written by
Shashidhar S A
Founder, HubForge Global · MERL & AI Innovator
- En-Intrapreneur
- Driving evaluation ecosystems through People, Process & Product
- Co-Chair, IOCE Comms & Storytelling
- Co-Lead, EvalYouth Global Champion Engagement
A HubForge playbook
The best thinking on dashboard design, from Tufte to the Financial Times, turned into a method you can practise here and now, with AI doing the heavy lifting and you keeping the judgement. For any field, any tool.
What brings you here today?
1Begin with a test
That is roughly how long a busy reader gives a dashboard before deciding whether it is worth their time. Try it. You will see one dashboard for five seconds, then answer one question.
A typical district services dashboard. Look for which region needs help most.
The Hills region resolves only 51% of requests on time, twenty points behind every other region.
Which region needs help most?
Why this matters now
Most dashboards don't fail loudly. They launch, get opened twice, and quietly die. The causes are almost always the same three.
Every metric goes on the page because it exists, not because anyone acts on it.
Twenty charts, five colours, three fonts. The one number that matters gets lost.
No definitions, no source, no date. One wrong figure and people return to spreadsheets.
The frameworks in this playbook give you the right dashboard: decades of research on how people read charts and make decisions. AI gives you that dashboard built and checked, faster, once you tell it what good looks like.
2The method
Meet Sahayak, a public helpline that takes thousands of requests a month. Its managers asked for "a dashboard". As you read the six stages of the DESIGN loop, watch their dashboard take shape beside you. The same steps work for a sales team, a clinic network or a school system.
3Keep this open
Sahayak showed the loop in motion. These are the reference pages to come back to while you build your own: the whole loop on one screen, the four kinds of dashboard, the metric card and the layout rules.
Who decides what, how often?
Stands on Patton's utilization-focused evaluation, Munzner's domain level and jobs to be done.
Which numbers answer those questions, exactly?
Stands on SMART indicators, results chains and Few's warning about deficient measures.
What does the reader see first, second, third?
Stands on Shneiderman's mantra, Few's single screen, Gestalt and Z and F reading.
Which chart shows each relationship most clearly?
Stands on the FT Visual Vocabulary, Tufte's data-ink, Knaflic and pre-attentive attributes.
Can everyone read it, and trust it?
Stands on WCAG 2.2, the Okabe and Ito palette and do-no-harm data ethics.
Is it used, and getting better?
Stands on usability testing, Few's thirteen mistakes and Munzner's validation.
Mixing these on one page is the most common design mistake of all. If you need two, build two pages. Tap a kind to see what it looks like.
Metric card, Sahayak example
closed_within_7d ÷ requests_closedHand your AI the theory of change and the data dictionary, and ask for six of these cards. Then ask which ones it could not compute from your data; that list is the real finding.
The inverted pyramid
Shneiderman's mantra, 1996
A one-screen wireframe
4Practise
Five tools to sharpen your eye. Each one teaches a principle by letting you do it, not read about it.
This chart went to a board meeting. It contains five classic mistakes. Click on each one you can find.
Good charts start with what you want the reader to see. Finish the sentence and the chart follows, using the Financial Times Visual Vocabulary.
I want my reader to see…
From Tufte's data-ink and Knaflic's "eliminate clutter, focus attention". Step through and watch a default chart become a clear one.
Monthly requests by region, January to June
The eye spots colour, size, position, length and enclosure before reading a word. Use one of them, once, on the thing that matters.
Bars start at zero, scales stay consistent across small multiples, and areas are never used for one-dimensional values.
About 1 in 12 men and 1 in 200 women have some form of colour vision deficiency. Switch the view, then switch the palette.
Stephen Few's thirteen common dashboard mistakes, turned into a self-check. Tick what your dashboard already avoids. The prompt on the right asks AI to critique the rest.
5Working with AI
AI can draft almost everything on a dashboard. It cannot be accountable for any of it. Here is how to split the work: eight jobs it does well, what you keep in each, and six promises that hold whatever the tool.
AI is a fast, tireless collaborator. It is not accountable for your numbers. You are.
Surfaces purpose, readers and decisions.
You keep: confirming with real usersTurns questions into metric cards.
You keep: signing off definitionsDrafts wireframes and reading order.
You keep: choosing what to cutMatches relationships to charts.
You keep: the final visual callWrites DAX, SQL, formulas or chart code.
You keep: recomputing the headlinesReviews against Few, Tufte and Knaflic.
You keep: weighing trade-offsFlags contrast, colour-only meaning and jargon.
You keep: testing with a contrast toolWrites takeaways, summaries and alt text.
You keep: checking every claimThe one it can't play: the person who answers for the numbers.
Nothing on the page you can't follow back to source data and a formula.
The top three figures are checked by a person before anyone else sees them.
Aggregate and de-identify first, unless your organisation has approved the tool.
AI fills a gap with a plausible definition. Give it yours.
Each dashboard has a named owner for numbers, wording and release.
One line on the page: what AI drafted, what people checked.
| Tool | AI help for dashboards | Best for |
|---|---|---|
| General assistants | Claude, ChatGPT, Gemini and Copilot draft metric cards, layouts, chart code and critiques, and analyse files | Design thinking, critique, code |
| Power BI with Copilot | Generates report pages, suggests DAX measures, writes narrative summaries, answers questions | Microsoft 365 teams |
| Tableau Pulse | Personalised metric summaries, proactive insights, alerts by email or Slack | Metric monitoring |
| Looker with Gemini | Conversational analytics, a visualisation assistant, calculated fields | Google Cloud teams |
| ThoughtSpot Spotter | Conversational analysis and AI-assisted dashboard creation | Search-style self-service |
| Metabase Metabot | SQL generation, chart summaries, query debugging | Open-source, lean teams |
| Spreadsheets | Copilot in Excel and Gemini in Sheets suggest formulas, pivots and charts | Small teams, trackers |
AI features in BI tools change every few months. Check your tool's release notes; the pact holds whatever the tool.
6Borrow these
Eighteen openings for any AI assistant, grouped by stage. Replace the highlighted words. Then push back on the first answer; the second one is usually better.
7Before you launch
Twelve checks, two for each stage of the loop. Tick them off for the dashboard you are building; the ring fills as you go.
Start with the purpose.
8Go deeper
The concepts, theories and practices this playbook stands on, each in plain words: where it comes from, why it matters, and how to use it on Monday.
The books, papers and standards behind the ideas above, if you want the originals.
written by
Founder, HubForge Global · MERL & AI Innovator