Most identities in an enterprise now satisfy none of them. OWASP’s NHI Top 10 puts non-human identities at somewhere between 45 and 100 times human identities in enterprise environments, with a global average nearer 144 to 1 once cloud-native deployments and smaller firms enter the count. Those ranges differ because they measure different populations, so treat the spread as the finding rather than picking a single number.
Service accounts do not change roles. OAuth grants have no manager. An AI agent acquires permissions at runtime and may spawn sub-agents that acquire more.
NIST’s National Cybersecurity Center of Excellence named the gap plainly in a concept paper published on 5 February 2026: AI agents are commonly treated as generic service accounts with no dedicated identity, authorization, or accountability controls.
A checklist for evaluating identity governance solutions still needs everything it always covered. It also needs to establish whether the platform can govern the majority population rather than the familiar one.
Does the Platform See Everything?
Start with discovery, because nothing downstream works without it.
Enterprise access rarely lives in one system. A single employee holds accounts in an identity provider, Active Directory, cloud applications, databases, development tools, and assorted business systems. A platform should consolidate that rather than making administrators inspect each environment separately.
Relationships matter more than inventories. Teams need to see which users have access, what permissions they hold, how those permissions were granted, and which resources they can reach. Identity graph modeling helps here, mapping connections between people, accounts, applications, roles, and entitlements in ways disconnected system reports rarely reveal. Linx describes automatic mapping of identities, entitlements, and access relationships across an environment as part of that approach.
Extend the discovery question to non-human identities explicitly. Ask whether the platform finds service accounts, API keys, OAuth grants, machine certificates, and agent credentials, and whether it can attribute an owner to each. Provisioning increasingly happens outside IT, since purpose-built AI tools now cover narrow tasks the general platforms handled poorly and a department lead can connect one to a business system in an afternoon.
Can It Handle Multiple Access Models?
Most enterprises run several at once, and a platform should not force convergence.
Role-based control defines permissions around job functions. Attribute-based rules account for department, location, or employment type. Just-in-time access reduces standing privilege by granting elevation only when needed.
Just-in-time deserves particular weight now. Palo Alto Networks reported organizations continuing to rely on permanent privileged access rather than time-bound controls, while expecting AI agent growth of 85% over twelve months. Standing privilege compounds faster when the population holding it grows at that rate.
A baseline capability checklist:
- Identity and entitlement discovery across cloud, on-premises, and SaaS
- Support for role-based, attribute-based, and just-in-time models
- Access certification and review workflows
- Automated policy enforcement and least-privilege controls
- Risk-based prioritization of excessive or inappropriate permissions
- Integration with existing identity, security, and business systems
- Automated remediation and auditable governance records
- Discovery, ownership, and lifecycle handling for non-human identities
Features demonstrate well and deploy badly. A capability that impresses in a product walkthrough may deliver little if it cannot handle your identity sources, applications, or approval structures.
Do Access Reviews Give Reviewers Enough to Decide?
This is where most certification programs quietly fail.
Managers receiving long permission lists without context approve access because they cannot determine why it exists or whether it remains appropriate. The certification completes. Nothing gets governed.
A platform should supply the identity involved, the resource, the privilege level, relevant relationships, and enough surrounding information to support an actual decision. Risk-based prioritization makes the exercise manageable by directing attention toward higher-risk entitlements instead of treating every permission equally. Linx highlights contextual certification, automated recommendations, and in-platform remediation as part of its governance model.
Non-human identities break certification in a specific way worth raising with any vendor. Ask who certifies an agent’s access, and against what. There is no job description to compare entitlements against, no manager with firsthand knowledge, and no annual review conversation. Some platforms answer this with ownership attribution and usage-based evidence. Others have not addressed it at all.
Usage data helps considerably here. Sysdig’s analysis suggests both human and machine identities use roughly 2% of the permissions they hold, which means actual access patterns provide a far stronger certification signal than a static entitlement list.
Can It Enforce Least Privilege, Not Just Report On It?
Governance that only produces reports leaves the security team doing the work manually.
Look for automatic identification of ungoverned access, excessive permissions, and policy violations, paired with workflows that remove unnecessary access or trigger remediation without repeated manual intervention.
Control over that automation matters as much as the automation. Enterprises need clear rules about which changes happen automatically and which require human approval, backed by audit trails demonstrating what changed, why, and who authorized it. Organizations running autonomous systems well build fixed human decision points into the workflow, the approach government agencies took with their agentic AI deployments.
The non-human side needs enforcement most, because it accumulates fastest and gets reviewed least. GitGuardian’s State of Secrets Sprawl 2026 found 64% of secrets confirmed valid back in 2022 still unrevoked by January 2026. Nothing about that requires new technology to fix. It requires something that notices.
Credential behavior also differs in ways that affect detection. Machine identities bypass MFA by design, operate continuously, and produce none of the location, device, or timing signals human compromise detection relies on. Agent traffic sits outside assumptions older monitoring was built around, so quiet dashboards prove less than they used to.
What Should You Ask About Agent Identity Specifically?
