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:
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 lateHere 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.
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.
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.