AI Data Analysis: Extracting Insights

Why Prompting for Data Analysis Is Different

When you ask an AI to write a poem or draft an email, a slightly vague prompt still produces something usable. Data analysis is not forgiving. If you ask for "a summary of our sales," the model will guess what you mean by "summary," which columns matter, what time range to use, and what level of detail you expect. The result is often a generic paragraph that sounds plausible but tells you nothing you didn't already know.

The real problem is that AI models are inference engines, not database queries. They fill gaps in your instructions with their own assumptions. In data analysis, those assumptions are where errors hide. A prompt that works for a marketing blog post will produce misleading numbers or false trends when applied to a spreadsheet. This chapter teaches you how to structure prompts so the model acts like a disciplined analyst, not a confident guesser.

The Core Structure of an Analysis Prompt

Every strong data-analysis prompt contains five elements. Miss any one, and you risk a vague or incorrect answer.

• Context: What is the dataset? What does each row represent? What are the column names and their meaning?
• Task: What specific analysis do you want? (Summary, trend, comparison, recommendation)
• Scope: What subset of the data? (Time period, region, product category)
• Output format: How should the answer be structured? (Table, bullet points, paragraph, ranked list)
• Constraints: What should the model avoid? (Averaging without context, ignoring outliers, making causal claims)

Here is a weak prompt and why it fails:

Weak: "Analyze our sales data and tell me what's happening."

The model does not know your columns, your time range, or what "happening" means. It will invent a narrative. Now here is the same request, properly structured:

Strong: "I have a CSV file named sales_2024.csv with columns: date, region, product_category, units_sold, revenue, cost. Each row is one transaction. Analyze the data for Q1 2024 (January 1 to March 31). Provide: (1) total revenue and units sold per region, (2) the month-over-month percentage change in revenue, (3) the top 3 product categories by revenue, and (4) any notable anomalies (e.g., days with zero sales or extreme spikes). Present results as a table with clear column headers. Do not calculate averages across regions unless you also show the standard deviation."

Notice the difference. The strong prompt gives the model a map of the data, a precise task list, and explicit guardrails. The model can now act like a competent junior analyst rather than a fortune teller.

Worked Example: Sales Data Analysis

Let's walk through a realistic scenario. You work for a retail company, and you have a file called sales_2024.csv with the following columns:

• date — transaction date (YYYY-MM-DD)
• region — one of North, South, East, West
• product_category — Electronics, Apparel, Home, Grocery
• units_sold — integer
• revenue — USD, float
• cost — USD, float

You want to know which region underperformed in Q2 and why. Here is the full prompt you would paste into ChatGPT, Claude, or Gemini (all three handle this well):

You are a data analyst. I will give you a description of a dataset, and you will answer based on the data I provide in a follow-up message.

Dataset: sales_2024.csv
Columns: date (YYYY-MM-DD), region (North/South/East/West), product_category (Electronics/Apparel/Home/Grocery), units_sold (integer), revenue (USD float), cost (USD float).
Each row = one transaction.

Task: For Q2 2024 (April 1 to June 30):
1. Calculate total revenue and total units sold per region.
2. Calculate the percentage change in revenue from Q1 to Q2 for each region.
3. Identify the region with the lowest Q2 revenue.
4. For that region, break down revenue by product category and by month.
5. Give two possible reasons for the decline, based only on the data (e.g., a specific category dropped, or a specific month collapsed). Do not speculate about external factors like weather or marketing campaigns unless the data supports it.

Output format: Use a markdown table for step 1 and step 2. Use bullet points for steps 3–5. Round all currency to the nearest dollar.

Constraints: Do not use average revenue per transaction as a primary metric. If you use it, also report the median. Flag any month where revenue is more than 30% below the Q2 monthly average.