Standards work now gives buyers concrete language for this conversation.
NIST’s Center for AI Standards and Innovation launched its AI Agent Standards Initiative on 17 February 2026, covering authentication, authorization, auditing, non-repudiation, and prompt injection mitigation. The companion NCCoE project proposes adapting existing protocols, including OAuth 2.0 and 2.1, OIDC, and SPIFFE, for non-human agent workloads, alongside zero trust architecture and federated identity approaches. COSAiS is developing SP 800-53 control overlays covering single-agent and multi-agent systems.
Regulators have moved too. Singapore’s IMDA Model AI Governance Framework for Agentic AI, published in January 2026, requires each agent to carry a verifiable digital identity and an audit trail recording which agent acted under whose authorization. EU AI Act provisions on human oversight and robustness apply to autonomous agents in high-risk domains.
Practical questions that follow:
- Can the platform issue and govern distinct identities per agent, rather than treating agents as shared service accounts?
- Can it record delegation, meaning which human or system authorized a given agent action?
- Does it support short-lived, scoped credentials rather than long-lived keys?
- Can it detect an agent acquiring permissions at runtime or spawning sub-agents?
- How does its roadmap track the NCCoE and COSAiS work?
A vendor with no answer to the delegation question is selling a platform built for the previous identity population.
Will It Integrate and Scale?
Document your environment before evaluating anything.
List the identity providers, directories, SaaS applications, cloud environments, databases, development platforms, and security tools that matter. Then assess integration for depth as well as breadth. A connector importing basic account information may not support certification, entitlement management, policy enforcement, or remediation.
Scale deserves scrutiny in two dimensions now. Entitlement volume across thousands of applications, and identity count growing on a curve steeper than headcount. Palo Alto Networks projected machine identities rising 77% against human identities at 56%, which means a platform sized for today’s population may not hold.
Ask how quickly new applications and identity sources get incorporated. Governance cannot depend on long manual implementation cycles in an environment that changes weekly.
Is It Auditable and Operationally Controllable?
Decisions need to be explainable after the fact.
Teams should be able to establish who had access, what they received, why it was granted, whether anyone reviewed it, and what happened when policy changed. Reporting should cover certifications, policy decisions, remediation actions, exceptions, and access changes, and should surface patterns rather than dumping entitlement data.
Administrators need centralized control over approval processes, policy definition, outstanding action monitoring, and investigation of unusual access relationships.
Test all of it with realistic scenarios rather than reviewing features individually. Build sample cases covering an employee transfer, a contractor termination, privileged access provisioning, application onboarding, an over-permissioned account, and an AI agent requesting elevated access. Assess each platform on how it handles the case from discovery through remediation.
That last scenario separates vendors more sharply than anything else on the list.
How Should the Final Comparison Work?
Against your environment, not against the category.
A multinational with several directories and thousands of applications needs something different from a cloud-first company with a narrower footprint. Connect technical capability to governance outcomes: how much manual work disappears, how clearly access relationships render, how effectively least privilege holds, and how the platform fits existing processes.
Cost extends past licensing. Implementation, integration, administration, training, and ongoing maintenance all shape what running the program actually requires.
One organizational note worth anticipating. Executives frequently believe least privilege is enforced because they are looking at human access, while practitioners working with machine and automated systems report otherwise. A platform that reports on both populations separately settles that disagreement with evidence.
FAQs
Q. What is the most overlooked requirement in IGA evaluation?
Non-human identity coverage. Service accounts, tokens, and agents outnumber employees substantially and receive the least review.
Q. Who certifies an AI agent’s access?
An open question most programs have not answered. Look for platforms that attribute ownership, record delegation, and support usage-based evidence rather than manager attestation.
Q. Are there standards to evaluate against?
Emerging ones. NIST’s AI Agent Standards Initiative and the NCCoE concept paper on agent identity and authorization both launched in early 2026, with SP 800-53 control overlays in development.
Q. Does just-in-time access replace certification?
No. It reduces standing privilege, which shrinks what certification has to cover. Both still matter.
Q. How do you test a platform properly?
Run scenarios rather than feature reviews. Employee transfer, contractor termination, privileged provisioning, application onboarding, excessive permissions, and an agent requesting elevation.
Q. What matters more, integration breadth or depth?
Depth, usually. A shallow connector to fifty systems governs less than a full integration with the fifteen that hold your sensitive data.
The Bottom Line
Choosing an identity governance platform involves more than confirming it supports access reviews and role management. Visibility, access modeling, contextual certification, least-privilege enforcement, automation, integration, scalability, and auditability belong in one evaluation rather than as separate checkboxes.
What changed is who the governance applies to. A discipline built around employees now oversees a population mostly composed of accounts that never change jobs, never leave, never fail an MFA prompt, and increasingly decide at runtime what they need access to.
The right platform helps teams understand who and what has access, judge whether that access remains appropriate, and act when it does not. A capability-based checklist keeps that comparison honest, provided the checklist reflects the environment as it is now rather than the one governance tools were originally designed for.
Related: AI Is Breaking the Old Rules of Data Center Resilience
