Microsoft published its April 2026 Copilot Studio update on May 11, shipping agent governance controls, an expanded cost estimator, and MCP server integration for enterprise workflows. The update also coincides with the general availability of Agent 365, Microsoft’s centralized control plane for managing agents across environments.
Governance: Security Posture and Read-Only Analytics
The governance additions target a specific operational pain point: as organizations deploy agents across business units, IT teams lose visibility into what’s running, how it’s secured, and who has access. Copilot Studio now surfaces agent status directly in the authoring experience, showing authentication gaps and policy impacts at the source rather than requiring admins to chase them through separate dashboards.
A new Analytics Viewer role, now generally available, gives business stakeholders read-only access to agent performance data without granting configuration or publishing rights. Mohamed Arhab, Solution Architect at the City of Montreal, described the split: “The Analytics Viewer role allows us to provide meaningful performance insights to business and operational stakeholders while maintaining strict production governance,” according to Microsoft’s blog post.
Agent 365, which reached general availability on May 1, extends this further. It provides a single pane for agent inventory, permissions, behavior monitoring, and activity logging across Copilot Studio agents, Microsoft 365 agents, and partner ecosystem agents. For organizations running agents from multiple sources, this means shared policies and security controls applied consistently rather than per-platform.
Cost Estimator Expands to Dynamics 365
The agent usage estimator now includes Dynamics 365 agents, covering Sales Qualification Agent and Customer Service Agent alongside existing Copilot Studio scenarios. The tool forecasts Copilot credit consumption across both platforms in a single view, letting finance and IT teams model usage before scaling deployments.
This is a direct response to the cost predictability problem that has slowed enterprise agent adoption. When organizations can’t forecast what a fleet of agents will cost to run, procurement stalls at the pilot stage.
MCP Server Integration and Workflow Expansion
Workflows in Copilot Studio can now connect to MCP (Model Context Protocol) server-enabled tools in preview, allowing agents to take actions across external systems while staying within Microsoft’s security and compliance boundaries. A new centralized, admin-controlled environment for the Workflows Agent applies data loss prevention policies consistently across automations.
Agent nodes, which embed Copilot Studio agents directly inside workflow steps, let workflows delegate reasoning and decision-making to an agent at prescribed points while maintaining deterministic structure around them. Unifi, North America’s largest aviation ground handling provider, used this pattern to automate legal contract review by combining agents with deterministic workflows, reducing contract processing from days to minutes according to Microsoft.
Apps in Agents Goes GA
The apps-in-agents capability, which allows agents to surface interactive app experiences directly in Copilot Chat, is now generally available. Agents can pull in Power Apps, Dynamics 365, and partner applications (Adobe Express, Box, Figma, Monday.com, Wix) so users can review data, update records, or approve requests without leaving the conversation.
The Scale Problem Microsoft Is Solving
The April update signals where Microsoft sees the enterprise agent market heading: away from capability demonstrations and toward operational infrastructure. Agent governance, cost forecasting, and centralized policy management unblock production deployments at organizations running dozens or hundreds of agents across business units — the kind of infrastructure that makes IT teams say yes to scale, not just pilots. Microsoft is betting that the next phase of agent adoption depends on whether IT teams can say yes to scale, not whether the underlying model can do the task.