Now paste your actual CSV data into the chat (or upload the file if the tool supports it). The model will follow your instructions step by step. If you have a large file, you can ask the model to write a Python script using pandas to do the calculation, then interpret the output. Here is an example of that script request:

Write a Python script using pandas that reads sales_2024.csv, filters for Q2 2024, and outputs:
- revenue and units per region
- Q1-to-Q2 revenue change per region
- the lowest-revenue region
- that region's revenue by category and by month
Print results as formatted tables. Use only base pandas, no other libraries.

This approach works because you have defined the exact operations. The model can write the code, and you can run it locally to verify. This is the most reliable way to use AI for analysis — the model does the coding, you do the verification.

Analyzing Survey Responses: A Second Example

Surveys are different from sales data because the values are often text, not numbers. Here is a prompt for analyzing open-ended customer feedback:

I have a CSV file named survey_responses.csv with columns: respondent_id, age_group, satisfaction_score (1-5), comments (free text). Each row is one respondent.

Task: Analyze the comments from respondents who gave a satisfaction score of 1 or 2.
1. Group the comments into 5–7 themes (e.g., pricing, shipping, product quality, customer service).
2. For each theme, give the percentage of low-score comments that mention it.
3. Quote one representative comment per theme (verbatim, with respondent_id removed).
4. Rank the themes by frequency.

Output format: A table with columns: Theme, Percentage, Example Quote. Then a short paragraph summarizing the top 3 themes.

Constraints: Do not invent themes. Only use themes that appear in the actual comments. If a comment mentions multiple issues, count it under each relevant theme.

This prompt forces the model to ground its answer in the actual text. The constraint about not inventing themes is critical — without it, the model may produce generic categories like "customer experience" that are too vague to act on.

Step-by-Step Application You Can Follow

Here is a literal sequence you can run right now with any AI chat tool that supports file uploads (ChatGPT Plus, Claude Pro, or Gemini Advanced):

1. Open your AI tool and start a new conversation.
2. Upload your CSV file (or paste the first 50 rows if the file is small).
3. Paste the strong prompt from the sales example above, but adjust the column names to match your file.
4. Before sending, read your prompt aloud. Ask: "If I were a new analyst with no context, would I know exactly what to compute?" If the answer is no, add more detail.
5. Send the prompt. Review the output. Check one number manually — for example, sum the revenue for one region in your spreadsheet and compare it to the model's answer.
6. If the number is wrong, do not re-ask the same question. Instead, say: "The revenue for the North region should be $52,300 based on my manual sum. Recalculate using only rows where region = 'North' and date is between April 1 and June 30." This correction loop is normal.

This verification step is non-negotiable. AI models can make arithmetic errors, especially with large datasets. Always spot-check at least one number.

Common Mistakes

• Asking for "insights" without defining what an insight is. An insight is a non-obvious pattern that suggests an action. If you do not say that, the model will give you a summary of obvious facts.
• Not specifying the time range. The model will assume you mean the entire dataset, which may include irrelevant historical data.
• Trusting the first answer. Always verify one number manually. Models are excellent at language, but arithmetic on large datasets is error-prone.
• Using vague categories. If you ask for "trends," the model may invent categories like "seasonal patterns" without evidence. Force it to use only the columns you provided.
• Ignoring outliers. If you do not mention outliers, the model will often smooth over them. Ask explicitly: "Are there any days with zero sales or revenue spikes above 3 standard deviations from the mean?"

Practice Task

You have a CSV file called customer_reviews.csv with columns: review_id, date, product_name, rating (1–5), review_text (free text). Each row is one customer review.

Write a single prompt that asks the AI to:

1. Identify the top 5 themes mentioned in reviews with a rating of 1 or 2.
2. For each theme, give the percentage of negative reviews that mention it.
3. Quote one representative review per theme (without the review_id).
4. Rank the themes by frequency and suggest one actionable fix per theme, based only on the review text.

Your prompt must include: the dataset description, the exact columns, the task list, the output format (table + bullets), and a constraint that the model must not invent themes. After writing the prompt, paste it into your AI tool with a sample of 20 reviews (you can make them up) and verify that the themes match the actual text. This should take under 15 minutes.