Directing AI Through an EDA

The one idea: An AI can run a whole exploratory analysis from a single sentence — but it will confidently summarize data whose structure it has misread. Your leverage comes from directing the sequence (grain first, evidence at every step) and reviewing its output the way you'd review a junior analyst's.

⚠️ The trap: "I'll just ask the AI to explore the data and tell me what's interesting." That feels right and it's wrong. Slicing a big dataset hundreds of ways guarantees striking patterns that are pure chance (the vast search effect), delivered with total confidence — bring the hypothesis yourself, and re-test any candidate on holdout data the search never saw.

AI assistants have changed the mechanics of exploratory data analysis (EDA). What used to be an hour of typing groupby and describe is now a sentence: "explore this dataset and tell me what's interesting." The mechanics changed; the judgment didn't. An AI will happily explore a table whose grain it misunderstands, summarize columns full of sentinel values, and report a fascinating "insight" that is really a join bug in the file you gave it.

The analysts who get real leverage from AI don't delegate the whole EDA. They direct it. They sequence the exploration the way a good analyst would, demand verifiable output at each step, and review the AI's work exactly as they'd review a junior analyst's. This lesson gives you the prompt patterns and the review discipline.

Start with the grain interrogation

Everything in this track has said grain comes first. That applies doubly when an AI is doing the looking: an AI that assumes the wrong…