Over the past year I've watched Copilot and autonomous agents spread through Microsoft 365 tenants surprisingly fast, and agent governance turned into a real concern just as quickly. This isn't only an IT matter anymore. Every new agent automates some process, but it also brings new risks, possible data exposure and compliance trouble. In the organizations I work with, the tenants that handle agent governance systematically, from inventory all the way to retirement, avoid the worst pitfalls and the expensive surprises.

Your agent inventory lives in more than one admin portal

The first practical hurdle is building a complete inventory. In the Microsoft 365 admin center, “Manage agents” now shows a consolidated view of centrally registered Copilot agents along with their basic status. What it leaves out are many Power Platform-based agents, especially the ones built with Copilot Studio and spread over several environments. You'll find those in the Power Platform admin center, usually under Environments, but they don't always appear in tenant-wide lists. For my clients I run regular reviews in both portals, and I often use a PowerShell script to pull agent details from every Power Platform environment. You only get a realistic picture once that technical inventory has been matched against what the business teams and owners tell you.

Practical tip: Don't trust the agent list in the Microsoft 365 admin center on its own. If your organization has Power Platform makers, there are probably agents sitting in Copilot Studio environments that it doesn't show. Always cross-check.

Every agent needs an owner, a sponsor and a proper identity

Each agent should have a clearly assigned owner, and ideally a business sponsor too. I regularly come across agents that makers built “in the wild” and that became orphaned once the creator moved to another role or left the company. The result is abandoned bots, gaps in compliance and even billing problems with consumption-based agents. On the technical side, you can configure new Copilot Studio agents with an Entra Agent ID (this requires Entra ID Premium P1), which makes identity lifecycle and permission management much better. Moving existing agents to Entra Agent ID takes deliberate planning, and Microsoft's official guidance covers the details.

Data access and connector policies decide what your agents can actually do

The toughest governance questions are about data. Which content can the agent reach, and which actions may it carry out? Copilot Studio agents give you fine-grained control over knowledge sources and data access, and you should review these regularly in the Copilot Studio environment (my guide on knowledge source permissions goes into this). Allowed actions and connectors are controlled through Power Platform DLP policies, which you set up in the Power Platform admin center → Data policies. In my experience, enforcing DLP policies for Copilot Studio is mandatory for any agent that talks to external tools or APIs. With the connector and action governance improvements from 2025, this has become workable for most organizations.

Before any agent goes to production, I check:

This simple diagram shows how the main governance aspects connect:

flowchart TD
  Inventory["Inventory"] --> Owner["Owner/Sponsor"]
  Owner --> Identity["Agent Identity (Entra Agent ID)"]
  Inventory --> Data["Data Access/Knowledge Sources"]
  Data --> Actions["Actions/Connectors"]
  Actions --> Sharing["Sharing/Limits"]
  Inventory --> Monitoring["Monitoring/Audit"]
  Monitoring --> Lifecycle["Lifecycle (Review/Retire)"]
  Lifecycle --> Billing["Billing"]
  Owner --> Training["User Training"]

Figure: An overview of the main governance aspects for an agent.

Auditing, DSPM for AI and billing reviews keep agents in check over time

For ongoing monitoring, the Purview Audit portal (Compliance portal → Audit search) is now the best place to go, since it logs more and more Copilot and agent actions. If you need deeper visibility, particularly into data leaving sensitive repositories, I now consider DSPM for AI essential. It shows AI-driven data flows, permission use and policy violations in a single view, at a level of detail that larger organizations actually need. Keep an eye on pay-as-you-go agents, because unchecked they can produce unpleasant monthly bills. The relevant portals now let you track costs per agent, but I still recommend a monthly review, especially for agents that call external APIs. And once an agent is no longer needed, make sure it gets decommissioned properly. Abandoned agents are a security risk and a billing risk at the same time.

Agent governance is a team effort and a leadership issue

IT can no longer treat agent governance as an afterthought. Time and again I see that tenants with clear inventories, assigned owners and policies in place have a much easier time with compliance, security and costs. Transparency, clear sponsorship and regular reviews count for more than shiny technical features. Business units also need to be part of governance from the very start. If you want concrete next steps, take a look at agent management in the admin center, agent sharing, knowledge source permissions and my Ignite admin takeaways on Copilot agents. As always, stay practical and stay alert, because governing AI agents is now everyone's job.