AI ON BUSINESS DATA
How to ask good questions of your data
Name the thing, the measure, the period and the breakdown, and an assistant can answer exactly. Leave them out and it has to guess, or ask.
- "How are we doing?"
- "Show me the numbers"
- "Sales by thing"
- "Compare the regions" (which measure? which period?)
- "Orders by region this month" - thing, measure, period, breakdown
- "Rejected orders this week, by store"
- "Average price of products in Clothing, by brand"
- "Now only North" - a follow-up that narrows
Four parts to a good question
An assistant over business data turns your sentence into a request: which collection, which measure, which period, which breakdown, which filters. The more of those your sentence names, the less it has to guess. "How are we doing?" names none of them. "Orders by region this month" names all four - the thing (orders), the measure (count, implied), the period (this month), the breakdown (by region) - and gets an exact leaderboard back: North 395 · West 247 · East 168 · South 142. You do not need to write like a query. You need to say what you would say to a colleague who was about to pull the numbers for you.
The thing
Orders, products, customers, invoices: the collections the assistant knows. Use the business word; the assistant maps it. If you are not sure what it knows, ask - "what can I ask about?" - and it will list the collections and their fields, which is the vocabulary it is working from. Questions outside that vocabulary cannot be answered, and asking first saves a round of guessing.
The measure
Count, total, average, minimum, maximum. "Orders" alone means a count. "Revenue" means a sum. "Average price" means an average. Say which when it is not obvious, because "sales by region" could be a count of orders or a sum of their value, and the two rank regions differently. If you leave it out, a good assistant picks the obvious one and names it in the summary, so you can correct it in a follow-up.
The period
This month, last week, this quarter, the last ninety days, since the first of January. The assistant resolves these to dates, so plain phrases work. Leaving the period out usually gets you all time, which is rarely the question. For comparisons, say what to compare with: "against last month", "same week last year".
The breakdown
"By region", "by store", "by category and month". One breakdown gives a ranking or a trend; two give a pivot. The breakdown is what turns a number into a picture, and it is the part people most often forget and then ask for next. Asking for it up front saves a step.
| Pattern | Example | Shape you get |
|---|---|---|
| Thing + measure + period | "Orders this month" | A KPI: 952, against the previous period |
| + one breakdown | "Orders by region this month" | A leaderboard: North 395 · West 247 · East 168 · South 142 |
| + two breakdowns | "Orders by region and month" | A pivot with totals |
| + a condition | "Rejected orders this week, by store" | A leaderboard of the 36 |
| A trend | "Orders per week this quarter" | A line chart |
| "Show me" | "Show me the urgent orders" | A table of the 14 |
| A follow-up | "Now only North" · "As a table" | The same question, narrowed or reshaped |
Follow-ups
The second question can be short, because the assistant keeps the first. "Now only North" narrows the last result to one region. "By store" adds a breakdown to it. "As a table" reshapes it without changing it. "Why?" is the one that does not work on its own - an assistant can show you what is behind a number, but you have to say which number: "which stores in North declined most?" The suggested questions under each answer are follow-ups the result makes natural, and clicking one is the fastest way to continue.
When it asks back
"Which field do you mean by growth?" means the question used a word the data does not have. "Did you mean order date or ship date?" means two fields fit. These are not failures; they are the assistant refusing to guess, which is the behaviour you want from anything that produces numbers people act on. Answer the question and the request completes. If it keeps asking, the thing you want is probably not in the collections, and that is worth knowing too. Why grounded answers don't invent numbers explains why asking back is the right design.
A few habits make questions land first time. Use the words the dashboard uses, because those are the field names the assistant was given. Ask one thing per sentence; two questions in one get one of them answered. And when a question is really a comparison, say both sides: "this month against last" rather than "how does this month compare".
Reading the answer
Read the shape first. A KPI means you asked for one number; a leaderboard means one breakdown; a pivot means two. If the shape is not what you expected, your question was read differently from how you meant it, and the summary line will usually say how - "orders counted by order date" - so you can adjust. Then read the summary, then the numbers, then decide whether the next click is a suggested question or a widget on the dashboard. How an answer gets its shape covers the shapes in detail.