Semantic Layer vs. Data Catalog for AI: Why Metadata Isn't Meaning

Early in my career as a healthcare consultant (pre-Gen AI, to date myself), I was trying to prove that our software was working. The pitch was simple: we help health systems keep referrals in-network, and my job was to show that the numbers backed that up.

I had a referral report pulled from the health system and was cross-referencing it against ICD codes grouped by specialty. In-network referrals to gastroenterology came back at 90% in the system report. When I grouped the codes myself, I got 70%. Same data, same question, two different answers. I was new enough to analytics that I spent a long time assuming I was making the mistake.

The system had been lumping certain oncology codes into the gastro bucket because stomach cancer lives in that overlap. The system knew the label “gastroenterology,” but it didn’t know what we meant by it.

The problem is not niche. The problem is the core structural flaw sitting underneath most enterprise AI deployments today, and it has a name: metadata is not meaning.

A semantic layer is an abstraction layer between raw data sources and analytics tools that translates technical data structures into business-understandable terms, ensuring consistent metric definitions across every report, dashboard, and AI query. A true semantic layer is an execution engine, not a label store: it doesn’t describe what “gastroenterology” means and hope the model interprets it correctly. Strategy Mosaic mathematically governs the definition before the AI ever touches the data.

Most enterprise AI stacks do not have a semantic layer. They have metadata: column names, lineage, tags, and descriptions that describe data without governing how it is computed. When an AI model encounters that metadata, it does what models do: reads the description and guesses the intent.

A true semantic layer removes the guess. Business logic, what counts as revenue, how churn is calculated, which ICD codes belong to gastroenterology, is encoded and enforced at the semantic layer, so every downstream tool and AI agent computes from the same governed definition.

We are currently watching a massive rebranding exercise. Overnight, every data catalog, orchestrator, and governance tool has slapped a new label on their homepage, claiming to be “The Context Layer for AI.”

The pitch is compelling. AI models hallucinate because they lack enterprise context, so traditional catalog providers step in and say: we have your metadata, we know your data lineage, we are the context layer.

A catalog provides structural awareness but no business logic. The AI fills that gap by guessing. And because the model guesses instantly and confidently, the wrong answer is polished, formatted, and halfway to a board presentation before anyone thinks to question it.

Catalog vendors know this is a vulnerability, which is why we are suddenly seeing a rush to bolt “semantic features” onto existing data dictionaries. Tag your metrics with definitions to bridge the gap.

The real difference between annotating a catalog and building actual knowledge architecture is execution.

Bolted-on semantic features are just richer metadata. They tell the AI about the rule. A true semantic layer is an execution engine: it doesn’t hold the definition of “churn” and hope the model interprets it correctly, it mathematically governs the calculation before the AI ever touches it. If your semantic capabilities are an afterthought layered over a catalog, your AI is still doing the heavy lifting, translating definitions into data engineering on the fly, and guessing when definitions are incomplete.

The gastro referral problem is exactly this. The system had labels. It had lineage. What it didn’t have was a governed definition that said: gastroenterology means these codes, not those ones, regardless of how the model interprets the word.

Beyond accuracy, the absence of a semantic layer is a CFO-level financial risk that organizations haven’t priced in yet.

Every time an LLM has to reason through raw tables to figure out a metric, you are paying for the privilege. You are burning expensive API tokens to make a language model write SQL and guess at business rules your infrastructure should have already solved.

In Strategy Software’s testing, when an LLM generated complex queries against raw data in PostgreSQL, it returned the correct answer 30% of the time. When the same LLM was routed through Strategy Mosaic’s semantic layer, accuracy reached 100%. And because Mosaic handed the model a pre-calculated, governed answer instead of forcing it to reason through the math, token usage dropped by up to 50%.

An LLM making 1,000 analytics queries per day without a semantic layer is spending half its compute budget guessing. Routing those queries through a governed semantic layer improves the answer and can halve the bill.

Strategy Mosaic, Strategy Software’s universal semantic layer, is built on the principle that business logic should be defined once and enforced everywhere, not described in a catalog and re-interpreted by every tool that reads it.

Four specific capabilities close the gap between metadata and meaning:

1. Governed definitions enforced centrally. Row-level and column-level security, metric definitions, and business logic are embedded directly in the Mosaic semantic model. Every connected tool like Power BI, Tableau, Excel, AI agents computes from the same governed definition automatically. If gastroenterology means these ICD codes and not those, that definition is set in Mosaic once and applied everywhere, without duplication or drift.

2. AI agents grounded in semantics, not raw SQL. Strategy Mosaic exposes governed business semantics to AI agents through the Mosaic MCP integration. Agents operating through Mosaic work with authoritative metrics and defined relationships, not raw tables. Conversational analytics queries return accurate answers because the semantic model governs what each metric means before the model is asked anything.

3. Connectivity across 200+ data sources. Mosaic connects to 200+ data sources and applies consistent governance across all of them, without requiring data movement or a centralized warehouse. The semantic layer sits across your existing data infrastructure, not on top of a copy of it.

4. Infrastructure cost control. Mosaic’s in-memory engine selectively accelerates high-frequency workloads, reducing compute costs at the underlying data platform. Organizations running high-concurrency AI workloads see reduced Snowflake and Databricks compute bills because Mosaic serves governed, pre-calculated answers rather than pushing raw queries to the warehouse for every request.

The architecture question for enterprise AI teams is direct: are your AI models getting raw tables and guessing at business logic, or governed, pre-calculated answers built on definitions that reflect how your organization thinks? The first is a data dictionary with an AI pitch and the second is a semantic layer.

AI agents are only as reliable as the business logic they run on. If that logic lives in a label, you have a data dictionary. If it lives in an execution engine, you have a semantic layer and a meaningfully different AI outcome.

Strategy Software builds Strategy Mosaic as the governed foundation for enterprise AI precisely because of situations like the gastro referral problem. Not because catalog vendors are dishonest, but because metadata, no matter how rich, is not a substitute for governed meaning.