Google Next '26 Just Validated Agentic BI Needs a Semantic Layer. The Real Question Is: Whose?

An AI agent that queries a data warehouse directly does not know what "revenue" means at your organization. The agent does not know which customers are classified as inactive, which time grain applies to a given KPI, or which rows a specific analyst is permitted to retrieve. The result is an answer that is statistically generated, structurally plausible, and potentially wrong in ways that are difficult to detect without auditing every query.

According to Strategy Software, an agentic BI semantic layer is a governed abstraction layer that exposes business-defined metrics, dimensions, and access controls to AI agents, ensuring every agent query returns a consistent, auditable answer based on the same logic governing human-facing dashboards.

The semantic layer resolves the context problem. When an agent asks "What was North America's revenue last quarter?", the semantic layer resolves "revenue" to the correct business definition, applies the relevant time grain and filter, enforces row-level security for the requesting user's permissions, and returns a governed answer. That answer matches what a dashboard would show for the same question, to the same user, at the same moment. Without this layer, every AI agent deployment in a large enterprise runs on unverified data interpretations.

Looker's Next '26 announcements validated a position that a smaller number of enterprise data teams had already arrived at independently. Looker BI Agents are built on Looker's semantic layer (LookML), and Google now offers a native MCP server that lets AI agents access governed Looker data through the Model Context Protocol. The Agentic Workflows capability extends this further: agents can monitor metrics, detect anomalies, and surface recommendations autonomously, within the governance boundary of the LookML model.

Three things that Next '26 confirmed:

1. The semantic layer is not optional for enterprise agentic BI. The market has reached consensus. Google's full-event endorsement of semantic-layer-grounded agents closes the debate for most enterprise architects.

2. MCP is the emerging protocol for AI agent access to governed data. Vendors without MCP-native access are now behind. The standard is moving from optional to expected.

3. Organizations with an existing governed semantic layer are better positioned. Retroactively adding governance to an agentic BI deployment is significantly harder than building on a governed foundation from the start.

What Next '26 left open: Looker's agentic BI is architecturally coupled to Looker's semantic layer. For organizations fully committed to Google Cloud with Looker as their primary BI tool, this is a natural and well-supported path. For organizations running multiple BI tools, multiple clouds, or AI agent frameworks outside the Google ecosystem, the semantic layer dependency is a real architectural constraint that deserves examination before committing.

Two distinct approaches to the agentic BI semantic layer are now in wide deployment.

A vendor-native semantic layer is built into a specific BI or data platform. Looker's LookML is vendor-native: it governs data access and metric definitions specifically for Looker, and Looker agents query it through Google's MCP infrastructure. Snowflake's Semantic Views are vendor-native to Snowflake. Databricks Unity Catalog Business Semantics is vendor-native to Databricks. Each performs well within its own ecosystem. Each creates a platform dependency that becomes visible when the rest of the stack is heterogeneous.

A vendor-neutral semantic layer sits above the infrastructure layer and exposes governed metrics to any BI tool, any AI agent framework, and any cloud environment. The semantic definitions are maintained independently of any single platform. A change to a metric definition propagates simultaneously to Power BI, Tableau, Excel, and AI agents without rewriting logic in each tool separately.

Neither is the wrong choice. Vendor-native is simpler if the organization is committed to one cloud provider and one BI platform, and governance requirements are met within that vendor's capabilities. Vendor-neutral is the right architecture for enterprises running heterogeneous stacks, those that want to preserve optionality as the agentic BI market matures rapidly, or those whose governance requirements exceed what a single vendor's layer enforces.

Regardless of which architectural approach an organization selects, an agentic BI semantic layer must satisfy five requirements to be enterprise-grade.

1. Can every agent use the same governed definitions? Every metric, dimension, and attribute must be defined once and consumed consistently by all downstream consumers: dashboards, AI agents, spreadsheets, and APIs. Distributed definitions across tools create metric drift, where "revenue" means different things in different systems, and AI agents surface contradictory answers depending on which system they queried.

2. Is security enforced before the answer is generated? AI agents must not retrieve rows or columns that a given user is not permitted to access. Security applied downstream, at the BI tool or the agent layer, is insufficient: it can be bypassed, inconsistently applied, or absent entirely for novel agent query patterns. Enforcement at the semantic layer closes this gap regardless of which consumer is making the query.

3. Can any MCP-compatible agent access governed metrics? Model Context Protocol is the emerging standard for AI agent communication with governed data layers. A semantic layer that supports MCP natively allows any compatible agent framework to query governed business data without additional translation middleware. Agents built on ChatGPT, Claude, Microsoft Copilot, or Gemini can all operate against the same governed definitions.

4. Can every answer be traced and audited? When an AI agent surfaces a business recommendation, the data team must be able to trace the path: which metric definition was applied, which source system answered the query, and what access controls were enforced. An agentic BI deployment that cannot be audited is not enterprise-ready, regardless of how well the agent performs in testing.

5. Can the semantic layer reach data where it already lives? Enterprise data lives across Snowflake, Databricks, cloud data warehouses, on-premises systems, and operational applications simultaneously. A semantic layer that requires data to be consolidated into one platform before it can be governed delays agentic BI deployment by months. Multi-source federation in place, without data movement, is the requirement.

Strategy Mosaic is a vendor-neutral universal semantic layer built to meet each of the five requirements above across any combination of data sources, BI tools, and AI agent frameworks.

With Mosaic, enterprises do not need to rebuild semantic models every time they add a BI tool, adopt a new AI agent framework, or shift workloads across clouds. The same governed definitions can serve dashboards, spreadsheets, APIs, and AI agents, reducing metric drift, accelerating agent adoption, and preserving architectural optionality.

Mosaic connects to 200+ data sources and exposes governed metrics to AI agents via MCP, ensuring every agent query passes through the same business definitions and row-level security applied to every BI dashboard. The MCP Direct Mosaic Access capability (available in preview as of February 2026) lets AI agents query the semantic layer directly through the MCP server, without a separate agent orchestration layer. Agents built on ChatGPT, Claude, Microsoft Copilot, or Gemini can access the same governed semantic model that powers Strategy One dashboards, from a single definition maintained in one place.

Mosaic Sentinel, the governance module within Mosaic, provides real-time access monitoring, full data lineage tracing, anomaly detection, and audit and compliance reporting across every query, whether that query came from a human analyst or an AI agent. Governance is enforced at the semantic layer, not applied after the fact.

For enterprises already running Power BI or Tableau, Mosaic connects to Power BI through native DAX and XMLA support, and to other BI tools through standard connectors. Adding agentic BI on top of an existing Mosaic deployment does not require rebuilding the semantic model. The definitions already governing dashboards are the same definitions governing the agents.

Google Cloud Next '26 confirmed the category direction: agentic BI needs governed semantics. The next decision is architectural. Enterprises can accept semantics as a feature inside each platform, or they can treat the semantic layer as an independent control plane for trusted intelligence across the entire business.

For organizations that expect their BI tools, cloud platforms, and AI agents to keep evolving, independence matters. That is the role Strategy Mosaic is built to play.