That creates an important identity question.
Who gets to authorise an autonomous agent?
It sounds like a straightforward IAM problem.
It is not.
Traditional identity and access management was designed around people, applications and predefined processes. An employee requests access. A manager approves it. A policy determines what the employee can do.
Agentic AI changes that model.
An AI agent may be acting on behalf of an employee, a business process or another system. It may make decisions dynamically and initiate actions that were not explicitly defined when access was originally granted.
The question therefore becomes much bigger than authentication.
It becomes a question of authority, delegation, accountability and control.
1. Authentication Does Not Equal Authority
The first mistake organisations can make is assuming that once an AI agent has been authenticated, it can be trusted to act.
Authentication establishes identity.
It does not establish authority.
An agent may have a legitimate identity and still be making an unauthorised request.
Consider an AI agent supporting procurement.
It may be authorised to search suppliers, compare prices and prepare recommendations.
That does not mean it should be able to approve a purchase order or transfer funds.
The distinction is critical.
Knowing who the agent is does not answer what the agent should be allowed to do.
IAM therefore needs to move beyond the question of whether an agent can access a system.
It needs to determine whether the agent should be permitted to perform a specific action under a specific set of circumstances.
2. Who Is Actually Giving the Agent Authority?
This is where agentic AI introduces a new governance challenge.
An autonomous agent rarely operates completely independently.
It usually operates within a chain of authority.
A person may instruct an agent.
The agent may call another service.
That service may access an application.
The application may retrieve sensitive data.
So who authorised the final action?
Was it the employee?
The application owner?
The security policy?
The agent itself?
Or some combination of these?
Without a clear authority model, responsibility becomes difficult to establish.
Enterprises need to know not only which identity performed an action, but where the authority to perform that action originated.
This concept becomes increasingly important as AI agents move from assistants to autonomous operators.
3. Delegation Is the Missing Link
Most organisations already understand delegation.
Managers delegate responsibilities to employees.
Employees delegate certain tasks to applications and automated workflows.
Agentic AI takes delegation to another level.
An employee may ask an AI agent to manage a specific business process.
The agent then needs enough authority to complete the task.
But that authority should not automatically equal the employee's entire access profile.
If an employee can access ten systems, an agent working on one specific task may only need access to two.
This is where delegated authority becomes central to agentic IAM.
The organisation should be able to define:
- Who delegated the authority
- What authority was delegated
- Why it was delegated
- What actions are permitted
- Which resources can be accessed
- How long the authority remains valid
- What conditions apply
- When the authority should be revoked
The goal is simple.
An agent should receive the authority required to complete its task, not the authority available to the person behind it.
4. Should an Agent Ever Authorise Itself?
This question becomes increasingly relevant as agents become more autonomous.
Imagine an AI agent that identifies a problem and decides that it needs access to another system to resolve it.
What happens next?
Can it request additional permissions?
Can another automated system approve those permissions?
Can an agent delegate authority to another agent?
If the answer is yes, organisations need clear boundaries around machine initiated authorisation.
Otherwise, an agent could potentially become part of an automated privilege escalation chain.
The principle should be clear:
Autonomy should not mean unrestricted authority.
An agent can be autonomous in how it performs an approved task without being autonomous in determining the boundaries of its own authority.
That distinction will become increasingly important as agentic systems mature.
5. Least Privilege Needs to Follow the Agent
Least privilege has been a core IAM principle for years.
Agentic AI makes it more important.
An employee may have broad permissions because their role requires them to work across several business functions.
An AI agent usually has a much narrower purpose.
Its access should reflect that purpose.
For example, an agent responsible for customer service may need to:
- View customer records
- Update case information
- Generate responses
- Create support tickets
It may not need to:
- Export the customer database
- Modify billing information
- Delete accounts
- Change security policies
Giving the agent broader access simply because the underlying employee has it creates unnecessary risk.
The better approach is to establish task based authority.
The agent receives the permissions required for the task and nothing more.
6. Context Needs to Influence the Decision
Static permissions are unlikely to be sufficient for highly autonomous systems.
An agent may be allowed to perform an action in one situation but not another.
Consider an AI agent that can approve expense claims.
A normal expense below a defined threshold may be automatically approved.
A large transaction may require additional human approval.
A transaction involving a high risk supplier may require a completely different control.
The agent's identity has not changed.
The context has.
This means authorisation decisions need to consider more than identity and role.
They may also need to consider:
- The action being requested
- The sensitivity of the resource
- The value of the transaction
- The identity represented by the agent
- The business context
- Risk signals
- Time and location
- Previous activity
- Applicable policies
This is where dynamic authorisation becomes particularly valuable.
7. Human Oversight Should Be Based on Risk
There is a tendency to frame agentic AI governance as a choice between complete autonomy and human approval for everything.
