Skip to content
BI Pixie Workload in Microsoft Fabric
Generally Available

AI adoption

When people ask AI instead of opening a report, report views fall even though people still use the data. The AI adoption tab counts the questions people ask AI. It answers the question: how many people ask Copilot, Fabric data agents and other AI applications questions, and which semantic models those answers come from?

The tab is in the BI Pixie AI Readiness Dashboard, right after Overview. It appears only when Admin Monitoring is on, because BI Pixie learns who asked AI from your tenant's activity. To turn it on, see Admin Monitoring. The Workspace and Dates (UTC) filters at the top carry onto the tab as onto every other page.

Choosing which AI the page counts

Above the cards, the slicer AI this page counts offers three choices. It applies to every card and visual on the tab.

  • Data agents only counts the people who asked a Fabric data agent a question, wherever they asked it from.
  • Copilot only counts the people who asked Copilot in Power BI or Microsoft 365 Copilot a question.
  • Both counts all of the above, and also the people who asked Claude or another AI application a question about your semantic models.

A person counts once a day, however many questions they ask. Service principals and benchmark runs are left out, so the counts are people.

The cards

The cards cover the last 30 days and compare them with the 30 days before, so the Dates (UTC) filter does not move them. The line under each card name says what it counts under the slicer choice.

  • Daily Active Agents is the average number of people a day who asked AI a question.
  • Monthly Active Agents counts the people who asked AI a question in the last 30 days.
  • MAU counts the people who used a report with Pixies or asked AI a question in the last 30 days.
  • MAA/MAU is Monthly Active Agents as a share of MAU, with both counts under it, for example "41 of 228 people". It is the share of active people who asked AI.

While BI Pixie has counted AI questions for less than 60 days, the cards read, for example, "Counting began Oct 2, 2026" in place of the comparison, because the earlier 30 days are not fully counted yet.

Below the cards

  • People using reports and AI each day draws two lines over the Dates (UTC) range: the people who used a report with Pixies and the people who asked AI. As the gap closes, more of the use moves from reports to AI.
  • Where people ask AI counts the people in the last 30 days by where they asked: Copilot in Power BI, Fabric data agents, Microsoft 365 Copilot, and Other AI applications. Other AI applications shows only under Both.
  • Semantic models AI relies on lists the semantic models a Fabric data agent is built on and the semantic models Copilot was asked about, with their report users, AI users and AI share of use, the AI users as a share of everyone who used the semantic model. The Readiness Score and Readiness band columns come from your AI Readiness assessments. A semantic model nobody assessed has no score.
  • Fabric data agents in use lists every Fabric data agent in the tenant with its workspace, the people who asked it, its questions and its Data agent score. The score is empty for a Fabric data agent nobody assessed. Test chat questions counts the questions asked in the data agent's test chat while it was built. A Fabric data agent nobody asked in the range is listed with 0.

When nobody asked AI in the period, the tables say so for the slicer choice, for example "No one asked a Fabric data agent a question in this period."

What to do with it

  • Read a fall in report use on the other pages against MAA/MAU. If the share of people who ask AI rises as report use falls, people are moving to AI, not leaving.
  • When a semantic model has many AI users and a low AI Readiness Score, the most people get the least reliable answers from it. Improve that semantic model first, with the work list in its AI Readiness assessment.
  • A Fabric data agent with many people and an empty Data agent score has never been assessed. Assess it to learn whether its answers can be trusted.
  • A Fabric data agent with questions only in its test chat was built but never used. Ask its owner whether it is ready to share, or whether to retire it.