Why this is the last session
Notice the order of this track: dashboards, metrics, commissioning, operating models - and only now, tools. That was deliberate. A tool choice made before the operating model is a lottery ticket; made after, it is a checklist. Everything from a1-a5 becomes your evaluation criteria today. And the AI question belongs in the same session, because the honest answer to "can AI do our BI?" turns out to be: only as well as the semantic layer you built underneath it.
The landscape in one table 10 min live
Five families cover the market that matters to you. Power BI: the Microsoft-stack default, now part of Fabric, lowest seat cost of the big three and the strongest governance story. Tableau: the visualization-craft pedigree, owned by Salesforce. Looker: Google Cloud's play, built on LookML - a semantic modeling language that generates SQL so business users never write it (the industrial version of our mini-BI's Show SQL trick). Looker Studio: Google's free browser tool, light and fast to start. Metabase and Apache Superset: open source - free licenses, self-hosted, thinner governance.
| Tool | Stack fit | Semantic layer | Governance | Cost shape | Who it suits |
|---|---|---|---|---|---|
| Power BI | Microsoft 365 / Azure / Fabric | Semantic models (the renamed "datasets") | Strongest story: endorsement, RLS, labels | Low per seat, scales up with Fabric | Microsoft shops, compliance-heavy orgs |
| Tableau | Anything; Salesforce-owned | Present, lighter than rivals | Mature via Blueprint practices | Premium per seat | Viz-craft cultures, analyst-heavy teams |
| Looker | Google Cloud / BigQuery | LookML - the most rigorous: code-reviewed definitions | Strong, enforced in code | High, plus modeling time | Data-product teams, embedded analytics |
| Looker Studio | Google Workspace | Minimal | Thin | Free | Startups, marketing teams, first dashboards |
| Metabase / Superset | Self-hosted, any warehouse | Basic models | Thin - you build the rituals yourself | Free licenses, real hosting + ops cost | Engineering-led teams, budget-tight orgs |
LiveHow to shortlist in one meeting4 min▶
Four questions, asked in order, collapse the table to one or two candidates before the coffee goes cold:
- What stack are we already on? Microsoft 365 everywhere → Power BI starts with a lead it rarely loses. Google Workspace and BigQuery → Looker Studio now, Looker later. No dominant stack → Tableau and the open-source pair stay in the race.
- How much semantic-layer rigor do we need? If a3's one-number-one-truth problem is your daily pain, weight Looker's LookML and Power BI's semantic models heavily. If you have five dashboards total, rigor can wait.
- How many viewers versus builders? Hundreds of viewers make per-seat pricing the whole conversation - model the viewer tier, not the builder tier.
- What can IT actually run? Metabase and Superset are only free if you have engineers happy to host, patch, and upgrade them. No platform team, no open source - the license saving becomes an ops bill.
Self-studyWhy "best tool" is the wrong question3 min read▶
Three reasons the ranking-review framing misleads leaders:
- Switching costs dwarf license costs. Migrating 300 dashboards and retraining 40 builders costs more than any price gap between vendors. The tool you can adopt cleanly beats the tool that wins benchmarks.
- The talent market is a feature. Power BI and Tableau skills are abundant and certifiable (PL-300, Tableau's exams); niche tools mean every hire needs retraining. Your hiring pipeline is part of the architecture.
- The semantic layer outlives the viz tool. Companies swap charting front-ends every five to eight years; the definitions of revenue and churn persist across the swap. Invest in the definitions - they are the asset. The pixels are fashion.
AI in BI - what is real in 2026 10 min live
Here is the strongest signal that AI-in-BI crossed from demo to product: Copilot is now on the PL-300 exam. Microsoft tests four sub-skills - create a narrative visual with Copilot, use Copilot to create a new report page, use Copilot to suggest content for a new report page, and use Copilot to summarize the underlying semantic model. Vendors do not put vaporware on certification exams. Natural-language Q&A over a governed semantic model genuinely works; narrative visuals genuinely summarize a page in prose. The catch is the phrase in the middle: over a governed semantic model.
LiveWhat to pilot now4 min▶
Two AI pilots have a real payoff-to-risk ratio in 2026, and both lean on work you have already commissioned:
- Natural-language Q&A on your ONE certified model. Scope the pilot to the certified core from a5 - the model where definitions, RLS, and refresh are already audited. When an exec types "revenue by channel last quarter" and gets the governed number, trust compounds. Microsoft even ships a dedicated MS Learn module, "Get started with Copilot in Power BI" - a fine one-hour assignment for your champions.
- Narrative summaries for exec packs. The narrative visual turns a dashboard page into three paragraphs of prose - a genuine time-saver for the Monday pack. Rule: a human owner still signs the narrative before it circulates. AI drafts; the owner answers for it.
Success metric for both: not "wow" in the demo, but whether the answers match the certified dashboard for a full month.
Self-studyWhat to distrust3 min read▶
Three patterns should trip your alarm in any vendor pitch or internal experiment:
- AI over raw spreadsheets. Pointing a chatbot at an ungoverned export and asking for revenue gives you a fluent answer with no definition behind it. It will disagree with finance, and nobody will notice until the board deck.
- Screenshots of charts pasted into chatbots. The model reads pixels, guesses the numbers, and hallucinates precision. If the data never touched a governed model, the answer inherits nothing you can audit.
