Significant increase in Data Agent AI Query CU consumption since October 29

Hi PedroSergey10,

That article is based on a Microsoft message center post (MC1469820, Updates to Consumption Measurement for AI Capabilities in Microsoft Fabric). The updated version says ontologies move first, and Copilot, data agents and AI Functions follow later this year, with a separate notice before that happens. So a later start for data agents on your capacity is possible, but I haven't seen that follow-up notice published yet.

A few things I'd check:

1. Capacity Metrics app, filter to the LlmPlugin item kind and the AI Query operation, then drill into timepoints before and after the jump. That shows if it's more requests or more CU per request.
2. Conversation history on the Foundry side. Long threads get sent back as context and push input tokens up.
3. Ask your M365 admin to look in Message center for that post and any follow-up. That's where the effective date shows.

How it's metered today: https://learn.microsoft.com/en-us/fabric/fundamentals/data-agent-consumption

If cost per request clearly went up with nothing changed, I'd raise a support case with those timepoints. Curious what you find.

Hi PedroSergey10​ 

Thanks for sharing the details.

Microsoft's Data Agent consumption documentation confirms that AI Query consumption is based on the number of tokens processed. This includes not only the user's prompt but also supporting context such as Data Agent instructions, data-source instructions, example queries, and conversation history. Therefore, even if the agent configuration has not changed, the token consumption per request can vary depending on the context processed.

The documented consumption rates are:

• Input prompt: 100 CU seconds per 1,000 tokens
• Cached input prompt: 10 CU seconds per 1,000 tokens
• Output completion: 400 CU seconds per 1,000 tokens

Additionally, queries generated by the Data Agent are billed separately through the corresponding query engine. For example, Warehouse query execution is reported under the SQL Query operation.

To investigate the increase observed since October 29, I recommend:

1. Open the Microsoft Fabric Capacity Metrics app and filter for the LlmPlugin item kind and AI Query operation.
2. Compare the usage before and after October 29. Check whether the increase is associated with more requests, higher CU consumption per request, or both.
3. Compare the request patterns, conversation history, and context being sent to the agent. Longer conversation histories or larger supporting context can increase input token consumption.
4. Review the corresponding query-engine operations separately to determine whether any additional consumption is coming from generated query execution.
5. Check your Microsoft 365 admin center Message Center and the official Microsoft Fabric release notes for any announcement about changes to Data Agent consumption rates.

The Bismart article by itself does not confirm that the described consumption-measurement change applies to Fabric Data Agents. I would avoid attributing the October 29 increase to that change unless Microsoft has published a specific announcement or provided confirmation.

If the increase remains unexplained, I recommend raising a Microsoft support request with the capacity details, affected workspace and Data Agent IDs, exact UTC time range, and Capacity Metrics operation data.

Thank you.

Hi PedroSergey10​,

The checks in the earlier reply from ShafiUllah-bc​ are the right starting point. I have not seen the follow-up notice either, so I cannot say whether the metering change has reached your capacity. Two additions that may help you separate "more usage" from "more cost per request":

Capture before opening a case
- Capacity Metrics: AI Query CU per day for Oct 27-31, requests per day, CU per request
- Same query shape on both sides of the jump (same agent, same question set)
- Foundry side: conversation length / input tokens per call for the same period
- Capacity name, region and the exact timepoints where the jump starts

Contain while investigating
- Capacity alerts or a CU budget threshold for the workspace
- Keep agent instructions and data source scope tight so each call sends less context

If requests per day are flat and CU per request jumped on a specific timepoint with no configuration change, that is the evidence that makes a support case actionable.

Hi jubinsoni​ ShafiUllah-bc​ v-csrikanth​

Thank you all for your comments.

I would like to reiterate a key point from my original post: we have made no changes to our prompts, agent configuration or workflow.

We have also already compared the same query executed before and after October 29, and confirmed a significant increase in AI Query CU consumption per execution. The suggested checks to establish whether consumption per query has increased have therefore already been completed.

We would appreciate guidance that addresses this specific finding. Could Microsoft confirm whether any change to Data Agent consumption measurement or other service-side change around October 29 could explain the increase?

In particular, has the consumption-measurement change mentioned above been applied to Data Agents, and how can we verify whether and when it became effective for our capacity?

Thank you.