In short
A dashboard should answer a specific question for a specific person at a glance, then let them go deeper. Build it from the decisions it supports rather than from the data you happen to have. Show state and change together — a number without a comparison cannot be interpreted — and encode what needs attention in form as well as in value, so it reads before anything is examined.
Why most dashboards go unread
They are built from the inside out. Someone lists the metrics the system can produce, arranges them in a grid, and ships it. Every number is accurate and the screen answers nothing, because no one asked a question it responds to.
The fix is to start from the other end: who opens this, how often, and what are they deciding? A dashboard for a daily operational check is a different object from one reviewed monthly by a director, and trying to serve both produces something that serves neither.
Design for one question
The strongest dashboards answer one question prominently and support several others beneath.
Is everything running normally? — operational monitoring. Are we on track this month? — commercial review. What needs my attention? — a work queue.
Write the question down. Then every element either helps answer it, supports the follow-up, or gets cut. This single constraint removes more clutter than any visual guideline.
Structure: summary before detail
A dashboard is scanned, not read. Lay it out in three tiers.
The answer, immediately. Two to five numbers at the top that resolve the main question. Someone should be able to open the page, look for two seconds, and know whether to keep going.
The breakdown. Charts and comparisons explaining the headline. This is where a rise or fall becomes attributable.
The detail. Tables and records, for when someone needs to act on a specific case.
Most failed dashboards begin at tier three, presenting a table of everything and requiring the reader to compute the summary themselves.
A number needs a comparison
This is the most common single defect. "1,284 orders" is uninterpretable. Compared with what — last week, target, last year?
Every headline figure needs state and change together: the value, a comparison, and the direction. Choose the comparison to match the rhythm of the business — week over week for weekly patterns, year over year for anything seasonal. Month-to-date against last full month is the classic misleading pairing, because it compares a partial period with a complete one.
A sparkline beside a number costs almost nothing and turns a snapshot into a trend.
Encode attention in form, not only value
If something needs action, it should read before anyone examines it. Use a status pill, a severity stripe, a chip — form the eye catches during a scan.
Two rules. Semantic colour is separate from your brand accent; good, warning and critical need to be distinguishable from each other and from decoration. And never rely on colour alone — pair it with an icon, a label or position, both for accessibility and for anyone glancing at a screen in bright light.
Use it sparingly. If a third of the screen is red, nothing is urgent.
Choosing the right chart
Match the form to the question.
Change over time — a line chart. Comparison between categories — a bar chart, sorted by value rather than alphabetically. Composition — a stacked bar, not a pie, unless there are two or three parts. Relationship — a scatter plot. Density across two dimensions — a heatmap.
The common errors are pie charts with eight slices, dual axes that manufacture a correlation, and truncated y-axes that exaggerate a small change. All three mislead, and the last two do so in ways that look professional.
Give every chart the same care as type: a light grid, a clear baseline, direct labels rather than a legend where possible, and an emphasised endpoint on a trend.
What to cut
Metrics that change no decision. If a plausible reading would not alter anything, it is decoration.
Vanity totals. Cumulative all-time counters only ever go up.
Precision nobody needs. 1,284,392 reads worse than 1.28M at a glance. Keep exact figures for the detail tier.
Redundant framing. Card titles repeating the section heading, axis labels repeating the card title.
Live-updating numbers that flicker. Movement pulls the eye. Reserve it for things that genuinely warrant attention.
Empty, loading and error states
Dashboards spend real time in all three, and they are usually designed last or not at all.
Empty — a new account with no data yet. Say what will appear here and what to do to produce it. A blank grid reads as broken.
Loading — use skeletons matching the eventual layout so nothing jumps when data arrives.
Partial failure — when one panel's data source is down, that panel should say so rather than showing zero. A zero and a failure look identical and mean opposite things, and that confusion causes real operational mistakes.
The test before shipping
Show it to the person it is for and ask what they would do next. If they scan it and act, it works. If they ask what they are looking at, the summary tier has failed. If they say "interesting" and close it, the dashboard is decoration — accurate, attractive, and changing nothing.
If you are building an interface people will operate daily and want it reviewed before build, book a call.
Common questions
What makes a good dashboard?
It answers one specific question for one specific person at a glance, then supports going deeper. Build it from the decisions it needs to support rather than from whatever data the system can produce, and structure it in tiers: the answer first, then the breakdown, then the detail.
Why is a single number not enough on a dashboard?
Because it cannot be interpreted without a comparison. '1,284 orders' means nothing until you know whether that is above or below last week, the target, or the same period last year. Every headline figure needs state and change together, with a comparison matched to the rhythm of the business.
Which chart type should I use?
Lines for change over time, bars sorted by value for comparison between categories, stacked bars rather than pies for composition, scatter plots for relationships, and heatmaps for density across two dimensions. Avoid pie charts with many slices, dual axes that imply false correlation, and truncated y-axes that exaggerate small changes.
How do you decide what to remove from a dashboard?
Cut anything where a plausible change in the number would not alter a decision. Also cut cumulative vanity totals, precision nobody needs, titles that repeat their section heading, and live-updating figures that flicker — movement attracts the eye and should be reserved for things that genuinely need attention.
How should a dashboard handle missing data?
A panel whose data source has failed must say so rather than displaying zero, because a zero and a failure look identical and mean opposite things — which causes real operational mistakes. Empty states should explain what will appear and how to produce it, and loading states should use skeletons matching the final layout.
