learn-business-intelligence-with-phoebe / Leader session 3 of 6
Learn Business Intelligence with Phoebe · Leader track · Session 3 of 6

One number, one truth

Sales says $2.1M. Finance says $1.8M. Both dashboards are technically correct, and the meeting still burns forty minutes arguing about arithmetic instead of deciding anything. This session is the one no vendor course teaches leaders: why correct dashboards disagree, what a semantic layer actually is, and the one organizational fix - an owner and a written definition for every number the company runs on.

🟡 Leader track Leaders: execs · sponsors · managers No code · 45 minutes
0-3 · Welcome 3-16 · Why numbers disagree 16-31 · The semantic layer 31-45 · Governance + your turn
Part 0

The most expensive meeting in your company

Every organization has a version of it: two teams present the same metric with two different values, and the room debates whose number is real. Nobody lied. Nobody miscounted. The dashboards diverged for reasons that are completely knowable - and completely fixable, once someone with authority decides to fix them. That someone is you. This is the flagship session of the leader track, because it is the one BI problem that only leadership can solve.

Live - presented in session Self-study - read after class ▶ Mini-BI - interactive playground Official sources covered
★ What you walk out with today The four reasons two correct dashboards disagree, a plain-language picture of what a semantic layer is (the same idea hiding inside Power BI, Looker, and dbt), and a one-page metric charter you can enact next week: every load-bearing metric gets an owner, a written definition, and a certified home.
Part 1 · the meeting every leader has survived

Why two dashboards disagree 13 min live

Here is the anatomy of the classic fight. Sales reports $2.1M because their dashboard counts every order the moment it is booked, cancelled or not. Finance reports $1.8M because theirs counts only recognized revenue on completed orders, refreshed from last night's close. Same warehouse, same month, two defensible queries. The disagreement lives in the definitions, not the data.

Warehouse one orders table same rows for all Sales team's query keeps cancelled orders counts when booked refreshed hourly Finance team's query drops cancelled orders counts when recognized refreshed nightly Sales dashboard revenue: $2.1M Finance dashboard revenue: $1.8M filters · definitions joins · refresh Neither number is wrong. Both queries run on the same rows - the definitions differ. The fix is structural.
🔍 Click to zoom - one warehouse, two correct queries, two different truths
LiveThe four divergence points6 min

When two dashboards disagree, the gap almost always enters at one of exactly four places. Learn these and you can chair the "whose number is right" meeting in five minutes instead of forty:

  • Different filters. Are cancelled orders in or out? Refunds? Test accounts? Internal orders? Each yes/no silently moves the number.
  • Different definitions. Booked revenue vs recognized revenue. Active customer as "ordered this month" vs "has a live subscription". Same word, different formula.
  • Different joins. One query joins orders to order items, another joins through subscriptions - and rows quietly duplicate or drop along the way.
  • Different refresh times. Sales refreshed at 2pm, finance at midnight. On a fast-moving day, that alone explains a real gap.
Real world

The forty-minute audit. A leadership team we know ran the same "whose churn number is right" argument monthly for a year. The eventual answer: one dashboard counted paused subscriptions as churned, the other did not. Four hours of analyst time to diagnose, one sentence to fix - once someone wrote the definition down and made it official.

Self-studyAnatomy of a metric definition4 min read

A metric is not a name - it is a small contract with six clauses. If any clause is unwritten, two teams will fill it in differently. Here is the full anatomy, using Daybreak's "revenue" as the example:

FieldWhat it pins downDaybreak "revenue" example
NameThe one official labelNet product revenue
FormulaThe exact mathsum of quantity x unit price per order item
GrainWhat one row meansorder item, rolled up by order date
FiltersWhat is in and outexcludes cancelled orders and test accounts
OwnerWho arbitrates changesVP Finance
RefreshHow fresh it isnightly at 2am, complete through yesterday

Six short rows. Most metric fights in most companies trace back to one of these rows never having been written down.

Part 2 · covers Power BI semantic models, Looker LookML, dbt MetricFlow (concepts)

The semantic layer: define once, agree everywhere 15 min live

You could fix metric drift with discipline: memos, reminders, heroic analysts checking each other's SQL. It will not hold. The durable fix is structural: encode the business definitions once, in a layer that sits between the warehouse and every chart, so that no dashboard computes "revenue" on its own. The industry calls this a semantic layer - a governed dictionary that every tool is forced to read from.

