As AI agents proliferate across enterprise environments, the traditional governance playbook of written policies, acceptable-use guidelines and security audits is no longer enough. The tools are moving faster than the rules that govern them. That’s leading to risks way beyond shadow IT, according to Lynn Comp, vice president and global head of sales for the AI Center of Excellence at Intel.
"AI governance used to be, let's have policies. Let's look at safety. Let's see how the model interacts with humans," Comp told The Forecast. "Unfortunately, now AI governance has to be code."
In a video interview recorded in April at the 2026 .NEXT event in Chicago, Comp described how dramatically AI is changing as organizations move from experimenting with large language models to deploying autonomous agents that carry user credentials, access calendars and make decisions without a human in the loop.
Comp is a technology executive who has navigated the dotcom boom, the mobile rollout and the rise of the internet. She says the current AI wave is the fastest and most heavily marketed transition she has witnessed, and it’s creating a governance crisis: who owns, manages and controls AI agents that have become embedded in employee workflows, and what happens when the legal and policy frameworks built for a previous era simply do not apply.
Comp cited an MIT journal article that pegged this dramatic shift to sometime between December 2025 and January 2026. She described enterprise IT as if she was watching a crawling baby suddenly start to walk then feeling compelled to run. But the governance frameworks built for the crawling phase, Comp said, were not designed for what comes next.
Who Owns the Agent?
The ownership question cuts to the heart of the governance problem, according to Comp. She described a scenario that most enterprise IT and legal teams have not formally addressed: an employee builds and trains an AI agent that dramatically increases their output. When that employee leaves the company, does the agent leave with them?
"Probably not," Comp said. "So who manages the agents?"
The implied answer, that the company retains the agent, creates its own problem. If employees understand that any AI system they build at work belongs to the organization, the incentive to invest in building those tools on company time could drop sharply. The risk, Comp said, is baked into the ambiguity itself.
She see identity management as a parallel challenge. AI agents require their own user credentials to operate. Comp raised the question of whether a staff of agents should receive standard user IDs, or whether an entirely different class of identity and privilege management is needed. Enterprise IT teams built their access control frameworks around human users. Agents are not human, and the rules do not map cleanly.
Legal Frameworks Are Lagging
The governance gaps extend beyond enterprise IT into law. Comp noted that in the United States, attorney-client privilege and doctor-patient privilege do not apply when questions are posed to AI chatbots. Copyright protections do not extend to AI-generated content. For companies using AI tools to produce creative or strategic work, those gaps carry real exposure.
"Our policies, our laws, our governance," Comp said, "we are going to have to rethink things as dramatically as we did when people started getting connected with the internet."
The comparison to the internet's early years is deliberate. The legal and regulatory infrastructure that governs commerce, privacy and intellectual property on the web took years to develop, often trailing the technology by a decade or more. Comp's argument is that AI, moving faster and with higher commercial pressure than any previous wave of innovation, may not have that kind of runway.
A Hardware Analogy and a Working Model
The infrastructure underlying AI is undergoing its own transformation. Comp compared the current state of AI systems to the early days of mobile telephony, when the equipment required to run a single wireless signal filled a rack. The path from that moment to the smartphone in every pocket required a transition from programmable systems to application-specific integrated circuits, or ASICs, with hardened logic optimized for a single purpose. AI, she said, is on a similar trajectory toward optimization and proliferation.
Intel’s own IT team offers a working model for enterprises trying to navigate the transition responsibly. Comp highlighted Intel IT's practice of building a common AI interface that allows engineers to swap underlying models as better or more economical options emerge, avoiding lock-in to any single vendor or architecture. Intel also uses natural language processing to analyze technician logs from manufacturing facilities, detecting rising frustration in the written language as an early signal that equipment needs maintenance.
Perhaps most relevant to the governance challenge is Intel's approach to agent chaining technologies. Rather than opening them to the entire organization, Intel monitors a select group of trusted developers building with these tools, learning which use cases are safe and scalable before broader rollout.
"They watch how people are building with it, who they know are trustworthy and developers that are going to do it right," Comp said, "so that they know what they can replicate and they know what's going to be a popular use case."
Overcoming the fear and confusion Comp described at the start of the conversation requires preparation, vigilance and continuous improvement. Governance must be constantly refined and reinforced across IT systems.
Editor’s note: According to recent stories reported by The Forecast, building a successful AI and cloud governance strategy requires understanding what agentic AI means for IT operations, implementing strong governance practices, and learning how enterprise IT governs agentic AI at scale. Organizations must also address managing AI agent sprawl while navigating the unprecedented pace of IT innovation. For public sector organizations, minimizing risk of AI agents and understanding AI for government policies, challenges and solutions remain critical priorities.
