Fabric Data Agent has two meters—and background smoothing is not a spending limit

Microsoft Fabric Data Agent does not have one cost meter. It has at least two: the tokens consumed while the agent interprets a request and generates a response, and any SQL or other query work that the resulting plan executes. Microsoft’s Data Agent consumption documentation says generated queries are billed separately to the corresponding engine—for example, a warehouse query appears under the SQL Query operation.

That distinction changes how teams should run a pilot. A prompt that looks inexpensive at the agent layer can still trigger material warehouse work, while a cheap query can sit behind a large prompt assembled from instructions, examples, schema context and conversation history. Microsoft says all of those inputs contribute to token usage; the user’s question can be only a fraction of the billable input context in the same documentation.

Read both receipts

The published model rates are 100 CU seconds per 1,000 input tokens, 10 CU seconds for cached input, and 400 CU seconds per 1,000 output tokens. Microsoft’s worked example prices a request with 2,000 input tokens and 500 output tokens at 400 CU seconds, or 6.67 CU minutes, before query execution. Usage appears in the Capacity Metrics app as AI Query under the LlmPlugin item kind, according to Microsoft’s billing page.

The operational lesson is to record two numbers for every acceptance-test question: AI Query consumption and the cost of the generated workload on its target engine. Aggregate capacity charts alone cannot tell a team whether a regression came from larger context, more verbose answers or heavier SQL.

Smoothing is a scheduling rule, not a budget guardrail

Microsoft classifies the agent’s AI activity as a background job. Its example says a 6.67-CU-minute request is smoothed so that it consumes one CU minute in each hour; on an F64 capacity, the documentation estimates more than 13,824 such requests before exhaustion. It also warns that once the capacity is exhausted, all operations shut down. The broader Fabric operations catalog likewise lists AI query as a background operation on the Copilot and AI billing meter.

Teams should therefore test burst behavior, not just daily averages. A useful rollout gate is to replay a representative hour of questions, capture both meters, then model the same load with cache misses and larger instruction sets. Add an alert well before capacity exhaustion and a per-user or per-workload admission rule outside the smoothing mechanism.

There is a third boundary for regulated deployments: Microsoft says cross-geo processing can place Azure OpenAI work outside a user’s geographic region, compliance boundary or national-cloud instance when the tenant setting is enabled, while billing remains tied to the billing region. That makes the cost test inseparable from a data-residency review in the Data Agent consumption documentation.