Back to course overview
Module 3AI at work 12 min

Data analysis basics

Using AI on tables and numbers without a data team: safe patterns for summarizing, categorizing, and querying — and the arithmetic trap.

You don't need SQL or Python to get analytical help from AI — but numbers are where the green and yellow zones sit closest together, so you need the safe patterns.

What works well on pasted data

  • Describe and summarize: paste a table (CSV text is fine) → 'what stands out? top movers, outliers, anything odd.' Great first pass before a meeting.
  • Categorize free text at scale: 200 survey comments → 'assign each to one of these five themes; return a count per theme and three representative quotes each.' This is hours of intern work in a minute.
  • Explain someone else's spreadsheet: paste the formula → 'explain what this does in plain English and when it would break.'
  • Draft the formula or query you can't write: 'Excel formula: average of column C for rows where column A says Northeast and date in column B is within last 30 days.'

The arithmetic trap

A language model predicts text — it doesn't natively compute. Ask it to total 40 numbers in a pasted table and it will produce a confident sum that is sometimes right. Two escapes: (1) many tools can now run actual code on your data (look for 'analysis', 'code interpreter', or an attach-file feature) — sums from executed code are real sums; (2) keep arithmetic in your spreadsheet and use AI for what it's good at: which calculation, and what the result means.

To get a CSV out of Excel: File ▸ Save As ▸ CSV. If you're in a people-facing role, substitute anonymized data — headcount by department, PTO totals by month, survey comments with names removed. And if you don't have safe data handy, practice on this invented sample:

sample-returns.csvtext
category,units_returned,top_reason
Bedding,42,damaged in transit
Kitchenware,105,damaged in transit
Lighting,18,wrong item shipped
Rugs,27,color not as pictured
Furniture,61,damaged in transit
Curtains,12,changed mind
Tableware,33,arrived late
Prompt to try

Here are last month's product returns as CSV: [paste]. Without doing any arithmetic yourself, tell me: (1) what patterns you notice qualitatively, (2) which three calculations I should run in my spreadsheet to confirm them, with the exact formulas, (3) what result would confirm or kill each hypothesis.

This split — model proposes, spreadsheet computes, you interpret — is the reliable division of labor for non-analysts. Sam used it on returns data and found the 'damaged in transit' spike two weeks before finance did — that monthly returns review became the second of Sam's standing AI-assisted workflows, alongside the Monday ops summary.

Check the row count first

When you paste a big table, first ask: "How many data rows did you receive?" Long pastes get silently truncated, and every conclusion after that is about a table you didn't send. Ten seconds, saves an embarrassing meeting.

Where this leads

When you want governed, always-correct numbers conversationally — 'what was Northeast revenue last quarter?' answered from the actual warehouse — that's a semantic layer + agent. Building exactly that is what the Conversational Analytics Agent course is about, with products like Vigil and Meld keeping the underlying data trustworthy. Foundations gets you fluent; that course makes it production-grade.