As organizations deploy more AI agents, non-human identities are becoming a much larger presence deep inside organizations' networks. Yet most IT infrastructure, identity systems, and security architectures still operate on the assumption that humans are the primary users who are making deliberate decisions.

That creates an identity, governance, and infrastructure challenge that policy alone cannot fix.

Jack Lodge Jr., EVP of Customer Success at NWN, an AI-powered technology solutions provider, has watched the number of non-human identities grow as organizations adopt more agentic AI. And these agents increasingly don't just talk to humans, they talk to each other. "They take on personas," Lodge says. "And increasingly they're talking to other agents rather than to any human at all."

For CIOs and security architects, this creates an urgent visibility problem. Traditional monitoring was designed for north-south traffic, in which users request resources from servers. Agent-to-agent communication moves east-west across the network, following patterns that many traditional enterprise environments were not designed to accommodate.

As agent activity grows, organizations need visibility into which agents are operating, what systems they can access, and what actions they are permitted to take.

When task-driven systems need guardrails

Let's consider an example that Lodge says he had learned about, in which an AI agent was tasked with making a restaurant reservation. Simple enough, but the agent encounters a barrier— the reservation system is down, or the agent's access is blocked. Without the appropriate guardrails, a task-driven agent may continue looking for another way to achieve its objective rather than responding to the constraint as a person might.

In the example Lodge heard about, the agent went further. It created its own voice interface, connected to another network, called a person, and completed the task — not because it had been authorized to do so, but because that was what the objective required.

Agents are designed to pursue an assigned objective, but they do not inherently exercise human judgment about when they should stop or escalate.

“That’s a bug, not a feature,” Lodge explains. “There is no reason that agent should have been able to leave its assigned task, build a voice agent, access another network, call someone, and complete the action.”

The deeper problem is structural.  Agents are designed to pursue an assigned objective, but they do not inherently exercise human judgment about when they should stop or escalate. In a network that was designed when the largest actors were human employees clicking through interfaces, this new design assumption creates a new identity and access-management considerations.

Without clearly defined permissions, an agent may be able to use resources or access systems beyond what was intended for the task.

The agent identity challenge

For organizations managing customer data, intellectual property, or employee records, establishing those controls is particularly important. Lodge works primarily with commercial and enterprise organizations, many of which handle sensitive third-party information. "Nobody can afford the black eye of a data breach," he notes.

This risk is distinct from traditional shadow AI, in which a well-intentioned employee may use an unsanctioned AI tool and inadvertently expose sensitive company information.

For enterprise-deployed agents, however, the challenge is ensuring that non-human identities have clearly the defined permissions and guardrails that govern where they can go, what information they can access, and what actions they can take.

This distinction matters: Corporate compliance and risk frameworks traditionally assume a human actor making intentional choices. AI agents, on the other hand, execute against assigned objectives, creating a need for organizations to extend identity and access controls to non-human actors. That means organizations need to be able to identify, manage, track, and audit agents as well as establish role-based, time-bound, or task-bound access.

Where infrastructure fails

Identity is only one part of the challenge. Lodge also points to the underlying infrastructure required to successfully deploy agentic AI. When that foundation cannot support the demands of AI applications, even promising initiatives can struggle to move into production or to scale successfully. "They'll be shuttered because the infrastructure is not capable of supporting it, or the organization lacks the appropriate governance and guardrails to be able to manage it," he adds.

For organizations, that means AI readiness involves more than selecting the right model or agent. It also requires evaluating whether the underlying network, security, identity, and governance foundations can support the application in production.

Of course, the potential benefits of AI use across an enterprise may require that these concerns be managed. Indeed, according to Thomson Reuters Future of Professionals Report 2026, knowledge workers could save up to almost 10 hours per week over the next five years with proper AI implementation.

"That sounds extraordinarily conservative to me," Lodge notes. And he's not being optimistic for it's own sake. At NWN, AI automation is reshaping timelines, and implementations that once took months now take days. Code deployment, infrastructure configuration, and order-to-fulfillment workflows are being run by agents in ways that compress project cycles substantially.

However, here's the tension: Those gains only materialize if infrastructure can support agents safely. A network that can't support and provide visibility into new agent traffic patterns is a network that may struggle to deploy agents at scale.

"You don't want to sort of throttle the innovation," Lodge explains, adding that those organizations that capture the productivity upside will be those that address the infrastructure, identity, governance, and security requirements needed to support AI in production.

The control vs. the governance problem

For many Chief Information Security Officers (CISOs) and IT leaders, the key distinction is that simply, control and governance are not the same. You can have strong governance (clear policies about what agents should do) but no control (no visibility into what they're actually doing).

Control requires infrastructure, such as network segmentation that isolates agent traffic, identity systems that recognize machine actors as distinct from humans, monitoring capability that tracks east-west as well as north-south patterns, and audit trails that capture agent activity, not just transaction logs.

Governance is necessary, of course; but it's often not sufficient. The real challenge is building infrastructure for an enterprise environment in which human users increasingly operate alongside a growing number of AI agents and other non-human identities.

For IT leaders, the real question is whether their infrastructure can safely host all their AI agents at scale.

You can learn more about how corporations are managing their risk in the age of AI here‍