Data Governance and the Semantic Layer: Why They Must Work Together

Quick Answer

• Governance documentation alone cannot enforce definitions where queries actually run. AI agents querying raw data without governed semantics produce errors documentation cannot prevent — only computation-layer enforcement can.

• A semantic layer enforces business rules at query time, not just in documentation. Unlike a data catalog, a semantic layer runs definitions like refund exclusions and currency conversions automatically every time a query executes.

• Governed semantics are a prerequisite for production-ready AI, not a phase-two addition. Inconsistent definitions, unauthorized agent queries, and unauditable outputs cause most enterprise AI initiatives to stall before reaching production.

Most governance programs are failing AI initiatives. Not because they're badly run, but because they're governing the wrong thing.

The frameworks are sound: defined data owners, business glossaries, quality processes, catalog entries. Data teams have spent years building these foundations. But they share a flaw that didn't matter much when humans were in the loop and now matters enormously: the governance logic lives in documentation, not in computation. It describes what data should mean. It doesn't enforce it where queries run.

According to a recent Gartner data summit, only 6% of data and analytics leaders consider their organization fully AI-ready. That number makes more sense when you consider that most governance programs were built for a world where an analyst could ask a colleague to clarify a definition. AI agents can't make that call.

When an AI agent queries your data without governed semantics, it does its best with what it has. That's the problem.

An LLM asked to calculate company revenue with no governed context returned a figure off by 77% in a documented example. Three failures at once: revenues aggregated across currencies without conversion, refunds not excluded, rows double-counted because VAT rates had changed mid-year. None of these were data quality issues. All three were governance rules that existed in documentation but had no presence in the computation layer.

The human in the loop used to catch these. A data analyst could ask a colleague. An AI agent can't. As more enterprise workflows move toward autonomous query and answer, the gap between what governance says and what the system does becomes the gap between trusted AI and wrong AI.

"Governance should happen where the computation happens, not in documentation away from it. That's exactly what creates the gap."

— Charlotte Ledoux, Data Governance Consultant and Author, The Data Governance Playbook

That observation is the core argument. Everything else follows from it.

A semantic layer like Strategy Mosaic is an abstraction layer that converts technical data structures into business-understandable terms, ensuring that every metric, KPI, and dimension is defined once and applied consistently across every analytics tool, AI agent, and application connected to your data platform.

In practice: one definition of "revenue" generates the same SQL query whether Power BI, Tableau, an AI agent, or a custom application is asking for it. The calculation doesn't drift based on who wrote the query. When a data steward updates a definition, that change propagates automatically and every downstream consumer gets the updated calculation without anyone manually correcting dashboards.

This is the real difference between a semantic layer and a data catalog. The catalog documents what data means. The semantic layer enforces it. A catalog entry that says "revenue excludes refunds" has no effect on a query. A semantic layer that encodes that rule runs it every time.

We see this pattern constantly. In a recent Diageo finance proof-of-concept, the hardest part wasn't connecting the data. It was getting every tool to agree on the same KPI definitions. Once Strategy Mosaic enforced a single metric definition per KPI across all connected tools, the downstream effects were measurable: data access that previously required two weeks was available within hours, and 41% of compute costs were eliminated by deflecting predictable query patterns away from the database.

Strategy built Mosaic around this principle: governance and semantics should live in the same system. That means definitions, access controls, audit trails, and query generation all operate from the same layer rather than being distributed across separate tools.

Mosaic Sentinel runs the governance layer on top of this: real-time alerts on unauthorized or anomalous data access, a complete audit trail of every change and query, and usage insights across all connected analytics surfaces. These operate at the semantic layer, not downstream in individual tools, so data stewards get one governance view rather than checking each system separately.

For AI agents specifically, the semantic layer becomes the control surface. An agent operating through Strategy Mosaic can only query what it's authorized to access, using the same definitions as every other consumer. The currency conversion, refund exclusion, and VAT handling the Diageo example required are all encoded in the semantic layer and run automatically, there's no version of the 77% revenue error where those rules are in place.

The costs of not connecting governance to computation are real. They just don't show up as line items.

Teams spend days confirming a number is right before presenting it to leadership. Not because the data is missing, but because no one is sure which calculation to trust. Those interruptions compound. When sales and finance produce different revenue figures from the same underlying data, meetings turn into arbitration sessions. The root cause is almost never that two teams chose different definitions on purpose. It's that no one ever enforced a single one at the point of computation. A workshop to align doesn't fix this permanently. Encoding the agreed definition into the semantic layer does.

The AI cost is less visible but growing. A significant share of enterprise AI initiatives stall before reaching production because the data foundation isn't ready: inconsistent definitions in training data, agents querying what they shouldn't, outputs that can't be audited or explained. Governed semantics aren't a phase-two addition to an AI roadmap. They're the prerequisite.