Three Governance Questions Every Enterprise Needs to Answer About Agentic AI

I sat down recently with podcast host Mirko Peters earlier today for a conversation about Microsoft 365 governance, and the discussion kept circling back to the same theme: agentic AI is not a productivity rollout. It is a governance and operational transformation, and the organizations treating it like the next feature drop are setting themselves up for trouble.

While I don’t want to get into everything covered in the podcast (you’ll have to wait for the episode to go live), I did want to provide a quick summary of three of the areas we covered. The full episode will be out soon, but I wanted to share the highlights while the conversation was still fresh in my mind.

The governance concerns that matter most

One of the first questions Mirko asked was my primary governance concerns in the age of AI. I told him that when you start layering Copilot, Copilot Studio, autonomous agents, and workflow orchestration on top of Microsoft 365, the stakes rise quickly — but many of the fundamentals of information management governance remained essential to AI. I narrowed my concerns down to three areas:

• Uncontrolled access to organizational knowledge. Most enterprises have permission sprawl across SharePoint, Teams, OneDrive, and email. Humans discover bad permissions slowly. AI discovers them instantly. The question shifts from “Who has access to this file?” to “What can this agent infer, summarize, and expose across the organization?”
• Autonomous action without enough human oversight. Traditional automation followed rigid rules. Agentic systems are goal-oriented, and once they can create tickets, route approvals, update records, or interact with customers, the risk profile changes completely. The concern is no longer just hallucinations. It is hallucinations that can execute.
• Shadow AI and agent sprawl. This looks a lot like the early days of shadow IT or uncontrolled Power Platform growth. HR builds one agent, marketing builds five, sales builds ten, and nobody knows what exists, who owns it, or what business processes depend on it.

The bigger shift is that AI governance is no longer just an IT problem. It is operational governance, information governance, and risk management all at once. We talked at length about how every part of the organization needs to have ownership of the agents they create and use, not just IT.

Microsoft’s response

Mirko and I spent a good chunk of the conversation on Microsoft’s strategy. What struck me is how clearly Microsoft has framed agents as first-class enterprise actors. Agent 365 is essentially a management and governance control plane for AI agents. Through Microsoft Entra, agents now receive their own identities, lifecycle policies, and conditional access controls.

Microsoft has also started talking openly about “shadow agents,” which is the right move. Discovery and monitoring capabilities are being built into Agent 365, Defender, Intune, Purview, and the Microsoft 365 admin center. The strategic pattern is starting to emerge clearly: Copilot is the experience layer, Copilot Studio is the orchestration layer, and Agent 365 is the governance and operational layer.

That mirrors how Microsoft historically approached endpoint management, identity, and cloud governance. It is a familiar playbook applied to a new kind of digital workforce.

Where this is heading in 18-24 months

When Mirko asked about the longer view, I shared my thoughts around some of the major shifts that I see happening, but I’d like to expand that to 5 things:

• Agent identity becomes a standard part of enterprise architecture, sitting alongside users, devices, apps, and service accounts.
• Governance becomes more centralized, with formal AI governance councils spanning IT, security, legal, compliance, HR, operations, and business leadership.
• The market shifts from “Can we build agents?” to “Can we govern agent ecosystems?”
• Content governance finally becomes a board-level priority, because agents amplify the chaos that humans used to work around.
• Organizations start defining autonomy tiers for AI systems, classifying agents the way they classify applications or data sensitivity: informational, advisory, transactional, and autonomous operational.

The result of all of this is not unrestricted autonomy. This is about creating managed autonomy. And I think that Microsoft’s strategy reflects that pretty clearly, combining identity, governance, observability, policy enforcement, lifecycle management, and human oversight into a unified operational model.

Because the real enterprise challenge is not building one smart agent, but in managing ten thousand of them without losing control of your business.

I will share the link to the episode once it goes live. Thanks again to Mirko for a great conversation.