Gemini can now query BigQuery, Spanner and Looker as the user. Decide which questions need a certified agent

A regional sales director asks Gemini how last quarter’s margin compares with the plan. Until recently, the honest answer depended on whether someone had copied the right tables into an index the assistant could search, and when.

Google’s Data Cloud connectors, announced in preview on November 6, remove the copy. Gemini can now query BigQuery, the operational databases Spanner, AlloyDB and Cloud SQL, and Looker directly, in natural language, under the identity of the person asking.

How the connectors work

The connectors run in federated mode: they query data where it lives instead of ingesting or indexing it. A question goes through four steps. Google’s Knowledge Catalog finds the relevant data, maps the question to tables and business definitions, generates SQL and runs it with the user’s credentials, so existing row-level and column-level permissions apply. The catalog only reveals metadata the user is allowed to see.

Admins switch the connectors on in the Gemini admin console. Connections are read-only by default, and admins can set a default Google Cloud billing project for the BigQuery queries that Gemini runs. The BigQuery connector also reaches Apache Iceberg, Delta Lake and Parquet tables, and datasets held in AWS or Azure.

Two days earlier, Google said that its Knowledge Catalog made agents nearly four times more accurate, with up to 33% lower latency and up to 75% fewer tokens. Those are Google’s internal benchmarks, and they describe the catalog’s curated context, not the connectors in any particular company.

Two ways to let an assistant answer from company data

	Copy and index	Query in place, as the user
Where the data sits	A second copy in a search index	Only in the source system
Who may see what	Rebuilt in the index, often coarser	The source system's own row and column rules
Freshness	As of the last sync	As of the query
Main failure	Stale or overexposed copy	Wrong SQL on a correct, current table
Cost	Indexing and storage	Each question runs a query

Permissions solved, meaning not yet

Querying as the user fixes the problem that made many enterprise assistants unsafe: an index that showed everyone what only some people should see. That is a real improvement, and the same direction we noted when AWS moved AI apps onto live governed data.

It moves the risk to a different place. With live data and correct permissions, the remaining error is an answer that is confidently computed from the wrong join, the wrong definition of margin or the wrong fiscal calendar. That error is harder to spot, because the number looks authoritative.

Google’s own design recognizes this. It offers two paths. Direct connectors serve ad hoc questions. Conversational Analytics agents, built and certified by data teams for a defined scope, serve analyses that need to be right every time, and can be published in Gemini or exposed through an API.

Decide the split before people start asking

The useful governance decision is not whether to turn connectors on, but which questions belong on which path:

• Ad hoc is fine for exploration, sizing and questions whose answer a person will check before using.
• Certified agents are required for any number that goes into a board pack, a regulatory report, a price or a forecast. These should rest on trusted semantic definitions and verified queries that the data team owns.
• Billing needs an owner. Every question becomes a query. Setting the billing project is easy; deciding who reviews the bill when a popular question is asked a thousand times a day is the part teams forget.

The connectors are in preview, so this is the time to write that split down. Once a thousand employees are asking Gemini about margins, the habits will already be set.