Can you name every AI agent running in your production environment right now? When 82% of enterprises discover previously unknown AI agents in their environments, and 41% say it happened more than once, the odds are you have agents you cannot name and permissions you cannot trace.
This article is a diagnostic anchored to NIST, CSA, and Gartner, and one part of the broader governance and readiness picture. Any readiness assessment must start from why traditional IAM cannot simply be extended to agents. Microsoft Agent 365 reached general availability this year, embedding autonomous agents into the productivity suite. When it hits organisations without a baseline, governance debt compounds daily.
Is your IAM ready for AI agents? The assessment dimensions that matter
Traditional IAM was designed for humans. Agent identity spans provisioning through decommissioning, a different cadence because agents proliferate faster and often outlive the humans who created them. Only 18% of security leaders express high confidence their current IAM can handle agent identities.
Four dimensions anchor the evaluation: visibility, lifecycle, access model, and attestation — identity architecture fundamentals every readiness assessment must verify. An Agent Registry is the first gate. NIST’s AI RMF GOVERN function maps directly to this. Without an inventory, nothing else matters.
Visibility comes first. 78% of organisations have no documented policy for creating or removing an AI identity, making lifecycle the largest operational gap. None of these dimensions matter if the inventory is wrong. That inventory question is where maturity evaluation begins.
Measuring NHI governance maturity: from vaulting to continuous attestation
NHI governance maturity sits on a five-level spectrum. Level 0: no inventory. Level 1: credential vaulting. Level 2: automated credential rotation, the minimum viable marker. Level 3: behavioural baselining and continuous attestation. Level 4: full lifecycle governance with policy-as-code.
Most organisations self-assess one tier too high. “We use a vault” is Level 1, not Level 2. Only 23% have a formal strategy; another 37% make it up as they go.
Identity Security Posture Management enables Level 1 to 2, discovering shadow agents and mapping credential relationships. One Identity is among the vendors organisations encounter in this landscape.
The credential vaulting gap and what it costs you
Credential vaulting stores secrets. It does not answer: is this credential still needed? Has its permission scope drifted? Who owns it? When was it last reviewed?
Automated credential rotation is the Level 2 floor. Continuous attestation is the governance heartbeat vaulting lacks, too slow for the 80-to-1 NHI-to-human ratio when left to annual reviews.
GigaOm research, working with One Identity, has outlined this maturity framework. Answering “we use a vault” satisfies only the question of where secrets live. Whether the credential should still exist, whether its scope has drifted, who owns it: these remain unaddressed. And that gap is where the red flags show up.
Signals your organisation is not ready for production AI agents
The clearest signal is organisational: if no single person can name every agent in production and what each can access, you are not ready, regardless of tooling.
Red flags: static API keys in repositories, agents owned by a single Global Administrator, agents dormant for months with unrotated secrets, no process for removing an agent when a project ends. In one assessment, the longest-dormant agents held the highest-privilege scopes.
74% plan to run agentic AI within two years but only 21% have mature governance controls. 83% of security leaders at RSA 2026 agreed that business units are deploying agents faster than security teams can assess them. The practical readiness gate: before any production deployment, produce a complete Agent Registry, automate credential rotation for every entry, and name a human owner for each agent. Most cannot.
Build, buy, or bridge: a decision framework for agent identity governance
The assessment produces a diagnosis. What you do with it is a three-way decision: build on existing IAM, buy dedicated NHI governance, or bridge with an identity orchestration layer.
Build (extending Okta or Entra) suits teams with identity engineering capability. Buy (purpose-built NHI governance) suits a widening readiness gap. 57% now favour a blended approach, up from 51%.
Four factors shape the decision: time-to-readiness (build takes months, buy takes weeks), team capability (do you have identity engineers?), integration surface (multi-IdP, multi-cloud favours buy or bridge), and growth trajectory. Gartner projects that 40% of agentic AI projects will be cancelled by 2027. Each traces back to governance failures.
When the bridge option makes sense
Identity orchestration sits between traditional IAM and agent-native governance, abstracting policies across IdPs and providing a migration path rather than a rip-and-replace.
Bridge is strongest when you operate multiple IdPs, your IAM investment is too large to abandon but too rigid to extend, or you want to preserve flexibility while the NHI market matures. Bridge is the wrong answer with a single IdP, an urgent readiness gap, or no orchestration expertise.
