Persistence Analytics Group LLC

Neil P. Osnato

Authority Infrastructure Shaper

Enterprise AI is moving quickly from generating information to taking action.

The first wave of generative AI helped employees draft documents, summarize information, analyze data and answer questions. The next wave is different. Agentic systems are increasingly being designed to initiate transactions, interact with enterprise systems, coordinate workflows, make recommendations that trigger action and operate with decreasing levels of direct human intervention.

That transition creates an infrastructure problem that deserves more attention.

Organizations have spent decades building systems for identity, authentication and access control. Those systems answer important questions: Who is the user? Has that identity been authenticated? What systems or resources may that identity access?

Agentic AI introduces another question:

What authority has actually been delegated to the agent, under what conditions, for what purpose, and how can that authority be proven when the agent acts?

That is not quite the same problem.

Identity Is Not Authority

An enterprise can know exactly which AI agent executed an action and still lack a complete answer as to whether the agent was authorized to take that specific action.

Identity establishes who or what acted.

Authentication establishes that the identity is genuine.

Permissions establish what a system technically allows the identity to access or execute. Authority is different.

Authority establishes whether the action was legitimately delegated, whether the delegation remained valid when the action occurred, whether applicable boundaries were respected and whether the organization can produce evidence supporting that conclusion afterward.

This distinction becomes increasingly important as AI systems move across applications, data environments, financial systems, operational platforms and third-party services.

A technically permitted action is not automatically an authorized business action.

Capability is not authority. Permission is not authority.

The Delegation Chain Matters

Human organizations operate through chains of delegated authority every day.

A board delegates authority to executives. Executives delegate authority to employees. Policies, contracts, approval limits and operating procedures define what those people can do.

Agentic systems extend that chain.

An employee may instruct an AI agent. That agent may call another service. A secondary agent may execute part of the workflow. An external platform may complete the transaction.

The resulting chain can become difficult to reconstruct:

Who originally authorized the action?

What exactly was delegated?

Could the agent delegate further?

What monetary, operational or informational boundaries applied?

Were those boundaries still valid when the action occurred?

What happens when the agent's role, model, environment or instructions change?

Those questions cannot always be answered by examining the final transaction alone. Execution proves that something happened.

It does not necessarily prove that the action was authorized.

Agentic AI Needs an Authority-Evidence Layer

As enterprise autonomy increases, organizations may need an additional infrastructure layer between identity and execution.

That layer should establish a durable evidentiary relationship among:

Identity → Authentication → Authority → Action → Reliance

The purpose is not to slow AI down. It is to make autonomous action institutionally usable. A mature authority framework should allow an organization to determine, at minimum:

• what authority was delegated;

• who or what granted it;

• the scope and purpose of the delegation;

• applicable limits or conditions;

• the systems or actions covered;

• whether further delegation was permitted;

• when the authority began and ended;

• what material changes require revalidation; and

• what evidence proves that an executed action fell within the authorized boundary.

The objective is not another static permission table. The objective is a defensible chain of evidence.

Authority Must Be Revalidated

Delegated authority also cannot be treated as permanent simply because it was valid once. The environment surrounding an AI agent can change.

The underlying model may change. The workflow may expand. A new tool may be added. The agent may receive access to additional data. Transaction limits may increase. A vendor relationship may change. The business purpose supporting the original delegation may disappear.

In conventional governance, these events might trigger a new approval, policy review or control assessment.

Agentic systems need an equivalent concept. A useful rule is:

A previously authorized agent should not remain automatically authorized when a material fact supporting the original delegation changes.

That creates the need for explicit revalidation triggers.

The question is no longer simply, "Was this agent approved?" It becomes:

Does the evidence supporting that approval still hold?

Governance Will Move Closer to Infrastructure

Much of today's AI-governance discussion centers appropriately on model risk, cybersecurity, privacy, safety and human oversight.

Those issues remain essential.

But as AI systems become operational actors, governance will increasingly have to address the infrastructure of delegated authority itself.

CIOs, CISOs, CTOs, risk leaders and business owners will need to know not only whether an agent is secure and technically capable, but whether the organization can prove that its actions were legitimate within a defined authority chain.

That matters for financial transactions, procurement, customer interactions, data access, operational technology, software changes and virtually any other environment in which AI moves from advising humans to acting on their behalf.

The institutions that solve this problem will not necessarily be the ones that restrict agents the most.

They will be the ones that can safely delegate more because they can verify exactly what was delegated and prove whether execution remained within that boundary.

The Next Question for Enterprise AI

The first enterprise AI question was:

Can the system do the task?

The next was:

Can we trust the system to do the task?

As agentic AI matures, another question is becoming unavoidable:

Did the system actually have the authority to do it?

That question sits between governance and infrastructure.

And as autonomous systems begin making consequential decisions and taking consequential actions, that missing layer may become one of the most important foundations of enterprise AI.