Orders tables raw transactions Billing tables invoices, refunds Customer tables plans, subscriptions Semantic layer "revenue" defined ONCE filters, grain, owner Exec dashboard same $1.94M Team reports same $1.94M AI assistant same $1.94M The hourglass waist is the fix: change a definition once here, and every dashboard, report, and AI answer moves together.
🔍 Click to zoom - the hourglass: many sources, one definitions layer, many consumers
LiveThe same idea in every vendor6 min

Vendors brand this layer differently, and the branding hides how converged the industry actually is. Strip the names and all three are the same move: a governed metadata layer between warehouse and consumption, centralizing business definitions so every tool computes the same numbers.

VendorWhat they call itWho edits it
Microsoft Power BISemantic model - a star schema plus measures plus business terminology, positioned as the single source of truth. Renamed from "dataset" in November 2023, precisely to signal it carries meaning, not just data.BI developers, in the model layer
LookerLookML - a language defining dimensions, measures, views, and models. Looker generates the SQL from it, so business users explore freely and never write a query.The data team, in versioned LookML files
dbt Semantic LayerMetricFlow - metrics defined once in YAML on top of semantic models. The pitch is literally "define once, use everywhere", and any downstream tool can request the metric.Analytics engineers, in the dbt project
The leader takeaway You do not need to pick a vendor today. You need to know that whichever BI stack your team runs, this layer exists - and if nobody in your organization is deliberately curating it, your definitions are being set by accident, one dashboard at a time.
Self-studyWhat it costs - the honest trade-offs4 min read

A semantic layer is not free, and pretending otherwise is how these initiatives die in month three. Budget for three real costs:

  • Modeling time up front. Someone has to interview the business, reconcile the competing definitions, and encode the winners. For a mid-size company's top 20 metrics, that is weeks of skilled work, not days.
  • A gatekeeping function. Once definitions are governed, changing one requires a request and a review. That is the point - but teams used to editing their own dashboard formulas will feel slowed down, and will say so.
  • A slower first chart. Ad-hoc tools let anyone chart anything in minutes. Through a semantic layer, a brand-new metric must be defined before it can be charted. You are trading speed on the first chart for trust on every chart after it.

The trade is worth it for the numbers the company steers by. It is overkill for one-off explorations - which is exactly why Part 3 separates certified content from ad-hoc content instead of governing everything.

Part 3 · the org fix only leaders can enact

Metric governance for leaders 6 min live

The technology in Part 2 needs an organizational spine, and it is small enough to write on an index card: every key metric gets an owner, a written definition, and a certified home. The owner arbitrates disputes and approves changes. The definition is the six-field contract from Part 1. The certified home is the one dashboard where the official number lives - marked as certified, so everyone can tell it apart from the ad-hoc charts teams spin up for their own exploration. Ad-hoc content stays legal and fast; it just cannot masquerade as the official number. Then a light ritual keeps it alive: a quarterly metrics review where owners walk through changes, retire dead metrics, and settle open disputes.

LiveYour metric charter4 min

This is the whole program. If you enact nothing else from this course, enact this checklist:

  • List the load-bearing metrics. The 5 to 15 numbers that appear in board decks and steer real decisions. Not fifty - the ones people fight about.
  • Name one owner per metric. A person, not a committee. Ties go to whoever answers for the number when it is wrong.
  • Write the six-field definition. Name, formula, grain, filters, owner, refresh. One page for all of them.
  • Certify one home per metric. One dashboard is the official source; it gets the certified mark. Everything else is clearly ad-hoc.
  • Schedule the metrics review. Quarterly, 45 minutes, owners present. Definition changes happen here and nowhere else.
★ Why this is a leadership act Analysts can draft every one of these artifacts - but only leadership can make the certified number the one that settles arguments. The charter works the day an executive says, in a real meeting, "that is not the certified number, so we are not debating it."
Demo 1 of 2

Watch definitions matter ★ 12 min · follow along

Daybreak, our coffee-subscription company, has one warehouse and one innocent-looking question: what is our revenue? Watch how one definitional choice - do cancelled orders count? - moves the company's headline number. No code; just pick and read.

