TL;DR The agent itself isn’t the core issue. The problem is the context it consumes. Point a strong LLM at a data warehouse without a governed business context and you’ll get confident, fluent answers that are partially or completely wrong. A semantic layer for AI actually has two distinct context layers that most discussions blur together: a semantic model (what the data represents — entities, relationships, descriptions, synonyms) and metric definitions (how each business metric is precisely computed). They serve different purposes, and you need both. Governance is a third layer that sits on top of semantics. Without clear ownership, verification, lineage, and access controls tied to every metric, an agent can retrieve a perfectly valid definition of “revenue” and still choose the wrong one — because nothing indicates which definition the business officially endorses. The real moat is whether your semantic layer can be consumed from every surface — dashboards, agents, APIs, CLIs — instead of being trapped inside a single BI tool. As LLM capabilities become commoditized, the real differentiator is the governed context beneath them. Everyone is racing to drop an AI agent on top of their data. Almost nobody is asking whether that agent can be trusted to interpret what it sees. That sequence is backwards. And the common attempts to fix it — larger context windows, more complex reasoning, another eval — miss the point. The generative model stopped being the hardest part of agentic analytics months ago. Hooking an LLM up to a warehouse is a weekend task. The hard part is…