Research Workflows

AI for Data Analysis: Turn a Spreadsheet Into Clear Answers

Use AI for data analysis: frame the question, anonymise your spreadsheet, upload it to ChatGPT, Claude or Gemini, check the maths and write a clear brief.

Time to complete
About 2 hours first run, 30 to 45 minutes after that
Difficulty
Beginner
Tools needed
ChatGPT, Claude or GeminiExcel or Google Sheetsa CSV or Excel export of your data

You can use AI for data analysis without knowing code: upload a clean spreadsheet to ChatGPT, Claude or Gemini, ask a specific question, and let the tool run the calculations and draw the charts. The part most people skip is the part that makes the answers trustworthy: framing the question first, stripping out personal data and recomputing the key numbers yourself.

By the end you'll have a one-page findings brief from your own spreadsheet: three to five clear answers, supporting charts and next actions. It works for sales exports, survey results, expense logs and most other tables.

AI is your analyst's assistant here, not your analyst. It can read a column wrongly, quietly drop rows or describe a trend that isn't there, so every step ends with a checkpoint that catches those problems early.

What You'll Need

  • A spreadsheet export in CSV or Excel format: sales records, order logs, survey responses or transactions.
  • An AI chat tool that can analyse files, such as ChatGPT, Claude or Gemini.
  • Excel or Google Sheets for cleaning the file and checking the maths.
  • One clear decision the data should help you make.

New to spreadsheets and AI? Our guide to ChatGPT for Excel covers formulas and cleaning tricks that pair well with this workflow.

Step 1: Frame the Question Before You Open the File

Vague requests are why AI data analysis disappoints. "Analyse this" gets you a tour of averages. "Which states should we stop offering free delivery to?" gets you an answer you can act on.

Write down the decision, then turn it into two to four specific questions with this prompt.

I run [type of business or role]. I have a spreadsheet of [what the data is, e.g. 6 months of online orders] with these columns: [list column names].

The decision I need to make is: [describe the decision].

Help me turn this into 3 to 4 specific, answerable questions for the data. For each question:
1. Write it in one sentence.
2. Say which columns are needed to answer it.
3. Say what result would change my decision, and what result would not.
4. Flag any question this data cannot answer on its own.

Do not analyse anything yet.

Checkpoint: You have a short list of specific questions, each tied to a decision and to named columns. If a question needs data you don't have, drop it or note it as a gap.

Step 2: Clean and Anonymise the Data

Open the file yourself first. You have two jobs: make it readable, and make it safe.

Make it readable

OpenAI's help pages advise descriptive headers in the first row, one record per row, and no empty rows or multiple tables on one sheet. In practice:

  • Keep one header row with plain names such as "order_date", "state" and "amount_ngn".
  • Use one consistent date format.
  • Make money columns numbers only, without "₦" or commas typed in as text.

Make it safe

Most business data contains personal information: names, phone numbers, emails and addresses. In Nigeria, the Nigeria Data Protection Act 2023 (NDPA) governs how that data is handled. The Act says personal data should be collected for a specific purpose and not used in ways people didn't expect, and it sets conditions for sending data outside Nigeria, which is where most AI providers' servers sit. Fines can reach the higher of ₦10 million or 2 per cent of annual gross revenue for organisations the Nigeria Data Protection Commission treats as being of major importance. This isn't legal advice, but the safe habit is simple: don't upload what you don't need.

For most analysis you need none of the personal columns. Keep the useful parts in a coarser form: state instead of street address, a customer ID such as "C0412" instead of a name or phone number.

Unsure what to remove? Ask AI to plan the cleaning using only your headers:

Here are the column headers from my spreadsheet, with no data: [paste headers].

1. Mark each column as: needed for my questions, personal data, or not needed. My questions are: [paste from Step 1].
2. For each personal-data column, suggest whether to delete it, replace it with an ID, or reduce it (for example, address to state).
3. List any cleaning I should do in Excel or Google Sheets before uploading, such as date formats or text in number columns.

Checkpoint: Your file has one tidy table, plain headers, consistent dates, numeric money columns and no names, phone numbers, emails or street addresses. Save it as a new file such as "orders_clean.csv" and keep the original offline.

