[Intermediate] — the tools that let a whole organisation see the data, the semantic layer that keeps everyone’s numbers agreeing, and why most dashboards fail.
The kid version first
An analyst can’t personally answer every question for every person — so business intelligence tools let people see the data themselves, through dashboards and reports that update automatically. Done well, this frees the analyst from repetitive pulls and lets the business self-serve. Done badly — and it usually is — you get a graveyard of unused dashboards, and three teams reporting three different revenue numbers because everyone defined it differently.
What BI tools do
A BI (business intelligence) tool connects to the warehouse, lets you build charts and dashboards over it, and lets non-technical people explore and filter without writing SQL:
| Tool | Character |
|---|---|
| Tableau | The powerful incumbent — rich visualisation, steep-ish, expensive |
| Power BI | Microsoft’s — strong, cheaper, dominant in Microsoft shops |
| Looker | Code-first, built around a semantic layer (LookML) — governed, engineer-friendly |
| Metabase | Open-source, simple, great for self-serve and smaller teams |
| Superset | Open-source, the Airbnb-origin one |
| Mode / Hex | SQL-notebook hybrids — analyst-first, blend SQL, Python and viz |
They all generate SQL underneath and run it against the warehouse → SQL. The differences are governance, cost, and how technical the builder needs to be. The tool matters less than how you use it.
The semantic layer — the fix for “whose number is right?”
The most important BI concept, and the one that separates a trustworthy analytics setup from chaos.
The problem: without a shared definition, marketing calculates “active users” one way, product another, finance a third — and every dashboard disagrees, so nobody trusts any of them. Meetings become arguments about whose number is right instead of what to do.
The semantic layer is a central place where metrics are defined once — “revenue,” “active user,” “churn” each have one canonical definition, in code, that every dashboard and query inherits:
raw warehouse tables
│
SEMANTIC LAYER ← "revenue = SUM(amount) WHERE status='completed' AND is_test=false"
│ defined ONCE, governed, versioned
every dashboard, every query, same number
This is where BI meets dbt — the transformation layer and the semantic layer are how “one definition of every metric” actually gets enforced. LookML (Looker), dbt’s semantic layer, Cube, and others implement it. A metric defined in twelve dashboards is a metric defined twelve slightly-different ways; a metric defined in the semantic layer is defined once and trusted. → defining a metric precisely.
Self-serve analytics — the goal and the trap
The dream: business users answer their own questions, and analysts do high-value work instead of pulling numbers. Real, and worth pursuing — but it fails in predictable ways:
- Without governance, self-serve produces contradictory numbers — everyone builds their own slightly-wrong version. The semantic layer is the prerequisite, not an optional extra
- Non-analysts misinterpret — they read correlation as causation, ignore seasonality, or draw conclusions from a tiny segment. Self-serve gives people rope
- The dashboard graveyard — most dashboards built get used a few times and abandoned. Building dashboards nobody uses is a huge, common waste
Self-serve works when the metrics are governed (semantic layer), the common questions are anticipated, and users are taught enough to not mislead themselves. It’s a socio-technical problem, not just a tooling one.
Dashboard design — most are bad
A dashboard is a communication artefact, and the data-visualization principles apply with force. The common failures and their fixes:
- Too much — 40 charts, no hierarchy, no story. A dashboard should answer a specific set of questions for a specific audience, not display everything possible. Cut ruthlessly
- No context — a number with no comparison (vs target, vs last period, vs benchmark) is meaningless → comparability. Every metric needs a reference point
- Wrong chart — a pie chart with 12 slices, a line chart for categories, dual axes that mislead → choosing the right chart
- No clear action — the viewer should know what to do, or at least what to look at first. Lead with the headline
- Vanity over signal — big impressive cumulative numbers instead of the rates that show performance → vanity metrics
Match the format to the need, which is the decision most people skip:
| Need | Format |
|---|---|
| Monitor the same metrics continuously | Dashboard (live, glanceable) |
| Answer a specific one-off question | A report / analysis (narrative, a “so what”) → communicating |
| Explore freely | A BI exploration tool / notebook |
| A number checked daily by one person | An alert, not a dashboard |
A dashboard is for monitoring known metrics; a report is for answering a question. Building a dashboard to answer a one-time question (and a report for something checked daily) is a common mismatch that wastes effort and buries the insight.
Key insight
BI tools let an organisation see its own data, but the tool matters far less than the semantic layer — the central, coded, single definition of every metric that stops three teams from reporting three different revenue numbers. Self-serve analytics is a real goal that fails without that governance, plus anticipated questions and enough user education to prevent misinterpretation. And most dashboards fail as communication: too much, no context, no clear action, and built to answer one-time questions that a report should have answered. Design a dashboard for monitoring known metrics with reference points and a clear headline — and if nobody uses it, stop building it.
Related
- metrics and KPIs — what the semantic layer defines
- dbt — where the semantic layer lives
- data visualization — choosing charts that don’t mislead
- communicating analysis — dashboards vs reports
Source: [reference] — Sep 2026.
Software engineering
Infrastructure & security
Computer science
Hardware & engineering