Administration followed growing agent use

Microsoft's June update for Microsoft 365 Copilot described refreshed usage reporting and agent administration, including cost controls and better visibility into who used which agents. ContextIQ was intended to surface relevant organisational context in the flow of work. Project leaders could therefore connect agent usage with teams, activities and cost instead of treating it as one aggregate Copilot licence.

The controls only created value when they answered a governance question. A project could distinguish between an agent that summarised meeting notes, one that retrieved internal content and one that executed an action. These activities carried different risks and needed different boundaries for data access, budget and approval. Usage volume alone did not show whether work became faster or a decision improved.

The analytical model became part of the control surface

The Power BI update on 10 June emphasised secure collaboration on sensitive data, sensitivity labels and tenant controls for Azure Maps. It also expanded visual calculations, web editing in Power Query and Copilot in reports distributed through organisational apps. Analysis could move closer to the collaboration surface, while classification and model ownership still had to travel with the report.

For a project office, permission to view data was different from permission to interpret it. A user might be allowed to see a report even though an agent should not search the entire model or combine data in new ways. Model owners therefore needed to document measures, definitions and limits, while report owners controlled distribution. A Copilot summary should point back to the report and filter so its claim could be checked.

Gartner published a forecast on 11 June that guardian agents could account for 10–15 per cent of the agentic-AI market by 2030. In a webinar poll, 24 per cent of 147 IT leaders had deployed a few agents and four per cent more than a dozen. The sample was small and the prediction uncertain, but the finding illustrated the need for controls that follow actions while they occur rather than an audit long afterwards.

Three logs had to be read together

A practical governance model could connect the project log, agent log and analytical evidence. The project log recorded a decision and accountable person; the agent log recorded request, sources and action; the report recorded model, filter and time. A weekly sample review could reveal excessive access, costly agent runs and decisions without evidence. Automation then became measurable as a work process, while collaborators could inspect how each conclusion was reached.

Sources

Microsoft 365 Copilot Blog: What’s new in Microsoft 365 Copilot – June 2025, 30 June 2025.

Microsoft Power BI Blog: Power BI June 2025 Feature Summary, 10 June 2025.

Gartner: Guardian agents and controls for agentic AI, 11 June 2025.

Microsoft: 2025 Work Trend Index Annual Report, 23 April 2025.

For discussion

Which agent actions in a project require prior approval, continuous monitoring or review afterwards?