Most dashboards are art galleries for charts. They get curated, framed, hung on a wall, and quietly ignored. The CFO opens them on Monday morning, scrolls for thirty seconds, finds nothing actionable, and closes the tab. The marketing director glances at them in standup, mentions a number nobody can verify, and moves on. The CEO clicks through quarterly to make sure the dashboards still exist, and they always do, and the dashboards rarely change anyone's behavior.
The problem is that the dashboards were built around the data instead of the decision. Whoever built them started with here is what we have access to and ended with here is everything we can show. The output is a wall of fifty charts that collectively answer nothing because no specific question was asked.
Decision-driven dashboards work the opposite way. The starting question is what is the operator about to decide? The dashboard is built backward from that decision. Every chart, every number, every threshold either feeds the decision or it does not exist on the page. The result is a dashboard with three numbers above the fold, two supporting cuts, and a deliberately empty bottom of the screen, because the decision is already made by the time the operator scrolls. This essay is the practitioner playbook for building that dashboard. It lives in the Business Intelligence cluster, downstream of the data engineering covered in Source to Model, Observed.
Why the chart graveyard exists
The chart graveyard exists for a structural reason. Most dashboards get built by people who do not own the decision the dashboard is supposed to support. The data team builds what the data team can build. The visualization tool offers fifty chart types and the analyst tries every one of them. Stakeholders ask for additions and nobody removes anything because removing a chart feels political. Six months later the dashboard has eighty visual elements and every one of them is technically correct and collectively they tell the operator nothing.
This is the inverse of how an operator actually decides. An operator decides by ranking options. They ask should I increase the ad budget on Channel A or shift it to Channel B? The number that answers that question is one number, derived from the data, presented at a comparable scale, with a threshold for action. Everything else is context, and context belongs below the fold or in a drilldown, not above it.
The Notion portfolio frames the failure precisely: dashboards may look pretty, but nobody can explain which campaigns, offers, or customers actually drive profit. The pretty dashboard is the failure mode. The useful dashboard does not need to be pretty; it needs to make one specific decision faster.
The three numbers that change behavior
Three is the load-bearing constraint. Two numbers feel incomplete; the operator wonders what they are missing. Four numbers and the eye starts ranking, which means the fourth number always feels like clutter. Three numbers fits human working memory at a glance, leaves room for breathing space on the screen, and forces the dashboard owner to make hard ranking choices about what genuinely matters most.
Each of the three numbers carries three properties.
It is a number the operator can act on. If the number changes, the operator changes what they do. Cost per acquisition went from $80 to $95 last week is actionable; the operator can audit creatives, channels, or audiences and respond. Pageviews increased 8 percent is not actionable for most operators because it does not map to a budget decision or a campaign change.
It carries a threshold. The number means something only against a comparison: a target, a prior period, a benchmark, or a tripwire. CAC is $95 is data. CAC is $95 against a $80 target with a $120 walk-away threshold is a decision. The dashboard makes the threshold visible. Color, position, and visual weight communicate where the number sits in the action zones (green for safe, amber for watch, supernova orange for act-now).
It is fresh enough to act on. A number from last quarter is history. A number from yesterday is operational. The freshness floor depends on the decision cadence. Weekly budget decisions need daily data. Monthly retention decisions can run on weekly data. The dashboard names the freshness explicitly per the freshness SLA covered in Source to Model, Observed. Last updated 2 hours ago is part of the dashboard, not a footer afterthought.
The hierarchy of charts on the page
A decision-driven dashboard has four layers, each with a specific job.
Layer 1: the three hero numbers. Above the fold, large type, color-coded against thresholds, with the freshness timestamp visible. The first thing the operator sees, the first thing they react to. If the operator only ever looks at this layer, the dashboard has done its job.
Layer 2: trend. One sparkline or one small line chart per hero number, showing the last 30, 60, or 90 days. Trend gives context: is the number improving, deteriorating, or noisy? The trend chart is small. It is not the focal point. It exists to answer the second question the operator has, which is is this getting better or worse?
Layer 3: segment cut. One drilldown per hero number that breaks the metric down by the most actionable dimension. CAC by channel. LTV by acquisition cohort. Payback by campaign. The segment cut is what the operator clicks into when one of the hero numbers is off and they need to know why. It is not always visible; it expands when the operator engages.
Layer 4: deliberate emptiness below the fold. The space below the segment cuts is intentionally light. Maybe a row of secondary metrics for context. Maybe a link to the underlying tables for the analyst who wants to dig deeper. The point is that scrolling past the segment cut should feel like you are leaving the decision zone and entering the reference zone. The dashboard does not reward infinite scrolling with infinite charts, because every additional chart taxes attention without advancing the decision.
The Forensic Ad Audits work (case study) is the canonical example of this layering at work. The audit deliverable is not a 100-slide deck; it is a focused document with the three numbers that drove the wasted spend, the segment cut that proved which campaigns were responsible, and a prioritized action plan. The audit is short on purpose. The action is the deliverable.
