IDC white paper: managing AI agents through identity, with a sponsor-and-lifecycle model

In two minutes

• The paper says identity is the control plane that lets data controls and runtime controls work, because both need to know which agent is acting (pp. 3, 9).
• It rests on a panel of 13 senior security and identity leaders, backed by three IDC surveys, and names six problems: agents cannot be counted, human-built IAM and PAM tooling does not fit, privilege sprawls, human signals can be faked, governance lags, and shadow or vendor-embedded AI is outside control (pp. 7, 11-31).
• The one point of strong consensus is that every agent needs a named human sponsor and a managed lifecycle. Panel differences are framed as a maturity gap, not a disagreement (pp. 4, 32, 35).
• It proposes a definition-agnostic operating model built on an agent registry: intake gate, named sponsor, privilege bound to the grant, just-in-time access, revocation, audit, containment, and risk-tiered autonomy (pp. 37-42, Table 2).
• It splits 'agent identity' in two: knowing who each agent is can start now, while how agents prove identity (SPIFFE/SPIRE, IdP account types) can reasonably wait for standards to settle (p. 39).
• It offers a four-stage maturity model and an 18-area vendor demand map. Four areas are tiered Critical: governance framework, identity threat detection and response, agent lifecycle and ownership, and privilege right-sizing with just-in-time access (pp. 36, 46-48).
• It maps the model to the NIST AI RMF, the OWASP NHI and Agentic Top 10 lists, the CSA AI Controls Matrix and EU AI Act Article 12 (pp. 5-6, 38).

Key numbers

• 19.0% vs 19.5%: Abused non-human identities versus phished or stolen credentials as initial entry point in recent confirmed identity incidents (651 organizations) (p. 9).
• 77.3% vs 18.5%: Organizations highly confident they see all identities, versus those running discovery continuously or near real time (p. 11).
• 41.5%: Organizations whose NHI programs secure bots, RPA and AI agents, the least covered NHI type (pp. 11-12).
• 27.3%: Organizations with fully automated identity governance and access controls for AI agents (n = 860) (p. 16, figure 1).
• 18.0%: Organizations enforcing least privilege through just-in-time access with automated remediation (n = 829) (pp. 18-19, figure 2).
• 97.5% (67.9% actively): Organizations discussing giving agents more autonomy (n = 1,967) (p. 19).
• 14.5%: Highest share naming one role for AI governance ownership (dedicated AI governance team); no single role reaches 15% (p. 25).
• 47.4%: Organizations with an IAM failure or identity compromise recovery plan in place (n = 860) (p. 44, figure 8).

What to do with it

• Run automated discovery and build an authoritative agent inventory first, including agents running under user credentials, as the paper recommends (pp. 14, 49).
• Set up an agent registry with sponsor, purpose, scoped access, expiry and revocation for every agent, without waiting for architecture decisions (pp. 39, 42, 49).
• Replace vault-and-checkout PAM for agents with short-lived, task-scoped credentials, keeping governed standing access only where legacy systems require it (pp. 14, 17).
• Settle internal ownership of AI identity governance early, for example through an AI governance board (pp. 28, 49).
• Locate your organization on the four-stage maturity model and take the next step it lists (pp. 36, 49).
• Put the paper's three vendor questions to candidates: what can you prove in a pilot, what does exit look like, and who runs it if the team cannot (p. 46).

Read the full document if

• you must write or defend an agent identity and governance program and want a registry-based operating model (Table 2, p. 42).
• you are preparing vendor requirements and want the 18-area demand map and tiers (Table 4, pp. 47-48).
• you need survey figures on agent governance maturity, least privilege and AI governance ownership (pp. 16-31).

What it relies on

The core findings come from an IDC qualitative panel, "Converging Identity and AI": 13 senior security and identity leaders, 31 open-ended questions, coded responses (p. 51). Three IDC surveys add figures: the IAM Security Survey (May 2026, n = 860), the Unified AI Platforms and Governance Survey (March 2026, n = 744) and the Agentic AI Functional Use Case Study (February 2026, n = 1,967). The panel is small, and its maturity stages hold 2 to 5 people each (p. 36). The paper cites NIST, OWASP, CSA and EU AI Act sources. It is sponsored by GuidePoint Security, a cybersecurity services firm, and the paper recommends managed services and vendor support (pp. 43-44, 53). IDC states its endorsement is not implied (p. 54).

Our read

Useful as a structured argument and a vendor-requirements checklist for agent identity. The registry model and the split between knowing each agent and proving its identity are the most reusable parts. Panel evidence is small and qualitative, so treat the survey figures as the quantitative anchor.