AI ON BUSINESS DATA
How an answer gets its shape
Ask for one number and get a KPI tile. Ask for a ranking and get a leaderboard. The question decides the format, and the format decides what you see first.
- One numberKPI
- Two dimensionsPivot
- A rankingLeaderboard
- Change, comparison, shareChart
- The records themselvesTable
- Summary the finding in a sentence
- What you could do when the findings call for it
- Ask next the next questions, one click away
The format is part of the answer
Ask a colleague how orders are doing and they might say "952, up on last month" or hand you a table of every order; one of those is the answer and the other is homework. The shape of an answer decides what you see first and whether you have to work to see it. A conversational assistant that returned a paragraph for every question would be a search engine; one that returned a table for every question would be a database. The useful behaviour is to return the shape a person would choose, and the rule for choosing it is the same one dashboard designers use.
Five shapes
KPI. One number and its change against the previous period. "Orders this month" is 952 in a tile, not 952 in a sentence, because the tile carries the comparison and the eye reads it in a glance.
Pivot. Two dimensions and totals. "Orders by region and month" puts regions down the side and months across the top, so North's 395 is visible as 184, 122 and 89, and the totals column makes the regions comparable.
Leaderboard. A ranking with a bar beside each item. "Rank the regions" is North 395 · West 247 · East 168 · South 142, in order, with the gap between them visible.
Chart. Change over time is a line, a comparison is a bar, a share is a stacked bar, a distribution is buckets. The chart type follows the question, which is the subject of Choosing the right chart for business data.
Table. The records themselves, when the question is "show me". The 14 urgent orders, one row each, ready to act on or export.
| You ask for | Shape | Example |
|---|---|---|
| One number, often with its change | KPI | "Orders this month" - 952, against the previous period |
| Something broken down two ways | Pivot | "Orders by region and month" - North's 395 as 184, 122 and 89 |
| The top or bottom of something | Leaderboard | "Rank the regions by orders" - North 395 · West 247 · East 168 · South 142 |
| Change over time, a comparison, a share | Chart | "Orders per week this quarter" - a line |
| The records themselves | Table | "Show me the urgent orders" - 14 rows |
Around the shape
The region pivot shows North far ahead with 395 orders, followed by West with 247, East with 168 and South with 142.
- Check stock across North's stores.
- Review the rejected orders before the weekly call.
Three things frame the shape. A summary states the finding in a sentence - North leads order volume by a wide margin - so that the answer can be read before it is studied. What you could do appears when the findings call for it: check stock across North's stores, review the rejected orders. It is not a to-do list for every answer, only for the ones where the data points somewhere. And Ask next offers the follow-ups that the result makes natural, each one a click, so that the conversation continues from the answer rather than from a blank box.
How the choice is made
The shape is chosen from the structure of the request, not from the wording of the question. A request with one measure and no breakdown is a KPI. One measure and one dimension is a leaderboard if the dimension is categorical and sorted, a chart if the dimension is time. Two dimensions is a pivot. A request for records rather than a measure is a table. Because the request is built from a constrained vocabulary - collections, fields, filters, facets - its structure is known exactly, and the shape follows from it without guesswork. How can AI answer questions about business data? places this step in the full sequence.
Consistency with the dashboard
The rules are the same ones a dashboard designer applies, which means an answer in chat looks like a widget on the dashboard that would answer the same question. That is deliberate. When a question asked three times in chat becomes a widget, the widget is the shape the answers already had, and when a dashboard number needs explaining, the chat answer that explains it is in a familiar form. One visual language across both surfaces is what lets people move between them without relearning anything. Dashboards vs chat describes that handoff.
Shapes also compose. A leaderboard answer can carry a KPI above it - the total the ranking adds up to - and a table can sit under a chart as the records behind it. The assistant adds these when the question implies them, and leaves them out when they would be noise. A ranking of four regions does not need a table of 952 orders beneath it unless someone asks to see them.
When the shape is wrong
Sometimes the person wanted a table and got a leaderboard, or wanted the chart as a bar rather than a line. The fix is a follow-up - "show that as a table" - and a well-built assistant treats it as a re-shaping of the same result, not a new question. Being able to change the shape without re-asking is part of what makes the conversation feel like one, and it is a good test to run in a demo.