Skip to main content

Data Analysis

All levels
What you'll learn
  • Describe a dataset the way a model needs it — columns, types, and the grain
  • Run the five-step safe workflow: describe, ask, generate, execute yourself, interpret
  • Treat any number the model did not compute as unverified
  • State the four assumptions that silently change a result: nulls, dedup, time zones, currency
  • Keep sensitive data out of tools you have not vetted

You don't need to be a data scientist to get answers from data with AI — but you do need to keep it honest. AI writes the query and explains the result; you confirm the numbers are real.

What it helps with​

  • Write SQL / spreadsheet formulas / pandas from a plain-language question.
  • Explain what a query or result means in business terms.
  • Clean & transform messy data (describe the mess, get steps).
  • Suggest the right chart for the question.
  • Interpret output and propose the next question.

The safe workflow​

Guided walkthrough1 of 5
  1. Columns, types, and the GRAIN — one row equals what, exactly? Don't assume the model can guess your names or that order_id is unique. Most wrong queries are wrong about the grain, not the SQL.

Describe the grain, then ask — SQL with stated assumptions

My table orders has columns: order_id, customer_id, created_at, total_eur,
status. One row per order. Write SQL for monthly revenue from completed orders in
2025, and explain any assumptions.

Hand the output back — interpretation with caveats

Here is the result of that query. Interpret it in business terms in under 120
words, then list anything that could make these numbers misleading — nulls,
duplicate rows, time zone of created_at, mixed currencies, or partial months.
Label each as "checked" or "unverified".
Result:
[paste the rows]

Guardrails​

Never trust numbers it didn't compute

If the AI states a result without running the query on your data, that number may be invented. Run queries yourself; treat any figure you didn't compute as unverified. And remember correlation ≠ causation — ask it to flag associational claims. See Hallucinations.

  • Sanity-check magnitudes — implausible result? investigate, don't publish.
  • State assumptions (nulls, dedup, time zones, currency).
  • Don't paste sensitive/PII data into tools you haven't vetted (Privacy).

Those four assumptions are listed in a specific order because that is roughly how often they bite. Nulls silently shrink a denominator; duplicate rows silently inflate a sum; a timestamp stored in UTC and read as local time moves revenue across a month boundary; and a total column that mixes currencies produces a number that is arithmetically correct and economically meaningless.

Pro tip
  • Ask for the query AND a second query that sanity-checks it — a row count, a date range, a sum you can eyeball. Two cheap queries beat one elegant one you cannot verify.
  • When a result surprises you, suspect the join before you suspect the business. A fan-out join is the most common source of a number that is suddenly 3× too big.
  • For spreadsheet work, ask for the formula plus one worked example cell. If the example is wrong, you have found the bug without touching your real file.

Check yourself​

Check yourself

0/3
  1. What is the single most useful thing to tell a model before asking it for SQL?
  2. The model replies with a monthly revenue figure without you having run anything. What is that number?
  3. Your revenue total suddenly triples after you add a customer table to the query. What should you suspect first?

Takeaways​

Key takeaways
  • Describe columns, types, and the GRAIN first — grain errors beat syntax errors for frequency and cost
  • Always ask for the query's assumptions, and review those before the SQL
  • Run it yourself. A number the model did not compute is unverified by definition
  • Hand the real output back for interpretation — reading numbers is the part AI is reliably good at
  • Nulls, duplicate rows, time zones, and mixed currencies are the four quiet result-changers
  • Sanity-check magnitudes and row counts; suspect the join when a total jumps
  • Keep PII out of tools you have not vetted

Next​