Step 3: Upload to ChatGPT, Claude or Gemini

All three major assistants analyse spreadsheets by writing and running code behind the scenes. Here is what their help pages said in mid-November 2026:

Tool What it handles Worth knowing
ChatGPT Excel and CSV; OpenAI says spreadsheets can be up to about 50MB, depending on row size Up to 10 files per conversation. Bar, line, pie and scatter charts are interactive.
Claude CSV and Excel, up to 500MB per file and 20 files per chat Anthropic says Excel uploads need code execution and file creation turned on in settings.
Gemini Most file types, up to 100MB each and 10 files per prompt The Gemini web app can create charts from uploaded spreadsheets and let you edit them.

Free plans have tighter usage caps, and limits change often, so check inside the app. A small shop's six-month export is usually well within them. Any of the three will do; if your file is very large, split it by month.

Start the chat with context, not just the file:

I've attached [file name], an anonymised export of [what it is] covering [date range]. Each row is one [order/response/transaction].

My questions are:
1. [Question 1]
2. [Question 2]
3. [Question 3]

Before answering anything, confirm you can read the file and tell me the number of rows and columns you see.

Checkpoint: The AI reports a row and column count that matches what you see in Excel or Sheets. If it doesn't, stop and find out why before going further.

Step 4: Ask for a Data Dictionary and Sanity Checks

Before any analysis, make the AI show you how it understood each column.

Before analysing, please:
1. Build a data dictionary: for each column, give its name, what you think it means, its data type, the number of blank values and 3 example values.
2. Run sanity checks: date range covered, total of the amount column, number of unique values in key category columns, duplicate rows, negative or zero amounts, and any obviously odd values.
3. List every assumption you are making about the data.
Show the code or method you used for the totals.

Read it carefully. Did it treat "amount" as naira and not kobo? Did it read 03/05/2026 as 3 May or 5 March? Are refunds showing up as negative amounts that will distort the averages?

Checkpoint: The data dictionary matches your understanding of every column, and the total amount matches the total you get with a SUM formula in your own spreadsheet. Correct any misreading in the chat before moving on.

Step 5: Run the Analysis Prompts

Ask your questions one at a time; asking everything at once produces shallow answers. These three prompts cover most business needs.

Trends over time

Using the cleaned data, show how [metric, e.g. total sales in naira and number of orders] changed by [week/month]. Then:
- Describe the overall trend in two sentences.
- Point out the biggest rise and the biggest drop, with dates and figures.
- Say whether any period looks unusual because it has fewer days of data or a one-off event.
Show the table you used.

Segments

Break down [metric] by [segment column, e.g. state, product category or sales channel]. For each segment give the number of orders, total value, average order value and share of the total.
Rank them and tell me which segments together make up 80% of the value. Flag any segment with fewer than 20 orders as too small to draw conclusions from.

Outliers

Find outliers in [column]: values that are unusually high or low compared with the rest. List them with their row details (no personal data). For each, suggest whether it is likely a data error, a genuine one-off or a pattern worth investigating. Recalculate the main averages with and without the outliers.

Checkpoint: Each of your questions has an answer backed by a table you can see, and you know which findings rest on small numbers.

Step 6: Turn the Findings Into Charts

Charts help you spot mistakes as well as explain results. Ask for one simple chart per finding.

Create one chart for each of these findings: [list findings].
Rules:
- Use a line chart for changes over time and a bar chart for comparing categories.
- Title each chart with the finding in plain words, e.g. "Lagos orders grew every month from May to October".
- Label axes with units (₦, number of orders).
- Sort bars from largest to smallest.
- No 3D effects and no pie charts with more than 4 slices.
Give me the data table under each chart so I can rebuild it in Google Sheets.

Checkpoint: Every chart has a plain-English title that states the finding, labelled axes and a source table.

Step 7: Check the Maths Yourself

AI models can get numbers wrong, especially when describing results in words. Our explainer on AI hallucination covers why this happens. The fix is cheap: pick the three to five numbers you will actually quote and recompute them in your spreadsheet.

  • Totals: SUM on the amount column.
  • Counts by segment: a pivot table or COUNTIF.
  • Averages: AVERAGEIF, or total divided by count.
  • Month-on-month change: (this month minus last month) divided by last month.

Then ask the AI to reconcile any differences:

I recomputed some figures in Google Sheets and got different results:
- [Your figure] vs your [figure] for [metric]
Explain the difference step by step. Check for filtered rows, duplicates, date-range differences or rounding. Tell me which figure is correct and why.