Related:
Video Game Industry Gets Squeezed by AI
Video transcript:
Jason Lopez: AI tools. Who owns the productivity employees built into them? Who ultimately manages these tools, the employee or the company? Lynn Comp has been through enough technology transitions to know when an industry is running faster than it can think. And right now she says that's what's happening.
Lynn Comp: I see a lot of fear and confusion right now because I've been through a number of tech transitions and this one is happening faster and being marketed more. And there's more hype around it than I've seen in other tech transitions like dotcom, the mobile phone rollout or the internet. And so I think really the amount of stress that people are having over what is artificial intelligence, how is it going to affect me? Should I be using it? How should I be using it? Is probably the undercurrent of the entire tech industry right now.
Ken Kaplan: AI trust and safety?
Lynn Comp: Safety. AI governance, security. I actually had an IT person at a company that openly admitted when they saw OpenClaw, like what fresh hell is about to become shadow IT? There's a lot that is really starting to pop that's exciting because nobody really loves to have to do all of the grind themselves. They'd really like to have a personal assistant. But when something needs your user ID credentials, admin privileges, your credit cards, your calendar, you're taking a really interesting risk. I had somebody tell me today, "Oh, well, my information's all there on the darkness." Yeah. But you need to really think about it when you have something making autonomous decisions that is non-deterministic. So lots of great experiments, lots of excitement, but lots for us to really keep an eye on and try and figure out what does it mean to have admin privileges or the right privileges for a staff of agents.
They're humans. Is it really a user ID that you give them or is it something else? What do you do when an employee leaves the company? Do they take their agents with them? Probably not. So who manages the agents? So many things that we have not thought about. I had a blog post recently about how essentially, actually it was an MIT journal article that between December 2025 and January 2026, it's the same experience for IT that new parents have when their baby that was crawling starts walking. Because the whole idea of keeping them safe, it's a new ball game.
Ken Kaplan: You're making me think, if you have all these agents everywhere. Is it even possible to have sovereign AI?
Lynn Comp: There's a whole question. AI governance used to be, let's have policies. Let's look at safety. Let's see how the model interacts with humans. Unfortunately now AI governance has to be code. And you have to be ahead of questions around if someone has built an AI agent that has made them significantly more productive, is that the intellectual property of the company? Because if it is, then what do employees do when they've built and invested in this thing? It just by definition invites shadow IT. So there's just some things that we don't have worked out yet. I was talking with someone today at a lunch table about the fact that in the US, attorney-client privilege and doctor-patient privilege do not count when you're asking questions to chat bots. There is no copyright protection when you're using AI tools. So if you're trying to self-publish on Amazon, please don't use an AI image generator unless you actually can do a lot of work on the image to prove that it was yours. Because those three examples right there, it just says our policies, our laws, our governance, we are going to have to rethink things as dramatically as we did when people started getting connected with the internet.
Ken Kaplan: It feels like we are moving also towards people doing more AI on their own. Can you talk about where we're going with that?
Lynn Comp: Yeah. I used this analogy earlier today. And if you remember or you look on the internet, depending on if you're Gen Z or Gen X, when mobile phones and mobile telephony first came out, the systems that you were carrying were the size of a box. There were suitcases. If you go to Mobile World Congress when they're announcing the new wireless technology and they've just got it up and running, the motherboards are not phone size. They're rack size because they're still working out the kinks and they're still programming and tweaking the configurations for the wireless technology. With AI, we're in that kind of situation. So you got these big systems for it to proliferate and for it to really be everywhere the way that mobile telephony is, essentially it's going to have to go through optimization. You're going to have a bunch of heterogeneous types of solutions. You're going to have to go from programmable to ASICs in some cases, which is hardened logic. It only does one set of things. So what a phone call used to look like on something that was programmable is now a hardened set of logic that only does that wireless technology. So you're going to see something similar in terms of wireless, that evolution with AI and the proliferation of the different solution types.
Ken Kaplan: It goes back to what you said at the very beginning about fear. How are people getting through this time?
Lynn Comp: What Intel IT has done that I think is really, really smart. They've always been a eat your own dog food type of IT department. And they publish a lot about what they've implemented. So they've done everything from taking natural language processing to read tool logs in manufacturing facilities. Because as it turns out, when the technicians start getting really tense in their language on logging using the tool, that's because the system needs maintenance and it's about to go down. So now when they sense that the technicians are getting really pissed off, they know, oh, we should probably do proactive maintenance. So that's one use case. But they also do things where they have a common interface and underneath it you can swap models out. So as they find different models that are either more economical or better for a task, you don't get stuck on one model. And then they have also had a trial of things like some of the chaining, the agent chaining technologies. There's a few out there. And they watch how people are building with it, who they know are trustworthy and developers that are going to do it right so that they know what they can replicate and they know what's going to be a popular use case.
Related Articles
