Managing Agentic AI Through the Identity Control Plane: What Organizations Should Look for

Securing agentic AI runs through three control planes: identity, which governs who or what may act; data, which governs what an agent may know and touch; and runtime, which governs what an agent may do in the moment. All three matter, but they are not peers. Data controls and runtime controls each depend on knowing, with confidence, which agent is acting. Identity is the control plane that allows the other two to work.

A data permission must be scoped to some actor. A runtime guardrail must bind to some actor. An audit record must name some actor. When the identity beneath those controls is borrowed, shared or unknown, data controls scope to the wrong subject, runtime controls cannot be enforced or tuned per agent and the audit trail attributes actions to the wrong entity. Identity is not merely first among the three; it is the plane on which the other two operate.

All respondents in this IDC study cited non-human identities outpacing their identity and access management (IAM) programs, with ratios cited as high as 75 to 1. Non-human identities (e.g. service accounts, application identities, API keys, machine workloads and AI agents) already outnumber human identities in many environments and agentic AI accelerates the imbalance. Agents will push that figure higher.

IDC research shows that abused non-human identities (e.g. service accounts, API keys and tokens) were the initial entry point in 19.0% of the most recent identity incidents among 651 organizations reporting a confirmed incident, effectively tied with phished or stolen credentials at 19.5%. NHI compromise is no longer a theoretical exposure.

Among the security and identity leaders who participated in this research, there was strong consensus: every agent needs a named human sponsor and a managed life cycle. Where organizations differ most is in how far they have matured in their approach to agent identity, not in the destination they are maturing toward: a purpose-built, distinctly governed identity class for every agent. The variation across organizations is a maturity gap, not a standing disagreement.

The four critical capability areas identified in this IDC research are: (1) Governance, risk and policy framework for agentic AI; (2) AI-driven threat detection and identity threat response; (3) Agent and non-human identity life cycle and ownership; and (4) Privilege right-sizing and just-in-time access. These four areas reached the Critical tier, combining near-universal reach (12–13 out of 13 respondents) with the highest evidence volume.

IDC research identified six consistent pain points: (1) “I cannot see or count my agents” (most organizations cannot state an exact agent count); (2) “Agents break my identity and privileged-access tooling, built for humans”; (3) “Privilege is sprawling at machine speed”; (4) “I can no longer trust the human signal” (deepfake and synthetic identity threats); (5) “Governance and ownership lag adoption and no single vendor covers it”; (6) “There is risk I do not control: shadow AI and vendor-embedded AI.”

IDC’s four-stage maturity model helps organizations locate their current stage and take the next step it indicates, rather than attempting Stage 4 controls from a Stage 1 posture. Organizations converge on procedural governance, like a sponsor and a life cycle for every agent, long before they converge on the architecture of what an agent identity is. The destination is shared; the starting point is not.

GuidePoint sponsored this paper and commissioned the research from IDC. This IDC Whitepaper draws on the IDC qualitative study Converging Identity and AI: a three-session, asynchronous expert panel of 13 senior security and identity leaders (12 full participants, 1 partial) spanning financial services, healthcare, education and care services, technology and media, entertainment and media, hospitality, private-equity-backed enterprise, highly regulated enterprise and enterprise IT. Participants answered 31 open-ended items in their own words across three consecutive daily sessions.