There's a pattern I see repeatedly: someone pastes their data into an AI chat and asks "what can you tell me about this?" and gets back a description of the data they just shared.
"You have 247 rows. The average value is 1,432. There appears to be a spike in month 7."
Technically accurate. Almost entirely useless. A description is not an analysis. It's a recap.
Good data analysis answers the question: so what? It connects patterns in data to decisions you need to make. It tells you what the numbers mean, not just what they are.
Getting AI to do that well comes down to how you prompt.
The First Thing to Change: Tell It What Decision You're Facing
Most data prompts look like this: "Analyze this data." Or "What trends do you see?"
The model has no idea what you're trying to decide. So it describes. It looks for patterns. It generates observations. And they're almost never the ones you needed.
Change the framing: tell it the decision.
I'm trying to decide whether to [specific decision]. Here is the data:
[paste data]
Based on this data, what should I know that's relevant to this decision? Don't just describe the data — help me understand what it implies.
Now the model has a direction. It can filter the observations it makes toward the ones that actually matter for your question.
How to Share Data Well
AI can work with data shared in a few different formats. Here's what works and what doesn't.
Works well:
- Small-to-medium CSVs pasted directly as text (up to a few hundred rows)
- JSON payloads
- Summary tables
- Data described in structured paragraphs
- Markdown tables
Works poorly:
- Excel file screenshots (the model can see them but can't calculate from pixel values reliably)
- Very large datasets (paste a representative sample instead)
- Highly encoded or compressed formats
When pasting data, always include the column headers and explain what the columns mean:
Here is monthly sales data for 2025. Columns are:
- Month: calendar month (1-12)
- Revenue: total sales revenue in USD
- Orders: number of individual orders
- Returns: number of items returned
- CAC: customer acquisition cost that month in USD
[paste data]
Don't make the model guess what "col_3" means or whether "rev" is revenue or reviews.
The Insights Prompt vs. The Describe Prompt
The difference between these two prompts produces dramatically different output:
Describe prompt: "What can you tell me from this sales data?"
Insights prompt:
I'm reviewing our Q4 2025 sales performance to prepare for a budget planning meeting.
Here is the data: [paste]
Help me understand:
1. What's the most important thing happening in this data that I need to explain to leadership?
2. What's surprising or unexpected compared to what you'd typically expect?
3. What's missing from this data that I'd want to have before making budget decisions?
4. What follow-up analysis would be most valuable to do?
Same data. Completely different direction.
The multi-question format is deliberate. It forces the model to approach the data from several angles rather than landing on the first observation and stopping.
Asking for Hypotheses, Not Conclusions
One of the most useful applications of AI for data analysis is generating hypotheses — explanations for patterns — that you can then go verify.
Here's a pattern I'm seeing in the data: [describe pattern].
Give me 5 different hypotheses for what might be causing this. For each hypothesis, tell me what additional data would confirm or deny it.
This is particularly useful when you're stuck. You see something anomalous and don't know where to start. The model can rapidly generate explanations from multiple angles — some obvious, some you hadn't considered.
You're not accepting the hypotheses as conclusions. You're using them to structure your investigation.
Prompting for Comparisons That Matter
A lot of data analysis is comparative: this period vs. last period, this segment vs. that one, this channel vs. others. The default AI response to comparison requests tends to list differences. That's rarely what you need.
Push for the so-what:
Compare these two segments of customers: [paste segment A data] vs. [paste segment B data].
Don't just list the differences. Tell me:
1. What's the most significant difference and why does it matter?
2. What does this suggest about how we should treat these segments differently?
3. What's similar that I might have expected to be different?
Or for time-period comparisons:
Here is data from Q3 and Q4 2025: [paste]
I'm not just looking for "Q4 was higher." I want to understand the story — what changed between the quarters, what might have caused those changes, and what the Q4 trend suggests about Q1 2026.
Using AI to Write SQL or Spreadsheet Formulas
If you work with databases or spreadsheets, one of the highest-leverage uses of AI for data analysis is getting it to write the code for you.
For SQL:
I have a database table called "orders" with these columns:
- order_id (integer)
- customer_id (integer)
- order_date (date)
- total_amount (decimal)
- status (text: "completed", "refunded", "pending")
- product_category (text)
Write a SQL query that shows me: for each product category, the total revenue from completed orders in 2025, the number of unique customers, and the refund rate (refunded / total orders). Order by total revenue descending.
I'm using PostgreSQL.
For Excel/Google Sheets:
I have a Google Sheet where:
- Column A: customer names
- Column B: purchase dates (formatted as MM/DD/YYYY)
- Column C: purchase amounts
I want a formula in column D that shows how many days ago each purchase was made. Today's date should be dynamic, not hardcoded.
Also write a formula for a summary cell that shows the total revenue from purchases made in the last 30 days.
For these requests: always specify the database type or spreadsheet application, always describe the columns, and always ask it to explain what the formula does so you can verify the logic.
A Useful Framework: The "So What" Stack
When you have a data finding, push through three levels:
Level 1 — What: What does the data show? (Describe)
Level 2 — Why: Why might this be happening? (Interpret)
Level 3 — So what: What should we do, or investigate, because of this? (Recommend)
Most data questions get answered at Level 1. The decisions live at Level 3. Prompt explicitly for the levels you want:
Looking at this data, [describe what you've already found].
Help me think through:
- Why might this be happening? (Give me 3 possible explanations)
- What does this imply for our [product / marketing / operations] decisions?
- What would you recommend investigating or changing based on this?
What AI Can't Do With Data
Honest limitations:
It can't handle very large datasets. If you have 50,000 rows, paste a sample or share aggregate summaries. The model's context window isn't a substitute for a database.
It can't access real-time data. It works with what you give it. If your data is a month old and you need current insights, you need current data.
Its calculations can have errors. For any critical number — especially if you're presenting to leadership or making financial decisions — verify calculations independently. AI can make arithmetic errors, especially on large or complex calculations.
It doesn't know your business context. It might flag a 15% drop as alarming when for your business that's seasonal and expected. Always add the context that makes your numbers meaningful.
If you're interested in how to chain complex, multi-step analyses together, the Advanced Track covers prompt chaining — breaking analysis workflows into structured sequential steps that each build on the last.