AI Insights: ManageEngine’s Sujatha S Iyer

Enterprise AI has moved well past the question of whether to adopt it. The harder, more consequential questions are now operational: who governs what an AI agent does, how failures are contained before they propagate, and what it actually means to trust a system that makes decisions faster than any human team can review them. These are not philosophical concerns. They are infrastructure problems, and they are arriving at scale. 

The current wave of agentic AI is testing assumptions that most enterprise IT environments were never built to handle. Agents don’t just process requests; they act, call tools, access data, and trigger downstream workflows. As those systems become interconnected, the risk surface grows in ways that traditional perimeter security cannot address. The boundary between a model’s capability and an organization’s operational policy is blurring, and the gap between the two is where most of the real exposure lives. 

Earning autonomy 

The fundamental unit of risk in an agentic environment is not the model itself, but the scope it is allowed to operate within. Capability without constraint is where the exposure begins. 

“Autonomy should be introduced in layers. Low-risk, repetitive tasks can be handled with more freedom, but anything that affects identity, access, security settings, or critical workflows needs tighter guardrails. That means agents should operate within a clearly defined scope, with limited permissions, explicit policies, and human oversight where the impact is high,” says Iyer. 

That layered approach reflects a broader shift in how security practitioners are beginning to think about AI governance. The old model, where policy sat outside operations as a review layer, does not hold in environments where AI systems are making decisions in milliseconds. Governance has to be embedded in the architecture itself, not applied after the fact. For most enterprises, that is not a minor adjustment. It is a fundamental rethinking of how IT operating models are structured. 

“The best approach is to start with narrow use cases, measure where the system performs well, study failure patterns, and then expand carefully. The goal is not to maximize autonomy for its own sake; it is to build systems that can act independently without creating instability. At enterprise scale, control is what makes autonomy usable,” she adds. 

What makes this particularly difficult is the interconnection problem. Once AI systems begin talking to each other, accessing shared data sources, and triggering actions across platforms, the question of accountability becomes genuinely complex. Which agent acted? What context did it use? Which tools did it call, and did it stay within policy? Without answers to those questions, monitoring becomes guesswork, and the risk surface becomes effectively invisible. 

“The most important capabilities are identity-aware access control, observability, policy enforcement, and strong data boundaries. Enterprises need to know which model or agent acted, what context it used, which tools it called, what data it touched, and whether it stayed within policy. Without that visibility, you cannot monitor or secure the system,” Iyer explains. 

Quality and discipline 

The conversation in enterprise AI has been dominated by model quality for the past two years, with organizations chasing the most capable foundation models as though capability were the primary variable. The operational reality is more complicated. A highly capable model operating without observability, without clear data boundaries, and without fallback mechanisms is a liability, not an asset. 

“At scale, resilience matters just as much as intelligence. You need fallback paths, escalation mechanisms, and the ability to contain failures before they spread,” says Iyer. 

The complexity compounds further as AI becomes embedded in everyday IT operations. Deploying a model is no longer the hard part. The hard part is managing how models, agents, tools, workflows, and enterprise data all interact in a controlled way across an organization’s existing systems. Teams need visibility into how AI is using context, which actions it is triggering, and how decisions are flowing across connected infrastructure. 

“Governance can no longer sit outside operations as a separate checkbox. It has to be built into the system itself. The shift is from managing isolated tools to managing connected decision systems,” she notes. 

That shift is wider than most roadmaps currently account for, and closing it requires more than technology investment. It requires a different way of thinking about what production-grade AI actually demands, one where operational discipline carries as much weight as the capability of the model itself. 

Raising the bar 

There is also a physical dimension to this that rarely makes it into the governance conversation. AI at scale is an infrastructure problem as much as a software one. Power, bandwidth, latency, and cost become binding constraints once AI systems move from pilots into production workloads. Enterprises that have been designing around model performance are beginning to realize that the full-stack question, encompassing compute architecture, energy efficiency, and right-sizing models to actual task requirements, is just as consequential. 

“Organizations need to think beyond bigger models and start thinking in terms of better system design. Scalability has to be treated as a full-stack challenge, from models and data access to compute architecture and energy efficiency,” Iyer points out. 

As AI moves into real-time environments, including monitoring, incident response, and service operations, the tolerance for failure narrows considerably. Slow systems break workflows. Failures in connected environments propagate quickly when there are no controls in place. The standard shifts from research-grade to production-grade, and the expectations that come with that are unambiguous. 

“The expectation is shifting from ‘can the model do this?’ to ‘can the system do this reliably, repeatedly, and under real operating conditions?’ That is the bar AI now has to meet in enterprise IT,” she concludes.