Since November 4, any employee can create an agent identity blueprint in Microsoft Entra. This is the default setting, according to the Microsoft Entra Agent ID documentation. Nothing to enable, nothing to request. An administrator can, however, turn this right off in the Entra user settings.
The question for IT has changed. It is no longer about blocking, but about deciding who may create agents, under which rules and with what oversight.
What changed on November 4: creation moves to the business
An agent identity is one more actor in your directory. It has access, it acts, it leaves traces. When any employee can create an agent identity blueprint, the agent estate grows without anyone having decided it.
Until now, governance meant approving what came in. Now you have to organize what already exists and what will appear tomorrow. For IT, the role shifts: from gatekeeper to architect of the framework.
This is a change of posture. Agent governance is no longer only a security question. It is an adoption decision: you choose how AI usage will develop in your company.
Why blocking everything does not hold
The reflex is understandable, especially on the security side. But in many cases, a total block produces the opposite of the intended effect.
- Teams with a real need work around it: personal tools, external accounts, automations out of your sight
- You lose visibility, and with it the ability to measure what is used and what it delivers
- The business concludes that AI is a risk, not an opportunity, and adoption stops or goes elsewhere
Blocking gives an impression of control. Usage, meanwhile, carries on without you.
The opposite is no better. If you let things run, orphan agents will appear: created for a test or a one-off project, then forgotten. Their creator changes role or leaves the company. The agent stays, with its access, and nobody knows who owns it or whether it is still useful.
The right goal is not zero agents. It is zero invisible agents, and as many useful, tracked agents as possible.
The three decisions to make now
Between blocking and letting go, a simple framework exists: who creates, who approves, who deletes.
Who creates
Not every employee needs this right on day one. Decide by population: trained champions in each business line, identified project teams, IT. Start narrow, then widen as usage takes shape.
A useful test: whoever creates must be able to explain what the agent is for, which data it touches and who answers for it. If they cannot, the right is premature.
Restricting only makes sense if a clear path exists to obtain an agent. Without it, you fall back into workarounds.
Who approves
Not all agents need the same level of control. An agent that summarizes its user's documents does not have the same profile as one that reaches sensitive data or acts on behalf of a department.
Define two or three risk levels. For each, name who says yes: the manager, IT, security, legal or compliance. The aim: a fast answer for simple cases, a serious review for sensitive ones.
Document every decision. In six months, you will need to know why a given agent was accepted.
Who deletes
This is the most forgotten decision. Every agent needs a named owner, a person and not a team. The owner answers for the agent: its usefulness, its access, its end date.
The lifecycle comes down to three rules:
- An owner assigned at creation
- A regular review: is the agent still used, is its access still justified?
- A deletion or reassignment decided as soon as the owner changes role or leaves
Where to start this week
No need to wait for a six-month program. Five actions are enough to take back control.
- Check the current setting in the user settings of your Entra environment. Knowing whether the default applies to you is the prerequisite.
- Take stock of existing agents: who created them, with what access, for what use.
- Bring IT, security, the business lines and legal together for one hour. Decide who creates, who approves, who deletes for a first version.
- Name an owner for each existing agent. Those without one are your first priority.
- Announce the rule to teams, with a simple way to request an agent.
An imperfect first version beats the default setting. You adjust it with what usage teaches you.
To measure the effect, track two simple things: the number of agents with an owner and the number of agents actually in use. They show whether governance serves adoption.
Decide before the default does
The default setting has already decided for you: everyone can create. You can keep it, restrict it or frame it, but you have to decide. Setting the framework early keeps both adoption and control.
Sources: Microsoft Entra Agent ID documentation and Microsoft message MC1506733, "Review who can create agents in Microsoft Entra".
Do you have an AI program moving slower than planned? Let's talk to identify what is blocking it, realign stakeholders and get adoption moving again.