Survey finds why 84% of enterprises are adopting AgenticOps

Within the next 12 months, 84% of IT and network operations (NetOps) leaders expect their organizations to have an AI-led operating model, according to a new survey from Cisco. Already, enterprises are scaling AgenticOps and trusting agents with high levels of and even complete autonomy. 

The survey, conducted independently by Omdia, a division of Informa TechTarget, gathered insight from more than 1,000 IT and NetOPs leaders working at organizations with at least 500 employees. The results offer insight into the operational complexity, creating the need for AgenticOPs, and the guardrails enterprises view as necessary to continue scaling agentic AI for NetOps. 

In this Q&A, Joe Vaccaro, senior vice president/general manager, network platform and Assurance at Cisco, answers questions about the way AgenticOPs is taking shape in enterprises today and how it will drive the future of NetOPs.

Editor's note: The following interview was edited for length and clarity.

How is operational complexity putting pressure on traditional network operations in enterprises today?

Joe Vaccaro: AI workloads are now becoming an increasing part of enterprise infrastructure. It's changing network traffic patterns and adding to operational complexity. 

When we think about a network, we often oversimplify it to a single connection between a client and an application. But the reality is that a network is a highly choreographed orchestra that connects a client to the application across many segments of the network or across multiple networks.

Customers have seen a number of shifts related to that connectivity: the rise of cloud infrastructure replacing where workloads ran, the rise of internet-centric network connectivity replacing traditional corporate networks, and the rise of SaaS applications.

Joe Vaccaro

Now, AI is becoming a changing paradigm for how connections matter. As this complexity increases, traditional AIOps and human-centered operating models are reaching their practical limits in maintaining resilient network connectivity, which is paramount to delivering a great digital experience.

The survey found that 95% of enterprises say existing AIOPs tools can't keep up. Why are these tools not enough?

Vaccaro: When degradation happens or performance challenges arise, very rarely do they stay within a single domain. We saw that 92% of the organizations surveyed identified common issues spanning networking, cloud security and applications.

Customers today who are using different AIOps tools are often focused on a particular segment of the network or a particular domain. But connectivity issues can arise across many domains or segments of the network.

At the same time, reliance on the network is increasing. People talk about accessing their AI tools. Well, AI depends upon the network.

If your legacy AIOps tools are simply giving you recommendations but not actually able to drive action, then you're going to constantly be in a reactive position, trying to keep up with issues and trying to solve issues, versus: How do I spend time as an operations team trying to improve the network itself?

Enterprises are quickly moving to agentic-driven operations: AgenticOps. What does an AI-led operating model look like?

Vaccaro: In many ways, the math of a human team no longer works. In the survey, we found that the average organization receives over 4,000 monitoring alerts per day, 51% of which are networking-related. If you have a practitioner who can clear 21 alerts a day, you would need a team of 100 people just to clear your backlog.

The reality is that we know that they don't have that many people. A large majority of alerts go completely ignored or dismissed without finding the root cause. Now, some of that might be noise within the system, but there are issues that are going unaddressed that are creating frailty within the network itself.

What we see is that by using agentic AIOPs and thinking of the agents as teammates. These teammates, when provided with cross-domain context, can help offload and augment humans. They can go through this agentic loop and help you continuously sense your environment across domains. Then, they can reason across cross-domain telemetry in context to identify problems in real time, not just suggest remediation steps, but put them into practice with a single click.

In many ways, networking and network operations are the proving ground for agentic AI in many organizations. Because at the end of the day, running a resilient network that delivers performance is critical to how businesses and organizations need to operate.

What kind of NetOPs tasks are AI agents taking on in enterprises today?

Vaccaro: What we do see is that NetOps leaders are now more open to autonomy, but they're also very clear about the conditions required to scale it.

The leading organizations we work with are identifying specific use cases and recurring issues across their networks that have a very strong signature in terms of the root cause and a very deterministic path for how you'd resolve them. Then, as agents can identify those issues, they can apply that level of determinism for how they go about solving it.