Agents run across public clouds, on-premises systems, and private clouds, with 38% operating in hybrid configurations. The bridge option buys time while the vendor landscape consolidates.
The assessment is only as honest as you are
Most organisations fail at the first dimension: they cannot name every agent in their environment. Only 21% maintain a real-time inventory. Readiness is a diagnostic. Most organisations are one tier lower than they think.
Microsoft Agent 365 makes this urgent. The practical readiness gate is three questions: can you produce a complete Agent Registry? Can you demonstrate automated credential rotation for every entry? Can you name a human owner for each agent? If any answer is no, the assessment has done its job — and the next step is moving from assessment to action.
Frequently Asked Questions
What happens if we deploy AI agents into production without governance?
You deploy into a governance vacuum where nobody can name every agent, audit their permissions, or revoke access when a project ends. The immediate consequences are credential sprawl across repositories, ownership concentration on a single admin account creating a catastrophic blast radius, and dormant agents with unrotated secrets that become undetectable compromise vectors. Governance debt compounds daily.
How long does a readiness assessment typically take?
A candid initial assessment can be completed in two to four weeks. The visibility dimension, querying your identity graph for every non-human identity, often takes days once you have the right query points. The longer variable is the lifecycle and attestation dimensions, which require interviewing teams about provisioning processes they may never have documented. The bottleneck is rarely technical.
Is AI agent identity governance different from managing regular service accounts?
Yes, fundamentally. Service accounts are typically static, human-provisioned, and scoped to a known system boundary. AI agents operate autonomously, may chain permissions across multiple services, persist for years beyond the project that created them, and proliferate at machine speed. Traditional service account management was not designed for identities that create other identities or act without a human in the loop.
What should we do first if we discover we are at Level 0?
Start with inventory. Before buying any tooling or designing any policy, query Microsoft Graph API, AWS IAM, and every IdP in your environment and produce a single list of every non-human identity. Expect to find three to ten times more agents than you believed existed. That list, however incomplete, is your Agent Registry. Without it, nothing else matters.
How do I get board-level support for investing in NHI governance?
Frame it as a regulatory and operational risk question the board already understands. The EU AI Act Articles 9 and 10 demand documented access reviews for high-risk AI systems, and the Microsoft Agent 365 release means agents will be inside the productivity suite with or without the board’s approval. The choice is not whether to govern, it is whether to govern proactively or reactively under audit pressure.
Can existing SIEM or security observability tools fill the visibility gap?
No, not on their own. SIEM tools detect anomalous behaviour once an agent is already known and monitored. They cannot discover agents you have not inventoried, map credential ownership relationships, or identify which service principals have not been accessed in ninety days. Identity Security Posture Management tools are purpose-built for that discovery layer. SIEM is the detection layer that sits above it.
Is this only a concern for large enterprises?
No. The pattern is the same regardless of organisation size: a developer creates an API key for a prototype, the prototype ships, the key is committed to a repository, and nobody remembers it exists. Small organisations have fewer agents but the same governance gaps, and a single compromised credential can be existential when there is no security operations team to catch it. The ratio of agents to governance controls is what matters, not headcount.
What compliance frameworks require NHI governance?
Several frameworks now explicitly or implicitly demand it. The EU AI Act Articles 9 and 10 require risk management and data governance for AI systems, including access controls. ISO 27001:2022 Annex A controls for identity and access management apply to non-human identities. SOC 2 requires access review evidence for all identities, human and non-human. NIST AI RMF’s GOVERN function maps directly to knowing what agents exist and what they can access.
What is the Microsoft Agent 365 catalyst and how urgent is it really?
Microsoft is embedding autonomous AI agents directly into the Microsoft 365 productivity suite, meaning agents will operate inside your tenant with access to email, documents, calendars, and Teams conversations. For organisations that cannot produce a basic Agent Registry today, this means unknown agents with unknown permissions operating inside the collaboration environment within the current product cycle. The readiness clock runs out at general availability.
How often should we repeat the readiness assessment?
Quarterly at minimum, and after any significant agent deployment. Agent proliferation is not linear. A single automation project can spawn dozens of new non-human identities in a week, and the inventory that was accurate last quarter may be materially incomplete today. Continuous discovery tooling, once deployed, makes this cadence practical by replacing manual inventory efforts with automated posture scans.
