Analysis entered the conversation

Microsoft’s August update said Copilot Chat and Cowork could reason over Power BI reports and semantic models available to the user. The capability moved from public preview in June to worldwide rollout in August. Project teams could ask about schedule or cost without leaving their workflow.

Chats and individual responses could be shared as read-only links, and notebooks synchronised between the Copilot app and OneNote. Planner Agent could produce status reports in Frontier. Analysis and collaboration moved closer together, while report, filter, time and draft status still needed to be visible.

Action needed more than a good answer

Gartner advised on 20 August that a first agent pilot should be treated as a governance pilot. It stressed data access, allowed actions and human review before development. Although aimed at finance, the principle applied to project work where incorrect tasks or budget changes could spread quickly.

The project office could let Copilot explain a variance but require approval before Planner created or moved a task. The model owner owned the definition, the project leader the priority and the task owner the delivery. Automation then had an explicit chain of accountability.

A decision card preserved context

For every AI-assisted variance, the card could retain the question, Power BI report and filter, refresh time, answer, proposed task and approver. The team measured wrong answers, changes after refresh, reversed tasks and time to actual resolution. This connected insight with project outcomes without hiding uncertainty inside shared chat.

Controls in practice

Access to a semantic model did not mean every measure suited every question. Model owners needed to mark certified measures, business definitions and known limitations. If Copilot combined measures from different reports, its response had to show the relationship and warn when periods or units were not comparable.

Shared chat created a practical review path. A colleague could inspect the question and response without editing the original, while still opening the report to verify figures. The project could require this check before a proposal became a task and record both analyst and approver.

Sources

Microsoft 365 Copilot Blog: What’s new in Microsoft Copilot – August 2026, 31 August 2026.

Gartner: Pilot governance before scaling AI agents, 20 August 2026.

Microsoft Dynamics 365 Blog: One always-on AI at Work roadmap, 25 August 2026.

Gartner: Customers require access to a human agent, 4 August 2026.

For discussion

Which analytical finding may create a project task, and who approves the action?