Using AI.

How to use LLMs in data-analysis without giving up accountability.

The common failure is not technical. It is epistemic.

XP Lab’s analytical challenges are deliberately tool-agnostic: use search, docs, ChatGPT, Claude, or your notes. Just keep ownership of the thinking.

From a weak prompt to a strong one

The same SQL question framed two ways. Same effort, very different output.

Weak prompt

Write me a SQL query that finds duplicate orders.
  • No schema — the model invents column names that may not exist.
  • No definition of “duplicate” — the answer fits the model’s assumption, not your business.
  • No success criteria — you accept whatever you get back.

Strong prompt

Here’s our orders schema [paste]. Some duplicates are legit re-orders; others are data-entry errors. Suggest a query that flags both kinds, and call out anything you had to assume.
  • Real schema — no invented columns to debug later.
  • Defines “duplicate” — the answer matches your domain.
  • Asks for assumptions — easy to verify the reasoning.

Three principles to keep

These are the habits that make AI assistance productive instead of corrosive.

Use the tool

Ask better questions, generate options, and move faster through blank-page moments instead of pretending you work without help.

Interrogate the output

Check whether the query, explanation, or recommendation actually fits the data, the business question, and the constraints you were given.

Own the final answer

You are still responsible for the tradeoffs, the validations, and the explanation that goes into the submission.

What goes wrong when judgment is outsourced

  • Fast answers feel true

    A confident answer can feel persuasive before you have tested whether it is correct, complete, or relevant.

  • Weak taste stays hidden

    If you never challenge the output, you do not build the feel for what strong analysis and weak analysis look like.

  • Missing context compounds

    AI tools can miss business nuance, data caveats, and implied constraints unless you surface them explicitly.

  • The explanation still lands on you

    In interviews and at work, you are the one who has to defend the answer when someone asks why it should be trusted.