Run claims that go beyond your spreadsheet, such as "sales dipped because of fuel prices", through our AI fact-checker workflow before you repeat them.

Checkpoint: Every number you plan to quote matches your own spreadsheet calculation, or you understand and accept the difference.

Step 8: Write the Findings Brief

Finish with a one-page brief readable in two minutes. Our workflow on how to turn research into a clear brief goes deeper, but this prompt gets you a solid draft:

Write a one-page findings brief for [audience, e.g. me and my business partner] using only the verified figures below.
Structure:
1. Headline: the single most important finding in one sentence.
2. Three to five key findings, each with one figure and what it means.
3. What the data cannot tell us.
4. Recommended actions, each linked to a finding.
Use plain British English, no jargon and no figures that are not in this list.
Verified figures: [paste your checked numbers].

Checkpoint: You have a one-page brief where every figure has been checked and every action links back to a finding.

Worked Example: A Port Harcourt Online Homeware Shop

Ebere Nwosu runs an Instagram and WhatsApp homeware shop from Port Harcourt (Ebere, her shop and all figures here are invented for illustration). She delivers nationwide and wants to know whether to keep free delivery outside Rivers State.

Questions (Step 1): "Which states bring in enough value to justify free delivery?" and "Which first purchases lead to repeat orders?"

Cleaning (Step 2): Her sheet had 1,912 rows covering May to October 2026. She replaced phone numbers with customer IDs, cut addresses down to state, deleted names and removed 26 test orders.

Upload and sanity checks (Steps 3 and 4): Claude reported 1,886 rows and a total of about ₦41.3 million, which matched her SUM formula. But the data dictionary showed some dates read in American format, so early May orders appeared under January to April. She fixed it and re-uploaded.

Analysis (Step 5): Rivers, Lagos and the FCT made up 78 per cent of order value. Bayelsa and Delta averaged about ₦11,500 per order, too low for free delivery to pay for itself. One ₦640,000 order from an Abuja hotel inflated the FCT average; without it, FCT looked much like Lagos. Customers whose first purchase was a pot set were the most likely to reorder.

Maths check (Step 7): Her pivot table matched the state totals, but the repeat-customer figure didn't: trailing spaces in some IDs made one customer count as two. Fixed, the repeat rate fell from 31 to 27 per cent.

Brief (Step 8): Keep free delivery for Rivers, Lagos and FCT orders above ₦20,000; charge a flat delivery fee elsewhere; promote pot sets as the first-purchase offer. Total time: about two and a half hours.

Common Mistakes

  • Uploading the raw export. Personal data you don't need adds legal risk and nothing to the analysis. Anonymise first.
  • Skipping the data dictionary. Most bad AI analysis traces back to a misread column, usually dates, currency or refunds.
  • Asking "what's interesting in this data?" You'll get trivia. Ask questions tied to decisions.
  • Trusting small segments. A state with eight orders can show a 200 per cent rise because of one customer.
  • Quoting numbers you haven't recomputed. Check every figure you'll repeat.
  • Confusing correlation with cause. The data shows what happened, not why. Ask customers for the why; our workflow for writing survey questions with AI helps you do that properly.

FAQ

Can ChatGPT analyse Excel files?

Yes. ChatGPT can read Excel and CSV files, run calculations with code and draw charts. OpenAI says spreadsheets can be up to roughly 50MB, depending on row size. Free users get lower usage limits than paying users.

Is it safe to upload business data to AI tools?

It's much safer when you remove what you don't need. Delete personal details, check your plan's data and training settings, and never upload data customers wouldn't expect you to share. In Nigeria, the NDPA 2023 applies to personal data, including transfers abroad.

Which AI is best for data analysis?

For a typical small-business spreadsheet, all three do a good job. Clean data, specific questions and checked maths matter more than the brand.

Do I need to know Excel formulas?

Only a few. SUM, COUNTIF, AVERAGEIF and pivot tables cover Step 7, and AI can explain any formula as you go.

Next Step

Data tells you what your customers do; it rarely tells you who they are or why they buy. Once you have your findings brief, use it as the starting point for a buyer persona built from real customer data. And if analysis keeps slipping down your to-do list, give it a fixed monthly slot with our time blocking with AI workflow.

More in Research Workflows.

More workflows