How to pick the three numbers
Picking the three is the work. Most teams skip this and inherit whichever metrics the analytics tool defaults to (sessions, bounce rate, conversion rate). Default metrics do not earn their place because they were not chosen for a specific decision.
The selection process runs in three steps.
Step 1: name the decision. What does the operator decide weekly that this dashboard should support? How much do we spend on paid acquisition next week, and where? That is one decision. Which segment of customers is at the highest churn risk this month? That is another. One dashboard, one decision. A dashboard that supports three different decisions becomes three dashboards, not one. Multi-decision dashboards is how you end up with the chart graveyard.
Step 2: list every metric that could feed that decision. Cost per acquisition, lifetime value, payback period, contribution margin, channel mix, audience segment, attribution-corrected revenue, cohort retention, refund rate, and so on. Generate the full list without filtering. Twenty candidate metrics is normal at this stage.
Step 3: pressure-test each metric by removing it. For each candidate, ask: if I removed this number, would the operator still know what to do? If the answer is yes, the metric is not load-bearing. Cut it. The metrics that survive the cut are the ones that genuinely change the decision. By the end of the test, you usually have three to five metrics. Pick the three that have the lowest correlation with each other (the three that carry independent information). Those are your hero numbers.
The three numbers will look different per category. For ecommerce, the typical trio is contribution margin per order, blended CAC, and 30-day repeat purchase rate. For B2B SaaS, the typical trio is qualified pipeline value, CAC payback period, and net revenue retention. For services agencies, the typical trio is utilization rate, gross margin per project, and pipeline coverage. The trio depends on what the operator decides; copy-paste from another company's dashboard is the same chart-graveyard mistake one level up.
Case evidence
Home decor cohort dashboard. The DTC metal-art brand had ad-platform dashboards full of CTR, CPM, and ROAS numbers nobody was acting on. The decision-driven rebuild surfaced three: cost per acquisition by buyer segment, average order value by gift-occasion, and repeat purchase within 90 days. The 45-and-over female iPhone gift-buyer segment was visible at the top within a week. Spend rotated to that audience. The brand moved from $10K to $150K monthly. See Home Decor E-Com.
Forensic audit deliverable. The audit format is itself a decision-driven dashboard. The deliverable is a short document with three load-bearing numbers (typically wasted spend amount, audience overlap percentage, and campaign sprawl count), one segment cut that proves the diagnosis, and a prioritized action plan. The audit is intentionally short because the action is the product, not the report. See Forensic Ad Audits.
Solar portfolio brain. Fifty-plus solar accounts at one agency, each previously running its own per-account dashboard with twenty-plus metrics. The portfolio brain unified them with three numbers per account (cost per qualified install, install close rate, and gross margin per install) plus a portfolio-level rollup. Account managers stopped dashboard-shopping and started decision-shopping. New accounts onboarded faster because the playbook was clear. See Solar Portfolio Intelligence.
Anti-patterns to avoid
Vanity metrics on the hero row. Impressions, sessions, raw revenue without margin context, follower count. These look like leadership metrics and behave like decoration. Promote a vanity number to the hero row only when its movement genuinely changes the operator's behavior. Most of the time it does not.
Single-number dashboards with no threshold. A revenue number with no target is a status report. The threshold is what makes the number actionable. Revenue is $87K against the $100K target, with a $75K floor is decision-grade. Revenue is $87K is data.
Multi-decision dashboards. One screen that tries to support budget allocation, retention strategy, and product roadmap will support none of them well. Split it into three dashboards. Each one supports one decision. Each one has its own three numbers. Operators learn to navigate three single-purpose dashboards faster than they navigate one Frankenstein.
Dashboards owned by the team that built them, not the team that decides. A dashboard owned by the data team will drift toward what the data team finds interesting. A dashboard owned by the marketing director or the CFO will drift toward what they need to decide. Ownership matters. The operator who decides is the one with standing to demand changes and accept that other charts get cut.
Where to start
Three starting points.
Easiest, do today. Pick the most frequent decision your operator makes. Write the three numbers that, if shown, would change what they do this week. Put those three numbers on a sticky note. If you cannot name them in one sentence each, the work is naming them, not building the dashboard. Most operators do not have those three numbers; they have a vague sense of metrics that come up in meetings. The sticky note is the artifact.
Medium, this week. Build a v1 dashboard with just those three numbers, a 30-day trend per number, and one segment cut per number. Skip every other chart. Make it ugly if you have to; ship it functional first. The first time the dashboard is the only thing the operator looks at on Monday morning, you have proof the decision-driven structure works.
Hardest, this month. Wire the data layer so the three numbers refresh daily, the thresholds are explicit and visualized, and the freshness SLA is monitored per Source to Model, Observed. Add the action zones (color-coded thresholds, alerts when numbers cross). Now the dashboard is not just a screen the operator opens; it is an instrument that pages them when something genuinely needs attention. That is the version that compounds.
A dashboard built around the decision is shorter, sharper, and used. A dashboard built around the data is longer, prettier, and ignored. The work is the metric selection, not the visualization. For the upstream data engineering, see Source to Model, Observed. For why most attribution dashboards lie about which channels actually deserve credit, see Incrementality over Last-Click.