Databricks Unity Catalog: A Universal Semantic Layer

TL;DR: At Data + AI Summit 2025, Databricks extended Unity Catalog into a universal semantic layer for data and AI. The headline change is the Databricks Unity Catalog universal semantic layer, where business metrics become first-class assets defined once in the catalog and reused everywhere (Databricks, 2025). The signal for the rest of the industry is bigger than one product: the catalog is becoming the governance core for AI, not just a place to track tables.

For years, the catalog was plumbing. It tracked where tables lived, who could read them, and how data moved. Useful, but invisible to most of the people asking business questions. The 2025 Unity Catalog announcements change that role. Databricks is pushing business meaning, not just data location, into the catalog. That shift matters because AI agents are about to become major consumers of enterprise data, and they need to know what a number means before they can answer with it.

What is the Databricks Unity Catalog universal semantic layer?

A semantic layer is a shared definition of business concepts, like metrics, dimensions, and the relationships between them, that sits between raw tables and the tools querying them. It answers the question “what counts as revenue here?” once, so every report and every model agrees.

The new piece is Unity Catalog Metrics, in Public Preview on AWS, Azure, and GCP at announcement and reaching general availability later in 2025 (Databricks, 2025). It makes business metrics first-class assets in the lakehouse. Instead of defining “monthly active users” inside one dashboard, you define it in the catalog, and it becomes reusable across AI/BI Dashboards, Genie, notebooks, SQL, and Lakeflow jobs.

The design choice that matters: these metrics live at the data layer, not the BI layer. Databricks describes the contrast directly. Metrics defined only in the BI layer limit reuse and integration, while defining them in the catalog makes the semantics reusable across every workload (Databricks, 2025). The metrics are also fully addressable via SQL, so the definition does not depend on which tool you happen to open.

Why move metrics out of the BI tool?

Inconsistent metric definitions are an old, expensive problem. When “active customer” means one thing in the finance dashboard and another in the sales report, leadership spends meetings reconciling numbers instead of acting on them. Databricks names this directly: inconsistent definitions across tools and teams cause confusion, misalignment, and a lack of trust in data (Databricks, 2025).

The traditional fix was a BI tool’s own semantic layer. It works, until you use a second tool. A definition built in Power BI is locked in Power BI; it is not available in SQL or another platform. Databricks contrasts its approach against exactly this: unlike proprietary BI semantic layers, Unity Catalog Metrics are addressable via SQL so everyone sees the same metric regardless of the tool they choose (Databricks, 2025).

There is a customer voice attached to the claim. Richard Masters, VP of Data and AI at Virgin Atlantic, said Unity Catalog Metrics gives the company a central place to define business KPIs and standardize semantics across teams, so everyone works from the same trusted definitions across dashboards, SQL, and AI applications (Databricks, 2025).

How does this connect to AI agents?

Here is the part that explains why a metrics feature shipped inside a governance product. AI agents answer questions by querying data they do not inherently understand. Ask an agent “how did churn trend last quarter?” and it has to know which table, which filter, and which formula counts as churn. Get any of that wrong and the answer is confidently incorrect.

A catalog-level semantic layer closes that gap. The Databricks documentation is explicit that standardized metric definitions give AI tools the context they need to interpret data accurately, with governance enforced through Unity Catalog (Databricks Documentation, 2026). The same docs introduce agent metadata: synonyms, display names, and formatting rules that help AI tools read data in business terms rather than raw column names.

Databricks also built the metrics into its AI surfaces so an agent and a dashboard draw from one definition. In the company’s own framing, the catalog now connects everything through a single open layer, whether you are building AI agents, delivering BI dashboards, or sharing data across organizations (Databricks, 2025). The governance does not get bolted on after the AI; it travels with the data into the AI.

The bigger signal: catalogs are becoming the AI governance core

Step back from the feature list and the pattern is clear. Databricks is repositioning the catalog from a passive inventory into the place where business meaning, access control, and AI context all live together. The 2025 release pairs the semantic layer with a curated internal Discover marketplace of certified data products organized by business domain, plus attribute-based access control and automated data quality monitoring (Databricks, 2025).

Read together, those pieces describe a single idea: the catalog is where an organization decides what data means, who can use it, and whether it can be trusted, for humans and agents alike. That is a different job than the catalog had five years ago.

This is the same conviction behind SemanticOS, with one important difference in scope. A catalog like Unity Catalog governs structured assets inside one platform: tables, metrics, dashboards. Most institutional knowledge lives outside it, in Slack threads, contracts, wikis, and tickets. SemanticOS builds a knowledge-graph and AI-search layer across those fragmented tools so people and AI agents can find and reason over institutional knowledge, not just structured rows. Databricks extending Unity Catalog toward business semantics is a strong signal that the connective, meaning-bearing layer, rather than raw storage, is where the next decade of enterprise AI gets decided.

A concrete example: one definition, every surface

Picture Vantage Health, a mid-size health insurer. Its analysts had defined “active member” three different ways across three dashboards. Finance counted anyone with a paid premium that month. Operations counted anyone with a claim in the last 90 days. The customer team counted anyone who had logged into the portal. Every quarterly review opened with a fifteen-minute argument about whose number was right.

Vantage Health moves the definition into Unity Catalog Metrics. “Active member” becomes one governed metric, addressable via SQL, with lineage and auditing attached. The finance dashboard, the operations report, and the analyst’s notebook all pull the same definition.

The payoff lands when Vantage Health turns on an AI assistant for its care managers. A care manager asks, in plain language, how many active members in a region missed a follow-up. Because the agent reads the same catalog metric, plus the agent metadata that maps “active member” to the governed definition, it returns the number that matches the official dashboard. No reconciliation meeting. The semantic layer did the agreeing in advance.

What the catalog still cannot answer is why the definition changed last spring, or which regulatory note prompted it. That reasoning lives in a compliance memo and a Slack thread, outside the lakehouse. Connecting those sources back to the metric is the job of a knowledge-graph layer that sits above the individual tools.

Key takeaways

• The Databricks Unity Catalog universal semantic layer makes business metrics first-class assets defined once in the catalog and reused across dashboards, SQL, notebooks, and AI tools (Databricks, 2025).
• Defining metrics at the data layer instead of inside a BI tool keeps one trusted definition usable across every workload and external engine (Databricks, 2025).
• Governed metrics give AI agents the business context to answer questions consistently, with governance enforced through the catalog (Databricks Documentation, 2026).
• The broader signal is that catalogs are becoming the governance core for AI, where meaning, access, and trust live together.
• A catalog governs structured assets within one platform; a knowledge graph like SemanticOS connects the surrounding context across many tools so people and agents can reason over knowledge no catalog holds.