Asking your data
The AI-assisted analysis loop: plan with the model, compute where numbers are checkable, and never let a statistic into your findings without a verified source.
Analysts now have three ways to put AI on data, and knowing which one you're using — and its failure mode — matters more than any prompt template:
- Paste-and-ask — drop a table into the chat and ask questions. Great for small summaries (dozens to a few hundred rows), instant, zero setup. Failure mode: on arithmetic across many rows, models make confident small errors — a sum that's 2% off, a miscounted group. Anything that will be quoted upward gets recomputed in a spreadsheet.
- AI-with-tools — assistants that run actual code or queries on your file (analysis/code-execution modes, or an analytics agent wired to a governed semantic layer like the one the Conversational Analytics Agent course builds). The numbers are computed, not guessed — verify the logic (did it filter the way you meant?) rather than the arithmetic. You won't need this mode in this course — it's here so you recognize it; every lab in this course runs in a chat window plus your spreadsheet.
- AI-as-planner — the underrated one. Don't give AI the data at all; give it the question and ask for the analysis plan: what cuts, what comparisons, what would confirm or kill each hypothesis. Then you execute in your spreadsheet. Best thinking-per-token in the whole toolkit.
The loop that keeps you honest
I'm investigating [question]. Here's my hypothesis: [H]. Here's what data I have: [describe tables, columns, time range - no data yet]. Design the analysis: (1) the 3-5 cuts of the data that would confirm or kill the hypothesis, in the order I should run them; (2) for each cut, what result would support H and what result would kill it - committed BEFORE we look; (3) what alternative explanations could produce the same pattern, and what cut would distinguish them.
Committing 'what would kill it' before looking is the cheapest anti-bias device in analysis. It turns 'find support for my theory' into 'run the test' — and AI is a genuinely good test-designer when you ask before showing it the data.
Then the loop: run a cut → paste the result table back → ask 'what do you see, what would you check next?' → verify any number that surprises you → next cut. AI accelerates every step; the spreadsheet remains the source of truth; and your notes record which numbers came from where. That last habit has a name in Module 2's third lesson — the chain of custody — and it's what makes your findings survive hostile questions.
The subtlest analysis errors aren't arithmetic — they're grain errors (a concept Data Foundations drills): averaging weekly rates as if weeks were equal-sized, counting refund tickets when the question is refund dollars, comparing 'orders' from one table with 'order lines' from another. Before any comparison, say out loud what one row means in each table. AI makes grain mistakes fluently; so do dashboards; the analyst who checks grain wins arguments against both.