Enterprise governance was built for decisions made by people and for software that behaves consistently. That's not how AI agents work. They are non-deterministic, easy to build and able to act on their own.
That mismatch can cause challenges for governance. Organizations can't take a week to set a static policy, as that approach simply can't keep up with the volume of autonomous actions from agentic AI.
The governance gap is becoming a barrier to scaling AI. Organizations that can't govern agents at runtime risk falling behind peers that can.
TechTarget interviewed Blake Brannon, chief innovation officer at OneTrust, a privacy and data governance company, about this topic and how CIOs can effectively govern AI agents. Here's what he had to say.
Editor's note: The following interview was edited for length and clarity.
Is "bring your own AI" becoming the new "bring your own device," or are organizations steering people toward sanctioned AI tools rather than letting them choose their own tools and models?
Brannon: The premise back then, when you had a BlackBerry that you carried around for work, and you had a personal iPhone or an Android, and you were like, "Why am I carrying two devices around? Can't I just put my work data on my personal phone?"
A lot of the parallels that I see today. When I think about AI, I think of an explosion of new technology that puts organizations on their heels about how to address shadow AI. How do we secure the data that's going on those things, and approve it?
You have a lot of parallels, I think, to that IT’s answer can't be "no, let's block it all." It's got to be: "How do we safely move forward and adopt this new technology?"
Why is governance becoming a major barrier to enterprise AI success, particularly as organizations adopt AI agents?
Brannon: Fundamentally, governance was always something that had a lot of assumptions, and it was built and designed for governing things humans did, or software that humans built that was deterministic. And all those assumptions are now gone, right?
AI is non-deterministic. Everyone's a builder. Everything built is connected, and all these things act autonomously. What that means is that governing programs today take weeks of time to go through. But we've got to figure out how to go from weeks to seconds or milliseconds, right? Because an agent within a company will reason in real time. It will connect to things and act in real time.
So that governing approach, where you used to be able to govern something upfront that was deterministically decided, must happen at runtime. Every organization expects that within the next two years, there will be more agents than humans working inside their company. That means there are more autonomous actions that have to be analyzed and decisions made about how they'll be governed and whether they're permissible.
That means organizations must not just solve getting the AI to functionally work or be intelligent; they've got to make sure it doesn't introduce security problems in the organization.
If you can't solve governance at runtime, you won't be able to scale AI in your organization, and I think every organization sees that as a competitive disadvantage. If I can't scale effectively using AI autonomously in my organization, I'll be at a competitive disadvantage to my peers who do figure out how to do that.
What does a successful AI agent need to get right, and where does governance fit in?
Brannon: It's got to have intelligence, security and governance solved. The intelligence quotient is: did it do the thing it was supposed to do? It also can't leak data to the wrong person. It can't be jailbroken and create a backdoor into the organization. That's the security quotient. It also has to act permissibly, meaning it must adhere to our organization's governing policies.
The reason I call it a quotient is that it's math that is multiplied together. If you forget any of those, or they're zero, the value of that AI agent or the ROI of that AI use case is going to be zero, so you have to solve all three.
What is the intersection of compliance and governance?
Brannon: Most people start out with: what's the regulatory requirement for AI specifically? And that is almost the wrong question. The right question is: what existing regulations and compliance obligations do we have that are likely rooted in actions you take as an organization or in data use in general?
AI is just an engine, just like compute historically has always been an engine that is still doing that.
This means many existing regulations still govern your use of data in AI, with no new AI-specific requirements or regulations added. That means organizations still have to adhere to many requirements for their use of AI, regardless of whether it's new or different.
What we see happening, though, in practice, whether you call it compliance, regulation, a framework of best practices, or governance traditionally, these are kind of all on a bit of a collision course. A simple way to think of it is, if you have an AI agent that you're trying to give a set of governing instructions to, it's kind of like you can't have three different parents telling you different things. You need one boss who says: here are your things to follow; here are your guardrails and guidelines.
“If you have agents that are operating inside an organization and taking action, most of the time those agents are inheriting the person's identity to access the systems. But while they inherit their identity, they don't necessarily inherit their trust.”
Blake Brannon
Chief innovation officer at OneTrust
Where should human oversight remain essential, and how can organizations decide when an agent needs approval versus when it can act independently?
Brannon: I think every organization and every use case is going to have a bit of a uniqueness to the answer.
The framework to think about that is going to be: what is the risk of the decision, what is the reversibility of the decision, and what, if anything, are specific requirements that require human-specific intuition. Most organizations treat any use case in any scenario as kind of a snowflake today. It's almost like every new use case for something they look at as a unique beast. But to go from weeks of time to do that to seconds, one of the key principles you have to start to realize is that every decision can be broken into fundamental sub-decisions that are being made. Those sub-decisions probably have a lot of parallels to decisions you made yesterday, decisions you made about similar projects.
So, here's the decision that we've already made, 28 times in the past three weeks. Let's reuse that decision without a human having to make it again.
What's the intersection or requirement of identity when we're talking about agentic AI governance?
Brannon: Well, I think the simple way to think of this is, if you have agents that are operating inside an organization and taking action, most of the time those agents are inheriting the person's identity to access the systems. But while they inherit their identity, they don't necessarily inherit their trust.
A simple example is that most organizations would say, "Okay, I trust Blake the human to have access to the CRM to delete a record, an opportunity, a contact, but I don't quite trust Blake's agent operating on his behalf to have that authority yet." So, this means that when you're trying to govern and control the actions these agents are allowed to take, you have to understand the agent's identity. Is it inherited from a human, or is it something that is unique, like a non-human identity specific to that agent and their task? What permissions does that inherited or unique identity go with, and how does that tie into the governing of the agent to determine what is a permissive action or a non-permissive action?
These also carry what we call a notion of purpose intent. So, one of the core concepts you have to start solving for in agents, which are typically not there for humans, is the purpose of the action.
What practical guardrails should organizations establish before giving AI agents the authority to make decisions or take action?
Brannon: There are three core parts of the solution you've got to solve for here. One is you've got to be able to govern AI, you've got to have the ability to have visibility into all AI and then control it. How you do that is you've got to have four core steps to solve. One, you've got to have visibility into all the AI. The second one is you've got to have a policy system of truth, so a system of record for your policies, that effectively reflects the way you are making decisions as an organization. You've got to have a control mechanism to allow or deny any AI agent from doing anything, and then you've got to have a way to prove what's happened. You've got to have logs and evidence that can't be manipulated.
The way I always think about that is, no one should ever grade their own homework. An AI agent writing what it did is not evidence that something actually did or didn't happen because it can manipulate its own log, its own memory and its own thinking. So those are the four kinds of building blocks that you've got to have in place.
But conceptually, you've got to have a governing system that can determine how you decide as an organization. It can act and be where you can't be, which is every agent tool call at two a.m. on a Tuesday, and it's got to be able to prove what's happened.
Sean Michael Kerner is an IT consultant, technology enthusiast and tinkerer. He has pulled Token Ring, configured NetWare and been known to compile his own Linux kernel. He consults with industry and media organizations on technology issues.
