As enterprises move beyond AI prototypes and begin deploying AI agents into real business processes, governance is becoming a critical challenge.
In an interview at the Ai4 Conference in Las Vegas, OneTrust co-founder and Chief Innovation Officer Blake Brannon argues that organizations need a fundamentally new approach to AI governance—one built around independent oversight, runtime decision-making, deterministic controls, and policies that can evolve as AI systems change.
Rather than attempting to control every model or AI platform, Brannon says enterprises should focus on the moment agents interact with corporate data and systems, using “guardian agents” and an agent control plane to determine whether an action is permissible and, when necessary, stop it.
Core Takeaways
- Defining AI Governance: Effective AI governance starts with three questions: Do you know which AI agents and systems are operating inside the organization; do you know what they can access and do; and can you trust their actions?
- Guardian Agents and Independent Oversight: Brannon argues that enterprises need guardian agents that operate independently from business agents, evaluating their intent and actions rather than relying solely on the AI system to govern itself.
- Runtime Control Is Essential: Because AI agents are non-deterministic, traditional static permissions and guardrails are not enough. Enterprises need deterministic control points that can allow or block an agent’s actions in real time based on purpose, context, risk, cybersecurity requirements, and compliance obligations.
- AI Governance as Policy as Code: Organizations will need to translate cybersecurity, compliance, and business policies into machine-readable rules that power an agent control plane, while continually updating those policies as corporate risk tolerance, regulations, systems, and AI behavior evolve.
Key Quotes
AI Governance Comes Down to Three Questions
“If you are doing AI governance, you’ve got to be able to answer three really simple questions. You’ve got to be able to say, I know what AI agents and systems I have in my organization. I know what they’re connected to, what they have access to do, and I trust what they’re doing.”
“The first thing you’ve got to do, it’s really simple. It’s two things: Is the agent action or the thing that the agent’s doing, is it something that is permissive? Is it okay for it to happen? The second thing you’ve got to have is a system of control to actually enforce or block that thing. So you’ve got to decide and you’ve got to enforce.”
Purpose and Context Determine Whether an Action Is Safe
“This is where cybersecurity and compliance have always been able to rely on dynamic or fixed type conditions or rules to govern the systems and the actions. And that doesn’t work anymore when you have non-deterministic agents. You have to figure out in runtime when it’s actually figuring out what to do.”
“One of the core context or principles that you have to take into notion to decide is what I would call purpose. So it’s like: what is that creating of the PDF actually being used for? And in that cross section of purpose is actually the answer to: is this okay or not okay?”
Guardian Agents Can Govern Business Agents
“Guardian agents sit in an architecture where you have these nested layers of roles. You have a guardian agent that by design, you want separate than the business agent that’s doing something. And the business agent has this open charter of here’s your objective, go solve and do this.”
“The guardian agent’s job is not to have an open charter of figuring out ‘how do I make that action okay by finding a loophole in the compliance logic and things like that.’ It has a much more focused charter that’s about, given this action I’m seeing, given the intent and context of why it’s wanting to do that, is this adhering and a desired action from an organization’s policies and cybersecurity practices? And in run time, you can effectively make that decision and you can say this is allowed and this is denied.”
Guardrails Aren’t Enough—Enterprises Need Hard Controls
“Agent guardrails are more instructions than hard controls. You give an agent a prompt, you give it specific instructions. It doesn’t quite always follow it. And it’s kind of similar to the way a human doesn’t always follow instructions. It might hallucinate.”
“Can we put gates in place that allow us to say, no matter how it thinks, what it thinks, why it thought, what it concluded, when it goes to take that action, we have a control point that is deterministic? That is in our ability to control, to say, no matter what is happening here, I know I can block or allow this action from happening.”
