Enterprise AI projects often fail at a deceptively basic task: turning a plain-language question into a number the business can trust. Data may be available in a warehouse, but revenue definitions, fiscal calendars, joins and filters frequently live in a separate semantic layer or in analysts’ working knowledge. That gap creates a governance risk when generative AI agents are asked to query production information.
Google Cloud, part of Alphabet Inc. (Nasdaq: GOOGL), has made LookML semantic-language indexing in Knowledge Catalog generally available for Looker Core on Google Cloud. The integration is designed to give Gemini Enterprise and custom AI agents access to the business logic behind a metric alongside broader catalog information such as schemas, lineage and governance policies.
LookML is the modeling language used in Looker to define measures, dimensions, relationships and calculations. Google said the service now automatically indexes resolved LookML models, Explores, Views, lineage and dashboards into Knowledge Catalog. Rather than only finding a table or dashboard, an agent can retrieve the definition of a metric, its join path and the tables from which it originates.
The distinction matters because data discovery and deterministic business semantics are different jobs. An agent working directly from raw tables may infer a revenue calculation, a date range or a relationship between tables incorrectly. Google cited common risks including schema hallucinations, incorrect accounting conditions and join fan-outs across data with different levels of detail. Those errors can yield a plausible response while still producing a number that conflicts with finance or operational reporting.
Google said synchronization is enabled by default and updates pushed to production LookML are reflected in Knowledge Catalog in near real time. The company also said the indexed material is immediately available to Gemini Enterprise without custom agent code or an additional catalog data store. Developers building their own agents can access the same context through the LookupContext API, according to the company.
The expected workflow combines Knowledge Catalog’s policy and discovery information with Looker’s explicit calculations. In a question about quarterly revenue by product category, for example, an agent could retrieve applicable policy documentation and data-governance labels while using the LookML definition to apply the appropriate fiscal calendar, completed-order condition and approved joins. Google said a resulting answer can include an Open in Looker Explore link so a user can inspect the measures, filters and underlying line items.
That does not remove the need for validation. Indexing business logic makes an agent’s context richer, but organizations still need to manage who can access data, approve the models and definitions that become available, and test the accuracy of responses against established reporting. Live synchronization also means changes to production models deserve the same change-control discipline they already require for dashboards and analytics.
For data leaders, the practical benefit is a way to reuse the semantic work already invested in governed BI instead of asking AI tools to reconstruct it from database metadata. Legacy text-to-SQL approaches often treated the warehouse schema as sufficient context and left teams to diagnose inaccurate results after the fact. Google Cloud is positioning the semantic layer as an operational control for AI agents; its value will depend on whether enterprises keep that layer current, well governed and connected to the business definitions their users recognize.