Editor's note — September 12, 2026: Clarified service identities and removed the prediction that model quality would not contribute to failures.
The Next AI Risk Isn't the Model. It's the Agent.
THE TAKE
For the last two years, organizations focused heavily on model governance.
That made sense.
The biggest fear was AI generating incorrect information, exposing sensitive data, or creating legal liability.
But I believe many organizations are now focused on the wrong problem.
The real governance challenge is shifting from AI content generation to AI action execution.
The moment an AI agent can:
- Access ERP systems
- Approve transactions
- Modify customer records
- Trigger workflow automation
- Interact with third-party applications
You have effectively created a new digital workforce.
And like any workforce, it requires identity management, segregation of duties, monitoring, logging, and accountability.
Model weaknesses and missing operating controls can both cause failures.
My recommendation is to evaluate them together before increasing an agent's authority.
Boards should ask about the model:
"Which model are we using?"
And also ask:
"What authority have we given our agents?"
THREE THINGS BOARDS SHOULD KNOW
AI agents create a new identity problem — Organizations now need to define what systems agents can access, what actions they can perform, how permissions are managed, and how activity is monitored.
Governance is moving into architecture — AI governance can no longer live only in policies, committees, and documentation; it must be enforced through operational controls.
Auditability will become a competitive advantage — Organizations that can reconstruct every AI decision, action, approval, and system interaction will be far better positioned during incidents, audits, and regulatory reviews.
THE QUESTION I'M HEARING
"Can we treat AI agents like service accounts?"
An agent may authenticate through a service account or another workload identity.
The identity mechanism does not determine the scope of permitted behavior.
An agent that interprets instructions and invokes tools still needs bounded authority, monitoring and a named owner.
The governance model needs to evolve beyond standard IAM. Every AI agent should have a defined owner, approved purpose, limited access, monitored activity, and clear escalation paths when it reaches the edge of its authority.
Boards should ask management for an AI agent inventory before approving broad deployment. What agents exist? What business process do they support? What data can they access? What actions can they take? Who is accountable when something goes wrong?
THE BOTTOM LINE
Model evaluation remains important.
Agent behavior adds another layer of operating risk.
If AI is going to act on behalf of your organization, your board needs visibility into who authorized it, what it can do, how it is monitored, and how its actions are audited.
Source and further reading: Primary reference. Framework guidance is not a statement that a particular law applies to every organization.