How to ask your data and get an answer you can use
The difference between a useful answer and a pretty one is usually in the question. Five patterns that work, three that always fail, and what to request alongside so you can trust the number.
After a few weeks with a connected assistant a pattern shows up: bad answers almost always come from questions that had no answer. Not because of a model limitation — because the question named no period, no scope and no criterion.
The three questions that always fail
| Bad question | Why it fails | Version that works |
|---|---|---|
| How are we doing? | Does not say about what, when, or against what | Sales and revenue for the last 7 days versus the previous 7 |
| Which channel is best? | Best by volume, by cost or by return? Different answers | Channels by sales, cost per sale and revenue, last 30 days |
| Why did it drop? | Asks for cause; data answers change | Show what changed between the two weeks by channel and product |
Data does not answer why. It answers what changed. The cause is still interpretation by someone who knows the business.
The five patterns that pay off
- 01Metric + period + comparison. "Revenue this week against the average of the previous four." The comparison is what turns a number into information.
- 02Explicit scope. "Approved sales only, in local currency, excluding renewals." A stated criterion is a checkable criterion.
- 03From aggregate to list. "Now show me the 20 people behind that number." That is where analysis becomes action.
- 04Contrast. "What did buyers do that non-buyers did not?" Difference teaches more than an average.
- 05Time order. "What happened before the cancellation, in order?" Sequence is what reveals a trigger.
The request that should follow every question
Ask for the working. "Also tell me the filter you used and how many rows went in." That changes the nature of the answer: instead of a number to believe, you get a number to check.
It is the habit that builds trust fastest — and finds data errors fastest. In most first-week discrepancies the model was right and the criterion was different: time zone, included statuses, currency, recurrence.
How to ask for what you cannot name
You do not need to know the table or column name. Describe the fact in plain language and let the assistant look: "is there any source here that records when a person opened the checkout?". Discovery is part of its job.
- Start by asking what exists: "which data sources are connected to this project?"
- Then ask for the vocabulary: "which event types appear in the last 30 days, and at what volume?"
- Only then ask the business question. Asking about something that does not exist produces invented answers that look right.