The key thing is we talk about trust. How do you build trust? In the survey, we found that 69% of organizations require detailed explainability for an agent-driven action. That might even be underrepresented in terms of importance. The reality is that to offload and delegate, you need to ensure the agent finds the root cause you would expect. You do that by showing your work. You need to be able to show the reasoning for how you got to that conclusion.

Over time, we will continue to build that level of trust and delegate more and more use cases to agents to scale the impact of network operation schemes.

The survey results indicate a high level of comfort with agent autonomy in NetOPs. Where are enterprise leaders comfortable reducing or eliminating human oversight?

Vaccaro: To drive autonomy, the guardrails are not an optional step. The survey noted nearly 100% of organizations require at least one guardrail to trust before AI acts.

What's an example of a guardrail? In general, agents relying upon an LLM give a probabilistic response. By nature, that's how LLMs operate. What we see is a probabilistic system to help you understand context and derive insights through sensing, reasoning and diagnosing, but you pair that with deterministic execution.

When you come to this conclusion, instead of probabilistically making up how you would solve it, I'm going to dictate: here's how you solve it based on our organizational standard operating procedures or runbooks. Then it becomes a system that provides an additional level of trust. That's a key point, both in the survey findings and in what we see in practice as organizations adopt an AgenticOps mindset.

There are many reports indicating that AI adoption is outpacing governance and guardrails. What are the risks if that is the case for enterprises granting AI autonomy in NetOps?

Vaccaro: It's important not to wipe autonomy across the entirety of network operations all at once. You need to identify specific use cases where you can understand how it's going to work in practice and then evolve with it.

Governance is important for understanding how these agents will respond and for showing their work before they move to the next step, to ensure that there is potentially human oversight or approval before they move into action.

But over time, as a level of trust and governance is established, we can begin to use the agents across a wider range of network operations tasks. That helps a team scale. I think we will continue to have humans at the center of network operations. That doesn't change. But now, as a team, they're able to focus their time on where they want to apply judgment rather than on keeping up with the flood of alerts that potentially hit them daily.

Scaling agentic AI requires trust, visibility and control. How can enterprises achieve those things as the adoption of AgenticOps accelerates?

Vaccaro: It starts with that visibility. If you only have visibility into, say, the wireless segment of your network, an agent will always think that's the root of the problem. The reality, though, is that an issue might surface in how people connect to your wireless network, but the root cause might be upstream within your DHCP service. Or it might be upstream in your WAN connectivity. Or it might be the application itself. Do you have the right level of cross-domain telemetry and an understanding of the system as a whole to give agents the broadest set of context to reason across?

The second is control -- the ability to have that level of deep intelligence.

And then, it's about trust. When an action has surfaced, do you have the ability to see the evidence chain for how they drew that conclusion, to understand the confidence and risk score, and to know, within these guardrails, which issues need to be elevated to human oversight and which can be taken fully autonomously? If the confidence or risk score is elevated, you need to surface it for a human-led investigation.

What can enterprise leaders expect for the future of NetOPs?

Vaccaro: This world is moving fast. At a practitioner level, their teams are already moving in this direction. In many ways, they're moving in this direction because they have no other choice. As a network operations team, their job is to operate the day-to-day heartbeat of their company across their corporate networks. And as business leaders, they need to empower their teams to leverage agentic operations to scale while maintaining governance and trust, ensuring a positive impact on their overall business and objectives.

One of the things that came out in the survey is that we're seeing AI traffic across networks doubling every six months. How do we use AI to be able to augment network operations? But we also need to think about how we augment networks as agentic traffic rapidly increases across the environment.

That's going to mean how do we think about agent security? How do we think about AI defense and model security? How do we think about identity across these agents? These will all be important in building confidence and trust in the systems. It's not just about how we operate our network, but also about how we secure and trust the agents operating alongside us.

Carrie Pallardy is a freelance journalist with experience writing in cybersecurity, technology and healthcare. She currently covers a wide range of issues relevant to today's CIOs and IT leaders.

Omdia is a division of Informa TechTarget. Its analysts have business relationships with technology vendors.