Governing Agentic AI: What Every Institution Needs to Know

A plain-language guide for leaders in healthcare, banking, and other sensitive industries

By Dr. Augusta Uwah, MD, MPH · October 27, 2026

Why are we talking about this now?

AI is moving beyond answering questions. AI agents can open records, connect to other applications, and take actions on someone’s behalf. That creates new opportunities—and new questions about access, permission, safety, and responsibility.

Banking: In October 2026, reports described major U.S. banks restricting outside AI agents that accessed customer accounts using shared login credentials. The concern is not simply whether a customer asked an agent for help, but whether a bank can verify what that agent is allowed to do—and prevent an unintended transfer or other action.

Healthcare: On October 21, 2026, Epic announced Connectors that will allow its AI assistants and locally built agents to exchange information and perform approved tasks in other apps. Epic says health systems will review what each Connector can read and change before enabling it. This is an example of expanding capabilities that require clear institutional oversight—not evidence of a breach.

The central finding

Approving an AI company is not the same as governing every agent it operates. Institutions need to know which agents are active, what they can access, what they actually do, and whether their results are correct.

What leaders should understand

An agent needs a named human owner—even if a vendor built it.
Permission to read information is different from permission to change, send, or delete it.
An action log is not enough: important results must be checked in the receiving system.
Outside agents and subcontracted agents can cross organizational boundaries.
Disabling an agent must also disable its credentials, connected tools, and delegated access.

Seven practical steps

Register each agent, including vendor and external agents.
Assign a responsible owner and document its approved purpose.
Set boundaries: what it may read, write, send, or execute.
Compare approved permissions with the access it actually has.
Record important actions, approvals, and independently verified outcomes.
Review changes, errors, and unusual behavior; escalate when necessary.
Remove access and verify shutdown when an agent is retired.

Bottom line

The goal is not to stop AI innovation. It is to make AI useful without losing institutional control: know the agent, limit its authority, verify its actions, and keep a human accountable.