BigQuery’s data agents add a prompt-data control that administrators cannot leave to users

The control hiding behind the chat interface

Google’s current BigQuery conversational-analytics documentation says the service analyzes anonymized prompt data across customer interactions to improve quality, accuracy and reliability. The processing is enabled by default, although Google says the data is not used to train or fine-tune its foundation models. That distinction matters, but it does not eliminate the governance decision: organizations still need to decide whether prompt ingestion fits their policy before users begin chatting with warehouse data. Google Cloud

Until November 23, 2026, customers that want to opt out must collect project numbers—or an organization ID for all existing BigQuery projects—and submit Google’s opt-out form. On and after that date, administrators can instead apply the hierarchical custom constraint constraints/gcp.disablePromptAnalytics at the organization, folder or project level. Re-enabling analytics requires removing that constraint if it was applied. Google Cloud

Why this belongs in the deployment checklist

BigQuery data agents are not generic chatbots. Administrators select tables, views, graphs and user-defined functions as knowledge sources, then add business context through metadata, instructions, verified queries and glossary terms. Conversations persist, and responses can include generated SQL, charts, reasoning and metadata about the sources used. Google also warns that direct conversations without an agent’s configured context can be less accurate. Google Cloud

That design makes prompts operational data. A question may contain business terminology, filters, customer categories or investigative hypotheses even when the underlying table permissions remain intact. Google’s anonymization statement reduces one class of concern; the default-on ingestion setting creates a separate administrative choice that should be recorded alongside source selection, IAM and verified-query review. Google Cloud

The practical mistake would be to treat the November policy constraint as a privacy toggle each analyst can choose. It is a resource-hierarchy control. Platform teams should therefore inventory projects where conversational analytics is enabled, decide centrally whether prompt analytics is permitted, and document the temporary form-based path if they need an opt-out before November 23. After that date, policy-as-code can make the decision consistent across folders and projects. Google Cloud

Capability is expanding faster than the checklist

A new Adswerve implementation overview shows why this control may become relevant quickly: it frames BigQuery conversational analytics as a way for business users to obtain answers, visualizations and multi-step analysis from selected warehouse sources, while object tables and supported AI functions extend questions to unstructured material. The partner’s central point is that agent configuration—not fluent output—is what separates a grounded analyst from a confident chatbot. Adswerve

The same principle applies to prompt handling. Before expanding access, an enterprise rollout needs two explicit configurations: what the agent may query, and whether the prompts used to query it may enter Google’s cross-customer analytics process. BigQuery now exposes both decisions, but only one appears in the agent-building flow. The other belongs with the administrators who govern the project hierarchy.