One message per panel; six panels at most
◈ 6 cardsA dashboard is the deployment phase's final report on one screen. It is designed from the analytic question outward: the question as its headline, three KPI tiles (total 6,140 · 12 orders · average 511.67, each against a target), one comparison (a sorted bar of average by province), one trend (the monthly line), one detail table (the province × channel pivot) and one filter (a channel slicer). Six panels at most; the eye reads top-left first; every panel carries exactly one message.
Start from the question, not from the data
Module 1 left Renata Kowalczyk with an analytic question — does average order value differ by province, by $100 or more? — and Modules 2 to 13 built the tools to answer it. The dashboard is where the answer lives: CRISP-DM's deployment phase, task produce the final report, on one screen instead of ten pages. That makes the first design rule easy to state and easy to break: a dashboard is graded on whether it answers the stated question, not on how many charts it holds, what colours it uses or whether the tiles animate.
So the question goes at the top, in words, as the headline. Everything below it is there because it helps answer it, and anything that does not help is left out.
Worked example — the Maple & Birch Q1 dashboard
Five kinds of panel are enough for almost any question, and the Q1 dashboard uses each once:
- KPI tiles — one number each, with its comparison: Total sales 6,000 (+2.3 %) · Orders 12 vs 10 · Average order 500 (+2.3 %). A tile with no target or no last-period figure is a number without a message.
- One comparison — the question is by province, so the centre panel is a bar chart of average amount by province, sorted: AB 1,150.00 · ON 498.75 · BC 453.33 · QC 161.67 (Module 8: categories sort by value, not alphabetically).
- One trend — a line of monthly sales: Jan 1,610 · Feb 2,560 · Mar 1,970 — context for whether Q1 is typical.
- One detail table — the province × channel pivot from Module 7, for the reader who wants the number behind the bar.
- One filter — a channel slicer, so Online and Store can be asked the same question separately.
Three tiles, a bar, a line, a table, a slicer: six panels plus a control. Lay them out in the order the eye travels — top-left first, then across, then down. The KPI row is the top; the comparison that answers the question is top-left of the body; the trend sits beside it; the detail table is at the bottom where only the curious go. Under each panel, one sentence: "AB averages $651 above ON — on two orders." That sentence is the panel's message, and the test of a panel is whether it has one. A panel with two messages is two panels; a panel with none is deleted.
The counter-example — "shows everything"
A second draft has fourteen charts: amount by province, by channel, by month, by province and channel, units by everything, a pie of the twelve orders, a histogram, a scatter of units against amount. Every one is a correct chart from Module 8. Together they answer no question, because the reader must work out which one matters, and the reader is Renata, with four minutes before her next meeting. Past about six panels a dashboard stops being read and starts being scrolled. Thoroughness is a property of the data preparation behind the screen, not of the screen.
Assign panels, then write the message lines
The figure is the panel grid. The questions below hand you panels and ask where they go and what they say.
CRISP-DM: deployment → produce final report — but the design decision (which comparison, which KPI) was made in business understanding when the question was written, which is why a dashboard built before the question is written never quite fits it.