Neither approach is practical.
If every agent action requires a human to approve it, the value of automation quickly disappears.
If no actions require oversight, organisations risk giving autonomous systems too much freedom.
The better approach is risk based human oversight.
Low risk actions can be automated.
Medium risk actions may require additional policy checks.
High risk actions may require explicit human approval.
For example, an agent could automatically retrieve an internal document but require human approval before sending confidential information externally.
The objective is not to keep humans in every loop.
It is to ensure that humans remain in control of decisions where the consequences justify intervention.
8. Every Authorisation Decision Needs to Be Traceable
When an AI agent performs an action, organisations need to be able to reconstruct the decision.
This becomes particularly important when something goes wrong.
A useful audit trail should answer:
Which agent acted?
Who or what did it represent?
What action did it request?
What policy evaluated the request?
What authority was delegated?
Why was the action permitted?
Which systems were involved?
What was the outcome?
Without this information, investigating an incident involving an autonomous agent can become extremely difficult.
The organisation may know what happened but not understand why it happened.
That is a governance problem.
Identity therefore becomes an important part of the organisation's AI accountability framework.
9. Agent to Agent Delegation Creates Another Layer of Risk
The first generation of enterprise AI agents may operate independently.
The next generation is likely to involve multiple agents working together.
One agent may specialise in customer data.
Another may handle financial analysis.
Another may interact with enterprise applications.
They may communicate and delegate tasks between themselves.
This creates a new identity chain.
Human → Agent A → Agent B → Application → Data
Every transition creates another opportunity for authority to be expanded, misinterpreted or lost.
Organisations therefore need to understand not just the identity of individual agents, but also the relationships between them.
An agent should not automatically be able to transfer its authority simply because it has access to another agent.
Delegation should be explicit, governed and traceable.
10. IAM Needs to Evolve With Agentic AI
Agentic AI does not make traditional IAM obsolete.
It makes IAM more important.
The fundamentals remain relevant:
- Authentication
- Authorisation
- Least privilege
- Privileged access
- Identity governance
- Lifecycle management
- Access reviews
- Auditability
What changes is the nature of the identities being governed.
An employee follows instructions.
An application follows predefined logic.
An AI agent can interpret a goal and determine how to achieve it.
That introduces a different level of autonomy.
IAM therefore needs to evolve from managing static access towards managing dynamic authority.
The Trevonix Perspective
At Trevonix, we believe the critical IAM question for agentic AI is not simply whether an AI agent can be authenticated.
It is whether its authority can be clearly established, appropriately limited and continuously verified.
For enterprises adopting agentic AI, we see several principles as essential.
Clear Agent Identity
Every autonomous agent should have a distinct and attributable identity.
Defined Ownership
Every agent should have an accountable owner responsible for its purpose, permissions and lifecycle.
Controlled Delegation
Authority delegated to an agent should be explicit, scoped and appropriate to the task.
Least Privilege
Agents should receive only the access and permissions necessary to perform their responsibilities.
Dynamic Authorisation
Access decisions should consider context, risk and the action being requested rather than relying entirely on static roles.
Risk Based Human Oversight
High impact decisions should have appropriate human controls without unnecessarily slowing low risk automation.
Continuous Auditability
Every significant agent action and authorisation decision should be traceable.
Together, these principles create a foundation for allowing AI agents to operate with greater autonomy without turning autonomy into unrestricted access.
The Bigger Question: Who Is Accountable?
The most important question around autonomous AI may not be technological.
It may be organisational.
If an AI agent makes a decision that causes financial loss, exposes sensitive information or disrupts a critical process, who is accountable?
The answer cannot simply be "the AI."
An AI agent does not own a business process.
An organisation does.
That means enterprises need a clear chain of accountability connecting the agent to its owner, its delegated authority and the policies governing its behaviour.
This is why identity and AI governance cannot remain separate conversations.
Identity establishes who or what is acting.
Governance establishes who is accountable.
Authorisation establishes what the identity is permitted to do.
Together, they create the foundation for responsible autonomy.
Final Thoughts
Agentic AI is changing the relationship between identity and access.
For years, IAM has focused on determining whether a person or application should be allowed to access a resource.
The next challenge is determining whether an autonomous agent should be allowed to take an action.
That requires more than authentication.
It requires clear identity, controlled delegation, least privilege, contextual authorisation, human oversight and continuous accountability.
The organisations that establish these principles early will be better positioned to scale agentic AI without allowing autonomy to outpace governance.
The question is no longer:
"Can we authenticate this AI agent?"
It is:
"Who gave this agent the authority to act, what exactly is it allowed to do, and can we prove that the decision was justified?"
That is where IAM meets agentic AI.
And that is where the next generation of identity governance begins.



