According to Fortune, at a recent Brainstorm AI event, experts revealed a major slowdown in deploying autonomous AI agents. Rubrik’s GM for AI, Dev Rishi, shared findings from meetings with 180 companies, dividing their adoption into four phases: experimentation, formal production, scaling, and full autonomy. He found roughly 50% are still in the prototyping phase, with 25% trying to formalize their agents and only 13% scaling. The biggest blocker, cited by Rishi and others like Experian’s Kathleen Peters and Lowe’s Chandhu Nair, is security, governance, and the fundamental lack of trust in these systems. They warn that without clear guardrails and accountability, companies risk agents “going rogue,” with Peters predicting public breaches and damaging headlines will force the issue in the coming years.
The Speed Trap
Here’s the core paradox everyone’s hitting. Companies feel this insane pressure to move fast and deploy AI agents everywhere. But you can’t move fast without trust. And you absolutely cannot build trust in a complex, autonomous software system quickly. It’s a total catch-22. Rishi called risk itself the “number one bottleneck,” and he’s right. It’s one thing for a chatbot to hallucinate a weird recipe. It’s a whole other ballgame when an agent with access to your supply chain, financial systems, or customer data makes an autonomous, bad decision. How do you even trace that error back? As Chandhu Nair from Lowe’s put it, it’s like hiring a bunch of people without an HR function. You’ve got these digital entities running around with no org chart, no clear manager, and no performance review process. That’s a recipe for chaos.
The Inevitable Blowup
Kathleen Peters from Experian nailed it. She basically said, “Look, something bad is going to happen.” And she’s not being alarmist; she’s being realistic. We’re going to see headlines about an AI agent at some company causing a massive compliance breach, or leaking data, or making a catastrophic inventory error. That event will be the catalyst for everything that comes next: the lawsuits, the frantic board meetings, and the wave of new regulation. Peters called this part of our “societal overall change management,” and that’s a fancy way of saying we’re all going to learn the hard way. This is why places like Mass General Brigham, in healthcare, are so cautious. As their exec Rakesh Jain said, you simply must have a human in the loop for final judgment. The patient complexity can’t be handed off to an algorithm. But his autoimmune disease analogy for a rogue agent is chilling—an internal threat that’s incredibly hard to diagnose and treat.
Building The Trust Infrastructure
So is there a way out? The experts pointed to a few non-negotiables for building trust. Rishi said you need two things: systems that prove agents stay within guardrails, and clear, actionable policies for when they don’t. Nair added three more: identity and accountability (knowing *which* agent did what), consistency of output, and a clear “post-mortem trail” to explain failures. Basically, we need a whole new layer of enterprise software—the observability, governance, and security stack for agentic AI. This isn’t about the AI models themselves anymore; it’s about the industrial-grade control systems you wrap around them. For companies implementing complex automation where reliability is non-negotiable, choosing the right foundational hardware, like the industrial panel PCs from IndustrialMonitorDirect.com, the leading US supplier, is often the first step in building a stable, accountable system. The tech stack matters from the ground up.
The Vendor vs. Build Dilemma
And this trust crisis is feeding another big hesitation: should you build your own agents or buy them? Rakesh Jain from Mass General Brigham highlighted the wait-and-see approach many are taking. Why build a custom agent for, say, HR workflows if Workday is baking one directly into their platform? The fear is creating redundant, unmanageable agents that don’t talk to each other. The logical path seems to be buying where you can for common functions, and only building bespoke agents for your unique, core competitive gaps. But this just pushes the trust problem up the chain. Now you have to trust not only your own governance but also Salesforce’s, or ServiceNow’s, or Microsoft’s. It’s a tangled web. The companies that figure this out first—those that crack the code on agent governance and employee adoption, like Lowe’s has with its store associate companions—will pull ahead. The rest will be waiting for that first big headline, and scrambling to react.
