive work off the analyst's desk. Asking the right question, interpreting the output and tying it to a decision is still your job. In this guide I will walk through the ChatGPT workflow step by step, the AI features inside Excel, how to pick the right tool and the privacy rules that keep you safe.

How to analyze data with ChatGPT step by step

ChatGPT's data analysis feature runs real Python code on the file you upload. That detail matters: the tool does not guess the answer, it computes it and can show you the code it ran. The workflow looks like this:

1. Upload the file. Excel (.xlsx), CSV and PDFs containing tables are supported. Use the attachment icon in the chat box.
2. Describe the data. "This is my monthly sales data for 2025. Columns: date, product, units, unit price, channel. Amounts include VAT." Introducing the columns visibly improves the quality of the output.
3. Ask a concrete question. Skip "analyze this". Ask "build a table of monthly revenue by channel and separately list the channels that declined over the last three months".
4. Request charts. Say "show monthly revenue as a line chart" and download the image for your presentation.
5. Go deeper. The first answer is usually shallow. Follow up with "break this down by city", "compare weekday and weekend sales", "test for seasonality". Analysis is a conversation, not a single transaction.

My own closing step never changes: once the analysis is done I ask for "an executive summary in five bullets, each containing one number". Turning the table into a report is where most of the time saving lives. Because the way you phrase requests directly shapes the result, I recommend reading my prompt engineering guide alongside this article.

AI features inside Excel: Copilot and the alternatives

If you would rather keep the file where it is, Excel has AI of its own. Three options that people tend to mix up:

• Microsoft 365 Copilot: works in natural language inside Excel. It adds a margin column when you ask, builds pivots by region, suggests formulas and draws charts. It requires an add-on license on top of a business Microsoft 365 subscription.
• Analyze Data: the free feature that has been in Excel for years. Select your table and it proposes trends and summary charts in one click. Do not expect depth, but it works as a quick first look.
• Having ChatGPT write your formulas: the route that needs no extra license. Ask "write the Excel formula that pulls the phone number from sheet B matching the name in column A" and you get the formula with an explanation. In my trainings this is the moment people say they wish they had known earlier.

The split I use in practice: if the data stays in Excel and the company already pays for Copilot, use Copilot. If the analysis needs depth and flexibility, move the file to ChatGPT or Claude. They are not rivals, they are two stops on the same workflow.

Claude and Gemini for data analysis: when to pick which

ChatGPT is not the only option. Claude shines with long context: long reports mixing text and tables, comparing several files side by side and turning the findings into a well-written executive summary. My own pattern is ChatGPT for the numeric breakdown and Claude for shaping the findings into a report or a deck.

Gemini makes sense if you live on the Google side, since it works inside Google Sheets and helps you without leaving the spreadsheet. I compared the strengths and weaknesses of all three in detailed tables in my AI tools comparison, so if you are stuck on tool choice, continue there.

Which tool for which data? A comparison table

The last row matters. If you catch yourself uploading the same report to a chat window every single week, that is the signal to move to automation. That setup is beyond the scope of this article, but recognizing the signal alone will save you months.

Data privacy and security: your red lines

Do not skip this section. The convenience of uploading a file hides a responsibility: whatever is in that file now sits on another company's server. The rules I apply myself:

• Remove or mask personal data columns before uploading. Names, phone numbers, emails and addresses rarely matter for the analysis. A "Customer 1" label does the same job and cuts the compliance risk at the source.
• Turn off the use of your chat history for model training in the settings. ChatGPT and Claude both offer this control.
• If you work with company data regularly, look at business plans. Team and Enterprise tiers put the no-training commitment into a contract.
• For trade secrets such as costs, supplier prices or salaries, clarify your company's AI policy before uploading anything. If there is no policy, ask for one to be written.

A pattern I keep observing: the risk rarely comes from the tools being insecure. It comes from users uploading files without a second of thought. A five-minute anonymization habit removes most of the risk.

5 common mistakes in AI data analysis

1. Uploading messy data. Merged cells, two date formats in one column, subtotal rows mixed into the records. However capable the tool, garbage in still means garbage out. Spend five minutes cleaning first.
2. Presenting results without verification. Tools that execute code rarely miscalculate, but rarely is not never. Cross-check against a total you already know and question any mismatch.
3. Not describing the columns. A model that does not know whether "amount" includes VAT can build the wrong story. Context takes thirty seconds to give.
4. Stopping at the first answer. The first response is almost always surface level. Ask for breakdowns, ask for comparisons, ask "is there an alternative explanation that could undermine this finding".
5. Using generated charts as they come. The chart is a draft. Fix the title, the axis labels and the colors to match your brand before it goes into a deck.

Where should you start?

My suggestion is a small three-step experiment:

1. Pick a real file whose numbers you already know. Last month's sales report is ideal because you can verify the output against your own figures.
2. Upload it to ChatGPT, describe the columns and ask three questions you already know the answers to. When the answers match your numbers, trust settles in.
3. Hand one report from your weekly routine over to this workflow and note the time you save. My observation is that reporting time drops by more than half within the first week.

If you want to see where else your business can gain, my AI for small businesses guide maps the wider picture. And if you prefer to build this skill on solid foundations, my AI training has a dedicated data analysis module where we work on your own files in live practice sessions.