Google Put the Metric Definition in the Agent’s Path. Analytics May Stop Arguing With Itself.

Google Cloud has made expanded semantic-model support generally available in Knowledge Catalog, allowing AI agents to retrieve approved metric formulas, filters, join paths, lineage and data-quality signals across six major data systems. The immediate commercial consequence is that an agent answering a revenue, conversion or customer-value question can use the same business logic as analysts instead of inventing a plausible query. Companies must now decide which definitions are authoritative—and whether putting them directly in an agent’s path can turn analytics governance into an operating advantage.

The capability is generally available

Google announced the release on October 29. Knowledge Catalog can now ingest semantic models and metric definitions from Looker LookML, BigQuery Graph, dbt MetricFlow, Snowflake Semantic Views, Databricks Metric Views and SAP Core Schema Notation. It connects those definitions to schemas, lineage, data profiles, glossary terms and quality signals in one context graph without moving the underlying data or rewriting the business logic in another language.

Gemini Enterprise, BigQuery Conversational Analytics and custom agents can query that graph. In practical terms, an agent can move from a phrase such as “net revenue” to the approved formula, mandatory filters, verified join path and physical tables that produce the number.

This is a deployed, generally available capability, not a research preview. However, Knowledge Studio—the natural-language workspace for exploring and packaging the graph—is still in preview. Google’s claim that grounding agents this way makes them nearly four times more accurate, up to 33% faster and up to 75% less token-intensive comes from internal benchmarks and has not been independently validated.

A metric can now travel with the question

Businesses already spend heavily defining measures in analytics systems. The failure occurs when those definitions do not reach the AI answering the next executive, customer or campaign question.

For marketing, the difference is concrete. “Conversion” may mean an order in one dashboard, a qualified lead in another and any completed form in a third. “Customer value” may include refunds, service costs or only booked revenue. An agent that selects tables by linguistic similarity can produce an elegant answer that is mathematically correct and commercially wrong.

Knowledge Catalog changes the retrieval step. The agent does not merely find data; it can retrieve the rules that make the data meaningful. That can shorten the distance between a question and a governed answer, while reducing the repeated analyst work required to explain which table, filter and denominator count.

The new possibility is a shared metric contract across human dashboards and agentic work. A planning agent, reporting assistant and campaign-analysis system could all begin with the same definition even when the underlying semantics originate in different platforms. Google documents this cross-platform ingestion; consistent downstream decisions remain a possibility businesses must prove in their own environments.

One graph does not create one truth

The capability exposes disagreement as well as agreement. Google says overlapping definitions remain side by side with their provenance rather than silently overwriting one another. That is safer than choosing invisibly, but it still leaves a human decision: which definition governs which business question?

Independent data-governance analysis from Atlan makes the distinction sharper. A cloud catalog can be authoritative for what a platform observes and enforces, while the definition a business actually approved may live in a review, policy or contract elsewhere. If two systems define revenue differently, giving an agent access to both does not resolve the conflict. It makes precedence a design requirement.

Self-learning context creates a second risk. Google says Gemini can enrich the graph from playbooks, meeting notes, chat threads and code, then keep context current as definitions change. That may reduce documentation decay, but generated context can also formalize ambiguous language unless owners review it, track provenance and control promotion into production.

This is the kind of distinction OrionPilot is built to surface: a faster answer is not automatically a better business decision. The operating discipline is to connect every important metric to an owner, an approved meaning and the action it is permitted to influence. AI can then accelerate analysis without quietly changing what success means.

What businesses should measure next

First, test accuracy on real questions, not demonstrations. Compare grounded and ungrounded agents across revenue, acquisition cost, retention and campaign performance, then record false joins, outdated definitions and disagreements with audited reports.

Second, measure the operating cost of context. Google reports lower latency and token use, but companies should validate those gains across mixed Looker, dbt, Snowflake, Databricks, SAP and BigQuery environments, including permission checks and stale-source failures.

Third, watch how conflicts are resolved. The commercially decisive evidence will be whether teams can assign authority, approve changes and trace every answer back to the definition and data quality state used at that moment.

The breakthrough is not that an agent can read more metadata. It is that business meaning can accompany the question all the way to execution. If the definitions are governed, that can make AI answers operationally trustworthy. If they are not, the context graph will scale disagreement with impressive efficiency.