Google Cloud Moves AI Agent Telemetry Into Customer-Owned BigQuery

On November 13, 2026, Google Cloud moved the monitoring and evaluation of enterprise AI agents into BigQuery tables that customers themselves own — a design choice that quietly shifts governance of workplace AI from vendor dashboards to existing warehouse permissions.

For HR and IT leaders weighing agent deployments, the practical consequence is concrete: telemetry, cost controls and evaluation results now sit inside the same identity-and-access-management perimeter as payroll and people data, rather than in a separate SaaS console. That consolidation matters as HR functions adopt agents for tasks such as policy Q&A, where auditability is not optional.

What changed, exactly?

Google Cloud expanded its agent monitoring stack across BigQuery and Looker for two populations: managed data agents (Conversational Analytics) and custom agents built with the Agent Development Kit (ADK) or LangGraph. Both feed a single telemetry path. Managed agents emit telemetry automatically; custom agents send structured telemetry through single-line SDK instrumentation.

Availability splits into two tiers:

• Generally available: BigQuery Studio's prebuilt dashboards tracking usage, active sessions, top knowledge sources and latency, plus Looker System Activity dashboards for Conversational Analytics in both Looker (Google Cloud core) and Looker (original), covering engagement and token usage.
• Preview: user ratings and query review in the Responses & feedback tab; daily token quotas; native first-party agent evaluations via API and a guided Colab notebook.

Can you monitor agents without reading employee conversations?

Largely, yes — and that is the sharpest privacy feature here. Google Cloud deliberately separates operational telemetry from conversation content. Cloud Monitoring supplies aggregate metrics (request volume, latency, token usage); Cloud Trace provides step-by-step spans per request. Both exclude customer core content: user prompts, model responses, internal reasoning and generated SQL.

The result resembles a facilities system that records how often a meeting room is used without recording what was said. Platform teams get visibility; nobody gets automatic access to the substance of an interaction.

Full conversation logging is opt-in, per agent, and available only to agent owners and editors. When enabled, a Google-managed service agent streams prompts, responses and turn-level telemetry through the BigQuery Storage Write API into a customer-owned agent_events table. The table inherits the dataset's IAM permissions, customer-managed encryption keys and VPC Service Controls — the standard BigQuery security stack, applied unchanged.

Privacy guardrails run per turn. If a knowledge source uses row-level security, column-level policy tags or dynamic data masking, logging skips that turn. If a logging error occurs, the user's chat session continues uninterrupted.

The asymmetry cuts both ways: custom agents using the open-source ADK plugin can retain full tool-invocation trees, time-series telemetry and multimodal payloads, while first-party conversation logging strips low-level tool arguments and internal SQL.

Do the cost controls actually bite?

They extend past reporting into enforcement. Every logged conversation turn records input, output, thought, cached and total token counts alongside total turn latency. Looker System Activity separately reports estimated input and output tokens across BI workloads.

Two active limits pair with that reporting:

• Per-query byte-scan caps — generally available, using dry-run estimates to block oversized BigQuery queries before on-demand scan charges accrue.
• Daily input and output token quotas — in preview for BigQuery and the Conversational Analytics API, applicable per project, user or agent through the console or gcloud CLI.

Token quotas are the more interesting control for HR deployments: a per-agent cap turns a runaway benefits-FAQ bot from a budget surprise into a bounded event.

How are agents tested before changes reach employees?

Google Cloud's open-source Prism tool evaluates Conversational Analytics agents grounded in BigQuery tables or Looker Explores. It supports reusable test suites, regular-expression assertions and multi-region agent discovery, focusing on SQL accuracy and answer quality before prompt, model or schema changes ship to users. Completed evaluation runs land in day-partitioned BigQuery tables, sitting beside production agent_events records for analysis in BigQuery and Looker.

For teams planning production rollouts, the design decisions are now explicit: where to draw the line between operational traces and full conversations, and which controls are GA versus still in preview.