Look at revenue by order status. The box below splits Daybreak's revenue by status. Some of that money sits on cancelled orders. Should "revenue" include it? Sales-style booked revenue says yes; finance-style recognized revenue says no. There is no math answer - it is a definition decision.

Now look at the single company number. The KPI box below shows total revenue as one figure. Whichever choice a team makes in step 1 silently moves this number - and this is the number that opens the board deck.

Press Show SQL on both boxes. Read the generated query. The definition IS the query: a one-line filter on status is the entire difference between the $2.1M meeting and the $1.8M meeting. Whoever controls that line controls the truth.

Real world

This is what a semantic layer would freeze. In a governed setup, the status filter lives in the definitions layer, written once by the metric owner. Neither sales nor finance could produce a different "revenue" without visibly requesting a definition change. The forty-minute meeting becomes structurally impossible.

Demo 2 of 2

Your turn: draft a definition card ★ 8 min · pen and paper

Pick the metric your organization fights about most - revenue, churn, active users, pipeline, headcount, utilization. Draft its six-field definition card right now, in the session. Incomplete is fine; the blanks you cannot fill are the most valuable output.

Name the fight. Which metric triggered the last "whose number is right" moment you personally sat through? Write its official name - or the two competing names, which is itself a finding.

Fill the six fields. Name, formula (in words, not code), grain, filters (what is in and out - be specific about cancellations, refunds, test data), owner, refresh. Prompt yourself: could a new analyst rebuild this number from my card alone?

Circle the blanks. Every field you could not fill is a divergence point waiting to fire. If you could not name an owner, you have found the root cause - a number nobody owns is a number everyone redefines.

Bring it to the metrics review This card is the unit of work for the whole governance program. Ten cards like it, owned and certified, and your organization has a semantic layer on paper - which is the correct first version. The tooling in Part 2 just automates what the cards agree on.
Homework

Try it yourself - this week ◐ 20-30 min total

Source material

Official sources covered

No commercial platform teaches semantic-layer governance to leaders - the vendors each document their own layer for practitioners. This session unifies those documents into the leader view. This page covers:

Microsoft · Power BI semantic model documentationPart 2 · the concept and the November 2023 rename; hands-on modeling stays in b4
Looker LookML + dbt Semantic Layer (MetricFlow) documentationPart 2 · the same governed-definitions idea across vendors, concept level
PL-300 · "semantic model" terminology and model designBuilder depth - measures in b4, certification mapping in b10
Check yourself

Three questions before you go 🎯 ◐ 90 seconds

1 · Sales shows $2.1M, finance shows $1.8M, and both queries ran correctly on the same warehouse. The most likely cause is...

Correct queries with different results almost always diverge at one of four points: filters, definitions, joins, or refresh timing. It is a definitions problem, not a fraud or infrastructure problem - which is why the fix is a written definition, not an investigation.

2 · What does a semantic layer actually store?

The semantic layer holds meaning, not data: what "revenue" is, which orders count, what one row means. Power BI semantic models, LookML, and dbt MetricFlow are all this same governed dictionary - define once, agree everywhere.

3 · Your organization wants "one version of the truth". The right first step is...

Tools automate agreements; they cannot create them. Until a named owner and a written definition exist, any platform will faithfully compute several versions of the truth. Governance first, technology second - and locking everything down just drives teams back to spreadsheets.

Leader session 3 cheat sheet · pin this

The classic fightTwo correct dashboards, two revenue numbers. The disagreement lives in definitions, not data.
Four divergence pointsFilters · definitions · joins · refresh times. Check them in that order and the mystery dies in minutes.
Metric anatomyName, formula, grain, filters, owner, refresh. Six fields; any blank one becomes a fight later.
Semantic layerBusiness definitions encoded once between warehouse and charts. Definitions, not data. The hourglass waist.
Vendor namesPower BI semantic model (ex "dataset", renamed Nov 2023) · Looker LookML · dbt MetricFlow. Same idea, three logos.
Honest costsModeling time, a gatekeeping function, slower first chart. Trading first-chart speed for every-chart trust.
Certified vs ad-hocOne certified home per metric; ad-hoc stays legal but can never pose as the official number.
The metric charterOwner + written definition + certified home per key metric, plus a quarterly metrics review. That is the whole program.