- "No semantic layer needed" pitches. Any tool claiming AI removes the need for modeled definitions is asking you to automate the forty-versions-of-revenue problem. The semantic layer is what the AI reads - remove it and the AI reads vibes.
Build vs buy vs wait - and the 90-day plan 6 min live
Sometimes the right tool decision is to not decide yet. Looker Studio or Metabase is genuinely enough while you have one data person, a handful of dashboards, and viewers in the dozens - free tools plus the a5 rituals beat an expensive platform with no operating model. You graduate when the trust tiers need enforcement the tool cannot provide, RLS becomes a compliance requirement, or viewer counts make governance manual work. And when you do buy, remember the real total cost: licenses + modeling time + enablement. The license line is often the smallest of the three.
LiveYour 90-day BI roadmap4 min▶
The whole leader track, compressed into three moves:
- Days 1-30 · Certify five metrics. Run a3's one-number-one-truth exercise on revenue, orders, active customers, AOV, churn. Named owner, checked definition, monitored refresh - the certified core exists.
- Days 31-60 · Pick the operating model. Draw a5's hub-and-spoke on the whiteboard, name the hub, name one champion per department, and put the trust-tier badges in front of viewers.
- Days 61-90 · Pilot AI Q&A on the certified core. Scope it to the governed model only, measure answer-match for a month, and report the result to the exec team - whatever it is. A failed pilot honestly reported builds more credibility than a demo that never ships.
Notice the order: definitions, then org, then AI. Every failed BI-plus-AI program you will ever read about ran this list backwards.
Choose for three companies ★ 12 min · argue from the table
Three companies, three recommendations. For each, justify the pick using only the Part 1 table - stack fit, semantic layer, governance, cost shape. If your reasoning does not cite a column, it is a preference, not a recommendation.
40-person startup on Google Workspace, one data-curious marketer. Pick: Looker Studio. Free, browser-based, native to the stack they already pay for. Governance needs are one certified sheet of definitions and a weekly ritual - a5 practices, no platform required. Graduate when viewer count or RLS needs say so.
2000-person Microsoft shop with compliance requirements. Pick: Power BI. The stack-fit column decides it (Microsoft 365 everywhere), and the governance column seals it: endorsement tiers, RLS, sensitivity labels - the entire a5 guardrail kit as native features, at the lowest per-seat cost of the big three. Fabric questions (DirectLake vs Import) become relevant at their data scale.
Data-product company selling embedded analytics to customers. Pick: Looker or Superset. Dashboards ARE the product here, so definitions must be code-reviewed like product code - LookML's whole reason to exist. Superset is the engineering-led alternative when license cost matters more than modeling rigor and the team can self-host.
The same table, run in reverse. A mid-size retailer chose a premium tool because the demo dazzled the exec team - no stack fit (they were all-Google), no builder on staff, no operating model. Eighteen months of shelfware later, they restarted on Looker Studio with five certified metrics and a champion network, and adoption finally moved. The tool was never the blocker; the missing columns were.
The AI reality check ★ 8 min · see the dependency
One last visit to the mini-BI, with new eyes. Our playground generates SQL from your clicks; Copilot generates it from your English. Different input, identical dependency: both are only as right as the semantic layer that translates the request.
Press Show SQL. Read the generated query. When an exec asks Copilot "monthly revenue this year", the AI must produce this - same joins, same definition of revenue, same date logic. The English changed; the target did not.
Now imagine the drift. If two teams held different definitions of revenue (a3's nightmare), the AI would fluently answer with one of them - and no one could tell which without reading the SQL. Governance is not slowing your AI down; it is the only thing making its answers checkable.
Say the sentence out loud. "The AI writes the query; the semantic model decides if it is true." If you retain one sentence from six sessions, make it that one.
Try it yourself - this week ◐ 20-30 min total
- Score your current BI tool against the four shortlist questions: stack fit, semantic-layer need, viewer count, IT capacity. Write one sentence: keep, graduate, or replace - and why.
- Find your org's one certified-worthy semantic model - the candidate for an AI Q&A pilot. If none qualifies, that IS the finding: run the 90-day roadmap first, days 1-30.
- Put the one-line AI policy in front of whoever owns data policy: AI answers questions only from certified semantic models.
- That completes the leader track - six sessions from "what is BI" to a tool decision and an AI policy. If the mini-BI's Show SQL button ever made you curious about what your builders actually do, builder session 1 starts exactly there.
Official sources covered
The finale covers the newest exam material of the whole track - Copilot entered the PL-300 skills list in the April 2026 study-guide update - plus the vendor landscape read leaders actually need:
Three questions before you go 🎯 ◐ 90 seconds
1 · What is the prerequisite for AI Q&A your execs can actually trust?
The AI writes the query; the semantic model decides if it is true. Connecting more ungoverned data makes the answers more fluent and less checkable - governance first, then automation.
2 · A 30-person team on Google Workspace wants its first dashboards this month, budget zero. Best starting point?
Stack fit plus cost shape decides it: Looker Studio costs nothing and lives where they already work. Governance rituals can come from a5's playbook until scale forces a graduation.
3 · Your exec team asks "which BI tool is the best?" What is the right reframe?
"Best" without context is the wrong question: switching costs dwarf license gaps, the talent market matters, and the semantic layer outlives the viz tool. The four shortlist questions turn a beauty contest into a decision.