How to ask an AI a question about your data (so the answer is usable)

“How many customers do we have?” sounds like a simple question. Give it to an AI assistant with a database and you will get a confident number. Give it to three colleagues and you may need three different numbers: product wants people who bought last month, marketing wants people it can still reach with ads, operations wants everyone who placed an order in August, even if it was later cancelled. On the same data these can be 2,141, 3,670 and 2,339 — that is the whole idea of our game Wrong Number.

An incomplete question leaves room for assumptions. An assistant may fill those gaps silently; it may also make a calculation or SQL error. Agreeing the definition is the first check, not a guarantee that the answer is correct.

Short version: before asking, say which decision the number is for, who or what to count, how to measure it, over which period, and what to exclude. After the answer, ask the AI to restate its definition.

The five-part question

1. Decision. What will you do with the number? (“Decide whether to extend the autumn promo.”) This lets the AI — and you — judge which definition fits.
2. Population. Who or what is counted? Buyers or accounts? Orders or order lines? Including test accounts and staff?
3. Measure. Count, sum, average, median, share? Of which amount — before or after discounts, before or after refunds?
4. Period. Exact dates, time zone, and whether an incomplete last period should be dropped.
5. Exclusions. Cancelled orders, refunds, internal traffic, one-off outliers.

Before and after

Vague:

What was our revenue last month?

Usable:

Decision: did the autumn promo pay off?

Orders: paid only (exclude cancelled); promo and non-promo side by side, with order counts.

Money: what customers paid after platform discounts; completed refunds for those orders separately, with the refund-data cutoff.

Period: 1 September inclusive to 1 October exclusive, order creation time, Europe/Amsterdam; flag any missing day.

Before calculating: restate the definition and ask me if anything is ambiguous.

The second prompt makes the status, amount, period and comparison explicit. It still does not establish whether the promo caused any change or paid for itself: that requires a suitable comparison and cost data.

Questions to ask after the answer

• “State the definition you used in one sentence.”
• “List every filter you applied and every one you considered but didn’t.”
• “Did any join change the number of rows?”
• “Show the numerator and the denominator separately.”
• “What would make this number wrong?”

These are the short form of our 7-question checklist. The AI Answer Checker picks the ones that apply to a specific answer and writes them as one message.

When the AI should ask you

A good assistant asks a clarifying question when the request is ambiguous. If it never does, add one line to your prompt: “If the definition is ambiguous, ask me before calculating.” Check whether the questions actually resolve the ambiguity; an assistant can still miss a necessary assumption.

To build these habits from the beginning, follow the data literacy for AI learning path, with a worked campaign example and practice you can try without an account.

FAQ

What is the best prompt for data analysis with ChatGPT or Claude?

There is no magic wording. What works is a complete question: the decision, the population, the measure, the period and the exclusions — plus a request to restate the definition and flag ambiguity. The prompt only makes your thinking visible; the thinking is the skill.

Why does the AI give different numbers to the same question?

Ambiguous definitions are one possible cause. Different data snapshots, generated queries or calculation errors can also change the answer. Fix the definition and data version, then inspect and reproduce the calculation.

Is this a prompt-engineering skill?

It overlaps with prompt design, but the underlying skill is defining the analysis: what must be decided before a number means anything. It works with a colleague, a dashboard or an AI. Use the metric definition template to make the brief explicit.

How can I practise asking better data questions?

Play Wrong Number (free, three questions): you see one database give three right answers to three versions of the same question. Then try Bigger baskets?: choose clarifying questions, define a basket, and test whether a marketing claim follows from the comparison. This first Atterna Case is free without an account and uses basic SQL.