Your AI agents are ready, but your identity infrastructure is not

Enterprise AI programs stall at the same place. The models are capable, and the use cases are defined. The business case is approved, but then the security review starts, and months later the agent is still in pilot.

Most organizations blame the security team, the security team blames the tools, and both diagnoses miss the point.

Governance architecture is the bottleneck. Until enterprise leaders recognize the difference between limitations in technical capability and gaps in appropriate governance, they will continue to invest in AI programs that are unable to move from experimentation to production.

You already have the tools

Walk into any large enterprise and you will find vaults, PAM platforms, and IAM systems built over years of compliance-driven investment. The infrastructure exists, and the spending has been consistent.

The governance layer is the critical missing piece providing the structure needed for these tools to operate on a machine scale. Non-human identities now outnumber employees 144 to 1 in cloud-native environments, growing at over 40 percent per year [1]. Tools answer the question of how to store and rotate a credential. Governance answers the harder questions: who owns this identity, what is it permitted to do, and when the agent using it is decommissioned, does anyone know to revoke it?

CyberArk's 2025 machine identity survey found that the top blockers which enterprises face are organizational, not technical, including identifying the business owner of a machine identity, locating where it is in use, and understanding who is authorized to access it [2]. No vault solves those problems.

AI agents make this load bearing

A traditional service account does one predictable thing: an AI agent reasons about its access needs, chains tool calls across systems, and generates thousands of API calls in the time it takes a human to complete a single authentication prompt [3]. Giving that agent a shared service account with broad permissions, because it is the path of least resistance, creates both a security risk and an accountability void. When the agent executes an unauthorized action, attributing responsibility between the human authorizer and the autonomous system becomes legally and operationally impossible.

80 percent of organizations already report that their AI agents have acted outside their intended scope [4]. Only 18 percent of security leaders are confident their current IAM can manage agent identities at all [5]. The emerging governance standard has shifted from least privilege to least agency; governance has restricted how much an agent can decide and act on independently beyond just what data it can access [6]. That is a harder problem, and it requires architectural thinking rather than tool deployment.

The market has already priced in these developments

SailPoint revised its total addressable market (TAM) estimate to 90 billion dollars in June 2026, up from 55 billion, citing agentic use cases as the driver [7]. Further, Zscaler's AI Protect suite generated over 100 million dollars in bookings in its first year [8].

Regulators are moving in parallel. DORA became applicable across European financial services in January 2025. The EU AI Act’s enforcement timeline now extends through August 2028. Over 60 percent of financial services AI risk control objectives depend directly on foundational identity security [9]. The compliance window for planning has closed.

What fixes it

The enterprises closing this gap are building a governance model with five components:

1. A canonical inventory of every non-human identity and its owner
2. A provisioning standard that issues unique identities to new agents rather than sharing existing ones
3. A lifecycle discipline that rotates, reviews, and decommissions on a defined schedule
4. An observability layer that makes agent actions attributable and revocable in real time
5. A cross-functional ownership structure that closes the gap between security, platform engineering, and AI program teams.

This is an architecture you design and a capability you build over time. That is precisely what makes it a durable advantage for the organizations that do it well and an escalating liability for those that do not.

Enterprise leaders face a key question: do we solve this before the incident, or after?

Sources

This source list is illustrative, not comprehensive. It surfaces the primary references used to support specific factual claims in the article and does not represent the full body of research underlying BCE's analysis. Sources were drawn from four categories: (1) vendor-published annual threat and market reports (CyberArk, Entro Labs, SailPoint); (2) independent analyst and research organization surveys (Cloud Security Alliance, Forrester); (3) company financial disclosures and investor communications (Zscaler, SailPoint, CyberArk); and (4) official regulatory instruments (EU Parliament). Where a statistic was synthesized across multiple inputs or inferred from directional survey data, the closest primary source is cited. Readers conducting independent due diligence should verify figures directly against the original publications, which may have been updated since the time of research.

[1] Entro Security, Non-Human Identity Security Report, H1 2025. Machine identity growth rates and NHI-to-human ratios in cloud-native environments.

[2] CyberArk, Identity Security Threat Landscape Report, 2025. Annual survey of enterprise security practitioners on machine identity management and deployment blockers.

[3] Anthropic, Zero Trust Principles for AI Agents, 2025. Technical guidance on agent authentication patterns and API call volume in autonomous workflows.

[4] Cloud Security Alliance (CSA), AI Safety and Security Report, 2026. Survey of enterprise AI deployment teams on agent behavior and scope management.

[5] J.P. Morgan Research, Brian Essex, CFA, State of Security: Key Cybersecurity Topics and Metrics, May 6, 2025.

[6] Anthropic, Zero Trust Principles for AI Agents, 2025; Forrester Research, AEGIS: The Agentic Enterprise Governance and Identity Security Framework, 2025. Conceptual framing of least agency as a successor standard to least privilege.

[7] SailPoint Technologies, Investor Day Presentation, June 2026. Revised total addressable market figures and growth rationale citing agentic identity demand.

[8] Zscaler, FY2025 Earnings Call Transcript and Investor Materials, 2025. First-year revenue and bookings figures for the AI Protect product line.

[9] SailPoint, Inc., Investor Day Presentation 2026, June 2026 See source [7]