Someone Has to Administer the Agents. Right Now, Almost Nobody Does.
November 4, 2026
Jillian, CMO— AI-assisted and reviewed prior to publication.
Almost every enterprise is running AI agents now, and almost none of them can say exactly how many, who built them, or what they can access. That gap is creating a new administrative function inside IT, one that looks a lot like the identity and access work admins already do, applied to software that acts on its own.
Why Does Nobody Know How Many Agents They're Running?
Because agents get created the way shadow IT always got created: fast, by whoever had a reason to build one, with no central registry to log it. The OutSystems 2026 State of AI Development report found that 96% of enterprises now use AI agents, but governance has not caught up with adoption. The same research found that 94% of organizations report concern that AI sprawl is increasing complexity, technical debt, and security risk, while only a small fraction have a centralized way to govern it.
That pattern shows up across more than one survey this year. A review of enterprise agent governance research covering OutSystems, SAP, Gartner, and AWS found deployment has outrun inventory at a similar scale across all four, with SAP's LeanIX research showing that nearly all companies have deployed or plan to deploy AI agents while fewer than half have visibility into an actual inventory of them. When a majority of organizations cannot count their own agents, nobody is positioned to answer the more important questions: what data can this agent touch, who approved it, and what happens when it misbehaves.
What Does an Agent Administrator Actually Do?
An agent administrator manages the identity, permissions, and lifecycle of AI agents the same way a traditional admin manages user accounts, except the "user" never sleeps, never asks questions, and can act across multiple systems in seconds. The role sits at the intersection of Microsoft 365 administration, identity and access management, and compliance.
Microsoft built its response to this gap in two pieces. At Ignite 2025, the company introduced the Copilot Control System, which the official Microsoft 365 Copilot blog describes as covering security and governance, management controls, and measurement and reporting, anchored by a new control plane called Microsoft Agent 365 built specifically for agent registration, access control, visualization, interoperability, and security. The company has kept building on it since: an update to the platform described in Microsoft's Agent 365 blog explains that administrators can now take agent-level governance actions directly from a unified control plane, including lifecycle management actions like deleting synchronized agents across platforms.
In practical terms, the job covers work like this:
- Registering every agent in a central inventory instead of letting them spin up inside individual teams unnoticed
- Setting and auditing what each agent can read, write, or trigger, the same discipline applied to human account permissions
- Reviewing usage and cost data to catch an agent that is doing more, or less, than it was approved to do
- Decommissioning agents on a schedule instead of letting them persist indefinitely after the project that created them ends
That last point matters more than it sounds. An agent nobody remembers creating is functionally identical to an orphaned account with standing access, and those are exactly the kind of gaps that turn into incidents.
Why This Is Becoming a Named Role, Not a Side Task
Because the skills involved are specific enough that generalist IT knowledge does not automatically cover them, and employers are starting to test for them directly. Microsoft itself created a credential for this exact gap. The company's own announcement of the Microsoft 365 Certified: Copilot and Agent Administration Fundamentals certification states that it is intended to show employers you understand how to configure, secure, and govern Copilot and agents across Microsoft 365, and that you know the fundamentals of identity, security, and compliance well enough to support AI adoption rather than just use AI yourself.
The exam, AB-900, is structured around that distinction. A breakdown of the exam's skills from CloudThat's AB-900 certification guide shows roughly a third of the test covers core Microsoft 365 administration, including users, groups, roles, admin centers, licensing, and permission models, with the rest built around the AI-specific administration work of securing and governing Copilot and agents. That split is the clearest signal available right now of what employers mean when a job posting says "AI administration skills." It is not prompt writing. It is identity, access, and governance applied to a new kind of account.
The job title market is catching up to the same shift from a different direction. Coverage of the emerging "AI Agent Manager" role from The Interview Guys describes it as a role now appearing on job boards at companies that need someone accountable for agent performance, where domain expertise in operations and process matters as much as technical depth. Whether the title on your offer letter says administrator, manager, or something else, the underlying accountability is the same: somebody has to own what the agents are doing, and that somebody increasingly needs to prove it on paper.
What This Means for Your Next Certification
If you already work in Microsoft 365 administration, identity, or helpdesk operations, this is the most direct AI-adjacent credential available to you right now, because it builds on skills you likely already have rather than asking you to become a data scientist. Forge University's AB-900: Copilot and Agent Administration Fundamentals prep covers the identity, security, and governance groundwork the exam tests, alongside the agent-specific administration skills that are new to almost everyone right now, regardless of experience level.
It also pairs naturally with identity-focused credentials if you want to go deeper. Agent governance is fundamentally an access control problem, and professionals who already hold or are pursuing identity administration skills will find the concepts here familiar rather than foreign. If you are not sure which certification path fits your current role and where you want to go next, the certification resources and curriculum overviews on the blog are a reasonable place to compare options before committing time to one.
None of this requires betting your career on a brand-new title that might not stick. It requires being the person on your team who can speak fluently about agent identity, permissions, and lifecycle when the question comes up in a security review or an audit, because that question is coming regardless of what the job posting ends up calling the role. If you want a study plan built around exactly that gap, you can start training whenever you're ready.
The sprawl numbers are not going to shrink on their own. Every enterprise that deploys an agent without a registry, an owner, and an offboarding plan is building the same kind of exposure that ungoverned accounts created for a decade, just faster and with less visibility into what happened. The organizations catching up are not hiring agent administrators because it is trendy. They are hiring them because the alternative is explaining to an auditor or a board why nobody can list what their own AI is doing.