Managing Agentic AI: Why the Control Plane Problem Is an AI Problem

Organizations face a significant challenge in governing AI agents because visibility into these systems is often lacking. A core issue is the inability to count or track the massive number of AI agents operating in production, as continuous discovery across IT and OT environments is not widely implemented. AI agents are deployed at scale, executing tasks and making decisions without human intervention, leading to a governance gap despite existing AI adoption.
A recent study by IDC found that fewer than one in five organizations run continuous discovery on their AI agents. The need for clear control structures stems from the fact that agentic AI involves three interconnected control planes: identity (who acts), data (what is touched), and runtime (what actions occur). The necessity of having an agent-specific identity is foundational for enforcing least privilege, auditing actions, and containing incidents.
The research identified several pain points related to managing agentic AI through the Identity Control Plane. Organizations model agents differently—some treat them as distinct identities, others reuse existing service accounts, and some omit identity management entirely. The study suggests a four-stage maturity model for agentic AI, where current organizations are generally in stages 1 or 2, and the expectation is that every agent requires a named human sponsor and managed lifecycle.

* Fewer than 1 in 5 organizations run continuous discovery on their agents.
* Just over 1 in 4 organizations have fully automated governance for AI agents.
* AI agents are executing tasks, accessing data, making decisions, and chaining actions at machine speed.
* Most organizations cannot confidently determine the number of agents in production.
* The IDC research identified six new pain points related to managing agentic AI through the Identity Control Plane.
* Identity is posited as the control plane upon which data and runtime rely for enforcing least privilege, auditing actions, and incident containment.
* Agentic AI maturity is modeled across four stages: today most organizations are at stages 1 or 2.
* The prerequisite for effective agent management involves an agent having its own scoped identity.

The narrative structure relies on juxtaposing the massive, uncontrolled growth of operational AI agents with the inherent difficulty of applying traditional governance models to unseen entities. The pattern suggests that as technology capability accelerates (agents operating at machine speed), the corresponding organizational governance infrastructure lags significantly behind. This creates a gap where trust in opaque systems is implicitly accepted until a concrete failure occurs, which is then framed as an identity management problem rather than an operational oversight.
The focus on identity as the central control plane—requiring agents to have distinct identities separate from human or service accounts—highlights a fundamental cognitive shift required for securing autonomous systems. The divergence in how organizations model agents (treating them as distinct entities versus reusing existing credentials) is not merely a technical choice but reflects differing levels of organizational maturity and the willingness to acknowledge non-human actors within established security paradigms.
The underlying implication is that risk management in the age of agentic AI moves from perimeter defense to internal, granular identity control. The challenge for leadership is shifting from asking "How do we secure the network?" to "How do we define, discover, and constrain the operational boundaries of every autonomous actor?" This requires a paradigm shift where visibility, ownership, and governance are established concurrently with deployment, rather than as retroactive fixes.
Bridge Questions: If identity is the control plane, what verifiable mechanisms exist today for establishing agent identities across disparate IT and OT environments? What specific metrics or frameworks can effectively map an organization's current stage on the agentic AI maturity curve? How should organizations account for the systemic risk associated with relying on non-human credentials during incident response?