Your Data Team Doesn’t Need More Dashboards — It Needs Semantic Governance

Every data team we talk to has more dashboards than it can name. Most also have a revenue number that differs depending on which dashboard you open, a "customer" definition that changed in 2023 without anyone updating the older reports, and a new request in the backlog for — another dashboard. The instinct to answer every question with a chart is understandable. It is also why the questions keep coming back.

Dashboards are the symptom

A dashboard is a frozen query with a picture on top. When two dashboards disagree, it is because two queries disagree, and the queries disagree because there is no single, enforced definition of the metric they both claim to show. Adding a third dashboard adds a third query. The problem is not a shortage of visualisation; it is that the meaning of the data — what a customer is, when revenue is recognised, which events count as active — lives in dozens of SQL files instead of one governed place.

Generative AI makes this worse before it makes it better. An LLM that writes SQL over an ungoverned warehouse will confidently produce a fourth revenue number. Natural-language analytics only works when the model is grounded in definitions that humans have agreed on.

What semantic governance means

Semantic governance is the practice of defining business concepts — entities, metrics, dimensions, and the rules that connect them — once, in a form that both people and systems consume, and enforcing that every report, model, and AI feature reads from that definition rather than re-deriving it.

In BigQuery terms, it has four layers:

1. Physical governance. Row-level and column-level security, data masking, and Dataplex Universal Catalog policy tags so sensitive columns are protected regardless of which tool reads them.
2. Curated models. A conformed layer — built in Dataform, tested with assertions — where customers, orders, and revenue exist exactly once with documented grain and business rules.
3. Metric definitions. Measures defined in a semantic layer the BI tools share. In Google's stack that is a Looker semantic model (LookML), which Looker Studio can now consume directly and which Gemini-powered conversational analytics in Looker uses to answer questions with the same definitions the dashboards use.
4. Discoverability. Descriptions, owners, and freshness on every curated table in the catalog, so an analyst — or an agent — picks the governed table rather than a raw one.

Google has been steering toward this with its Knowledge Catalog direction: a catalog that carries business glossary terms, lineage, and data-product definitions that AI features can read. The product names will keep shifting; the underlying idea — governed meaning as an input to AI — is the durable part.

Why this ties directly to AI readiness

Every generative-AI feature in BigQuery is a consumer of metadata:

• SQL generation in BigQuery Studio reads table and column descriptions. Empty descriptions mean guessed joins.
• Conversational analytics answers from a semantic model. No model, no trustworthy answers.
• The Data Engineering Agent builds pipelines from existing definitions. If the conformed layer is a mess, the agent reproduces the mess faster.
• RAG over warehouse data inherits row-level security — but only if it exists.

Teams that skipped governance to "move fast on AI" discover that the AI features are precisely as reliable as the metadata underneath them, which is to say not at all.

What to do instead of building the next dashboard

• Pick the three metrics that matter most and define each once: name, grain, formula, owner, and the conformed table it is computed from. Publish them in the catalog and the semantic layer.
• Retire duplicates. Find every dashboard that computes those metrics independently (INFORMATION_SCHEMA.JOBS will tell you which queries reference which tables) and point them at the governed definition or delete them.
• Add assertions to the conformed layer so a broken definition fails the pipeline rather than quietly changing a number.
• Fill in descriptions for every curated table and column. It is unglamorous work; it is also the single highest-leverage thing you can do for AI-assisted SQL.
• Only then turn on conversational analytics or agent features, scoped to the governed layer.

The result is fewer dashboards, not more — and a warehouse where the next question can be answered by a person or a model with the same, correct number.

Our data governance and compliance practice covers the first two layers; our AI and Gemini services cover the rest. Contact us if you would like to talk through where your estate stands.