Conversational Analytics in BigQuery and Looker: How to Build a Data Agent

Google's conversational analytics lets people ask data questions in natural language and get an answer computed by a query. It is built around data agents: a set of sources plus the context that tells the model what the data means. As of October 2026, according to Google Cloud's release notes:

• Looker: generally available since 12 November 2025 (Looker release notes).
• Conversational Analytics API: generally available for BigQuery and Looker since 23 June 2026 (API release notes).
• BigQuery: generally available since 1 July 2026 (Google Cloud blog).
• Data Studio (Looker Studio until April 2026): generally available since 30 July 2026 (Data Studio release notes).

This is the technical guide: setup, cost and data preparation. To see whether it fits your sales team, start with conversational analytics.

According to the BigQuery documentation, the agent generates and runs SQL, answers with text or charts, and can call BigQuery's AI functions for forecasting, anomaly detection or key drivers. It cannot write data (no DML or remote functions), and GQL graph queries need an Enterprise or Enterprise Plus reservation.

Location and data residency

An agent lives in the US multi-region, the EU multi-region or Global, and the location cannot be changed after saving; agents created before 4 June 2026 are Global. Google says it guarantees data residency at rest and for ML processing in the EU and US multi-regions (Google Cloud blog). Prompts are not used to train its foundation models, but anonymized prompt data is analyzed by default to improve the product, and you can opt out.

In Looker the agent does not write free SQL: it queries Explores, so it follows the joins, filters and permissions in LookML and reuses your metrics. It is the most governed option, provided the LookML is clean. According to the Looker documentation:

• An agent can use up to five Explores, and each query returns at most 50,000 rows.
• Users need access_data on the underlying models; sharing an agent does not grant it.
• An admin has to enable Gemini in Looker. It works on Looker (Google Cloud core) and Looker (original).
• Advanced Analytics, formerly Code Interpreter, runs Python for analyses SQL cannot do. It has been generally available since May 2026.
• Data at rest stays in the Looker instance; ML runs on a global service unless you use the EU localized processing option added in August 2026 (Looker release notes).

Verified queries, also called golden queries, have been generally available since August 2026. Google suggests no more than 30 to 50, hiding keys and technical fields, and preferring verified queries over long free-text instructions, which go stale as the model changes.

Since April 2026, agents are not created in Data Studio: they are built in BigQuery and published there (Data Studio release notes). Data Studio is the window and BigQuery the engine, so context, verified queries and cost limits all live on the BigQuery agent. If your dashboards read summary tables, point the agent at the same ones and both will give the same number.

The API puts a data agent inside your own app or the chat tool your team already uses. According to the API overview:

• It is served from geminidataanalytics.googleapis.com, with v1 endpoints for BigQuery, Looker and Data Studio, and a Python SDK.
• You create dataAgents with their context, and conversations to chat with them. In stateful chat Google stores the history; in stateless chat your app sends it each time.
• Responses include the generated SQL and chart specifications your app can render. Access is controlled with IAM.

The v1 release brought data residency, CMEK, VPC Service Controls and audit logs; AlloyDB, Cloud SQL and Spanner as sources are still in preview (API release notes).