AI Governance May Be Having Its SOC 2 Moment

On September 9, 2026, California enacted two laws aimed at strengthening independent oversight of artificial intelligence. Senate Bill 813 establishes a framework for independent verification organizations that can assess AI systems and models for compliance with state law. Assembly Bill 1405 creates a state registry for AI auditors and establishes standards for independence, transparency and integrity.

On September 18, Governor Gavin Newsom issued an executive order accelerating implementation of those oversight measures and directing recommendations on potential additional independent verification of frontier-model safety frameworks, transparency reports and risk assessments, as well as a potential emergency shutoff mechanism.

These developments do not create a general requirement for every organization to undergo an independent AI audit. They do, however, make the separation between readiness and independent evaluation more visible.

The most interesting part, from an assurance perspective, is not simply that auditors are being registered. It is that independence is becoming part of the architecture of AI assurance.

That matters.

An organization can design controls. A consultant can help it implement those controls. An internal team can collect evidence. But independent assurance only has value if the party providing that assurance has sufficient distance from the design and operation of what it is assessing.

That principle will be familiar to anyone who has worked in financial statement audit, SOC reporting or other mature assurance disciplines.

I spent years on the audit and assurance side of SOC engagements at Deloitte.

A large part of that work came down to a deceptively simple question: does the evidence actually demonstrate that the control operated as intended?

A policy was not enough.

A process description was not enough.

Even a technically sound control was not enough if the organization could not demonstrate how it operated over the relevant period.

That is one reason the distinction between readiness and assurance matters so much.

Those activities are complementary. They should not be collapsed into one.

Many organizations already have AI policies. They may define:

• Acceptable AI use
• Human oversight
• Prohibited use cases
• Privacy requirements
• Risk ownership
• Model-review processes
• Vendor requirements

That is useful.

But as AI moves from generating content to taking actions, policy alone becomes increasingly insufficient.

Consider an AI agent capable of:

• Calling APIs
• Modifying records
• Sending external communications
• Deploying code
• Changing permissions
• Initiating downstream workflows
• Accessing sensitive data

A policy may say that material actions require human approval. The assurance question is:

• Where is that approval actually enforced?
• Can the agent proceed if approval is absent?
• Is approval tied to the exact action being performed?
• Can it expire?
• Can the agent bypass the gate?
• Who approved the action?
• Can the organization reconstruct what happened afterward from evidence the agent itself cannot alter?

That is the difference between documented governance and operating governance.

The pattern increasingly looks like this:

This is not unique to AI. It is how assurance disciplines mature.

ISO/IEC 42001:2023 specifies requirements for an Artificial Intelligence Management System. It gives organizations a management-system framework for governing AI across areas such as leadership, planning, operation, performance evaluation and continual improvement.

ISO/IEC 42001 is about the organization's management system. It is not, by itself, proof that every individual AI system or control is effective, nor is it a general legal mandate.

ISO/IEC 42006:2025, published in 2025, goes one layer further. It defines additional requirements for bodies that audit and certify AI management systems under ISO/IEC 42001, including expectations around competence and credible certification.

That distinction matters. ISO/IEC 42006 is about the organizations that independently audit and certify those systems.

That is another signal that the AI-governance market is maturing beyond “do you have a policy?” toward: “Can this governance system be independently assessed?”

The practical lesson is not that every company needs to rush out and obtain a new AI certification tomorrow.

The lesson is that organizations should start building AI governance as though they may eventually need to prove it.

That means asking questions such as:

• Do our AI systems and agents have identifiable owners?
• Are non-human identities governed?
• Are permissions technically constrained?
• Are high-impact actions subject to meaningful approval?
• Can those approvals actually stop execution?
• Are AI actions monitored after deployment?
• Can we reconstruct significant decisions and actions?
• Is the evidence independent of the system being evaluated?
• Could an external assessor understand and test how these controls operate?

Organizations that wait until an auditor, regulator or enterprise customer asks these questions will likely find themselves reconstructing governance after the fact.

That is usually harder than designing for evidence from the beginning.

There is also an important line for consulting and assurance firms.

Helping an organization design its AI governance program is valuable.

Helping it identify control gaps is valuable.

Helping it become ready for independent assessment is valuable.

At EK Treasure, AI governance readiness work focuses on helping organizations establish controls and evidence that can withstand scrutiny; it does not substitute for an independent assurance conclusion.

But readiness work and independent assurance should not become indistinguishable.

The independence requirements appearing in California's emerging framework are a useful reminder of why.

“Independence is not an inconvenience in assurance. It is part of what makes assurance credible.”

For firms working in AI governance, this creates two distinct roles:

• Helping organizations build controls and evidence that can withstand scrutiny; and
• Independently evaluating those controls when the appropriate independence requirements are satisfied.

Those roles can exist in the same ecosystem. They should not be confused.

AI governance is not simply cybersecurity with a new label.

The risk surface is different.

• Models can behave probabilistically.
• Agents can chain tools and actions.
• Training and inference data introduce provenance concerns.
• Human oversight can become ceremonial.
• Multi-agent systems introduce delegated authority and identity questions that traditional software architectures did not always need to address.

So it would be a mistake to assume that SOC 2 provides the complete model for AI assurance.

But the historical progression is instructive.

AI governance appears to be moving in the same direction.

California's recent actions do not complete that transition. Neither does ISO/IEC 42001. But the direction is becoming easier to see.

For enterprise AI providers, the important question may soon stop being:

“Do you have an AI policy?”

and become:

“Can you demonstrate that your AI governance actually operates?”

That means organizations should think beyond policy creation and start designing for:

Identity. Authorization. Human approval. Monitoring. Evidence.

Because once independent AI assurance becomes common, the hardest part will not be finding an auditor. It will be having controls and evidence worth auditing.

“The hardest part will not be finding an auditor. It will be having controls and evidence worth auditing.”

AI governance may indeed be having its SOC 2 moment.