Fabric data agents have a capacity meter—but operators must join it to warehouse query costs

Microsoft’s Fabric documentation now gives administrators a concrete way to find data-agent consumption in the Capacity Metrics app: agent activity appears as an AI query operation under the LlmPlugin item kind, billed through the Copilot and AI meter and classified as a background operation. That is useful visibility, but it is not a complete per-question bill. Microsoft’s operations reference separately lists warehouse and SQL-endpoint queries as Data Warehouse operations.

One question can cross multiple meters

A Fabric data agent can choose among lakehouses, warehouses, semantic models, KQL databases, ontologies and Microsoft Graph, then generate and execute SQL, DAX, KQL or Graph queries. Microsoft says the agent uses the requesting user’s credentials, validates the generated query and keeps data-source connections read-only. It can attach up to five sources. Those product mechanics mean a single natural-language question can create at least two operational records: the LLM-side AI query and the downstream workload query selected by the agent. Microsoft’s data-agent architecture describes that routing and execution path.

For capacity owners, the practical consequence is that looking only at LlmPlugin understates the workload generated by an agent. Conversely, a rise in warehouse-query consumption cannot automatically be attributed to human-authored SQL. Chargeback and regression analysis need a correlation layer that joins agent interactions to the data-source operations they initiated.

Background smoothing changes the alerting window

Microsoft classifies Copilot operations as background jobs. Its consumption guide says background usage is smoothed across 24 hours, while interactive usage is smoothed over at least five minutes; the first throttling phase begins after a capacity has consumed the resources available for the next ten minutes. The same guide prices input, cached-input and output tokens at different CU-second rates and warns that broad deployment without measuring a pilot can exhaust capacity. The billing guide provides the rates and smoothing behavior.

That makes daily capacity headroom—not just prompt latency—the right first operating signal. Teams should baseline both LlmPlugin AI-query usage and the relevant warehouse, semantic-model or KQL meters during a controlled pilot, then test whether their attribution survives multi-source routing. The new meter identifies the reasoning layer; it does not eliminate the need to account for execution separately.