What Anthropic actually built

In its June 2026 post, "How Anthropic enables self-service data analytics with Claude" (authors Josh Cherry, Clement Peng, Johanne Jiao, Justin Leder, and Chen Chang), Anthropic's own data team describes what happened when they pointed Claude at their internal data warehouse to let employees ask analytics questions directly. The result, at first, was not reliable self-service. It was a new failure mode.

The team names the root problem plainly: "with hundreds of viable options in a data model (out of potentially millions of fields), the agent is unable to choose the correct fields that best answer a user's question." They call out three specific failure modes: concept-to-entity ambiguity (many fields could plausibly answer a question, only one is right), data staleness (models and definitions change, the agent's knowledge doesn't), and retrieval failure (the right information exists somewhere but the agent never finds it).

And they're direct about why this isn't primarily a model problem: "An agent that doesn't understand your business will answer what the user asked, but not what they meant. It won't know that 'the Q2 launch' refers to a specific product, that two teams define the same term differently, or that a question is being asked because a board meeting is on Thursday."

That sentence is worth reading twice. It's Anthropic, not a vendor, admitting that model capability runs out well before the actual failure point. What fails is business understanding: definitions, ownership, and shared meaning that no amount of additional training or a longer context window supplies on its own.

The fix Anthropic describes is a semantic layer

Anthropic's answer wasn't a bigger model or a cleverer prompt. It was infrastructure, and the description of it should sound familiar to anyone who has worked on data governance.

At the center sits what they call a governed semantic layer: compiled metric and dimension definitions that serve, in their words, as "the mandatory default path for every data question, same numbers as [the BI tool], joins/grain/filters baked in." Before an agent (or a person) writes raw SQL, it checks the semantic layer first. Underneath sits a data foundations layer of canonical, single-source-of-truth datasets, enforced by tooling and CI checks, with metadata treated as a "first-class product": column descriptions, ownership, and lineage, maintained rather than left to rot.

Above the semantic layer, they add a business context layer: a knowledge graph of company terminology, roadmaps, decision logs, and organizational structure, the layer that resolves whether "the Q2 launch" means one specific product or something else entirely. They also add procedural knowledge, what they call skills, that encode which sources to consult in what order. The numbers are stark: without skills, Claude's accuracy on their internal analytics evals didn't exceed 21 percent. With skills layered on the governed data foundation, accuracy went consistently above 95 percent, and today 95 percent of business analytics queries are automated via Claude at roughly 95 percent aggregate accuracy.

Read that stack back: canonical, governed data, a compiled layer of shared definitions sitting between raw data and any question asked of it, a layer of business terminology and context on top, all connected and kept current. That is a universal semantic layer, built the hard way, because Anthropic needed their own agents to be trustworthy.

The part most companies can't build themselves

Here's the catch. Anthropic could build this because they have the engineering resources to stand up a data foundations layer, a semantic compiler, a business knowledge graph, and a skills framework, then wire CI checks around all of it to keep it from drifting. That's a serious, sustained infrastructure investment, not a weekend project.

Most enterprises don't have that team, or that runway, sitting idle. What they have is exactly the raw material this stack is built on: business people who know what "active customer" means, analysts who know which definition of "profitability" is the right one, domain experts who know the relationships between customer, order, and claim, none of it written down anywhere an agent, or a new hire, could reliably consult. That's the layer a business glossary exists to capture: entities, their attributes, and the relationships between them, defined once by the people who actually understand them, agreed on across business and technical teams, and kept current as the business changes.

Why the business glossary is the backbone

This is the piece Ellie.ai is built around. Ellie's business glossary lets domain experts define entities, attributes, and relationships collaboratively, with folder structures that let different business units maintain their own definitions without stepping on each other, and with a direct line from those conceptual definitions down to logical and physical data models. Through Ellie's MCP server, an AI agent can query that glossary the same way Anthropic's internal agents query their semantic layer, checking a governed definition first, rather than guessing from a schema.

That combination, business-owned definitions, explicit relationships, and a live connection to the data itself, is what makes a semantic layer genuinely universal rather than another siloed catalog. Take the glossary away and you still have a schema. You don't have the meaning an agent, or Anthropic's own data team, needed before their accuracy numbers were worth trusting.

Anthropic spent real engineering effort building a semantic layer because they discovered, the hard way, that agentic AI without one tops out around 21 percent accuracy on real business questions. That's the clearest evidence yet, from the source least likely to overstate it, that a universal semantic layer isn't a nice-to-have layered on top of agentic AI. It's the precondition for it.