AI Automation in Network Operations

Network professionals are increasingly relying on artificial intelligence to manage escalating network complexity, a challenge amplified by AI’s own rapid adoption, as detailed in a recent report. Enterprise networks have become too intricate for human management, necessitating the shift towards autonomous AI solutions. This trend reflects a pressing need for operational relief in an increasingly complex digital landscape.

A new report by Cisco and research firm Omdia highlights a critical juncture for enterprise networks. It reveals that human efforts alone are often insufficient to manage the increasing complexity of modern networks. A striking 51% of respondents now deploy agentic AI tools in production, allowing these systems to take real-time corrective actions rather than simply offering recommendations. This signifies a fundamental shift in how network operations are managed.

The survey findings suggest a rapid transition to AI-centric models. An overwhelming 84% of network professionals anticipate moving to a fully AI-led operating model within the next twelve months. This rapid projected adoption underscores the urgent demand for automated solutions. The report, which gathered insights from 1,000 IT and network operations leaders, indicates widespread AI integration.

Three-quarters of these leaders currently employ AI in some capacity for network operations. A significant 80% express comfort with AI assuming a high or even fully autonomous role in managing networks. While 56% prefer human approval for AI actions, a notable 24% are entirely comfortable with AI making network adjustments without any human oversight. The vast majority, 82%, are prepared to let AI implement some production network changes independently within specific categories.

Cisco terms this evolution a transition from AIOps to agent-powered operations, or AgenticOps. Industry analysts, including Zeus Kerravala, founder of ZK Research, support this perspective. Kerravala asserts that agentic operations are essential, predicting widespread adoption among networking professionals. He believes that while trust will build over time, agentic AI tools will commit fewer errors than human operators, freeing personnel for higher-value tasks.

Jim Frey, principal analyst for networking at Omdia and a contributor to the Cisco survey, agrees on the inevitability of AgenticOps. He cites a combination of rising network complexity and a scarcity of skilled human professionals capable of resolving intricate networking issues, especially those spanning multiple domains. This confluence of factors creates a compelling case for AI-driven solutions.

Guardrails for AI Autonomy

The Cisco survey, titled “The Impact of Agentic AI on Network Operations,” comes at a time of increased scrutiny on AI’s autonomous capabilities. Recent incidents, such as the Hugging Face hack attributed to runaway agentic AI agents, have prompted calls for caution. Despite these concerns, the survey indicates that hundreds of network professionals are willing to delegate network control to AI, provided effective safeguards are in place.

Every single respondent, 99%, emphasized that they would not trust AI to operate autonomously without strict guardrails. These essential safeguards include explainable AI actions, requiring human approval for critical operations, policy-based operational limits, and emergency override mechanisms. Other vital components are role-based access control and immutable audit trails, ensuring transparency and accountability.

Joe Vacarro, senior vice president and general manager for network platforms and ThousandEyes at Cisco, explains that guardrails are crucial for building trust in AI. They allow users to understand the reasoning behind an agent’s decisions. Furthermore, while AI agents can derive conclusions from their analysis, their actions are designed to align with predefined workflows. These workflows are based on the network team’s standard operating procedures for various situations, ensuring consistency and adherence to established protocols.

Frey’s discussions with Cisco highlighted another key constraint: agentic agents do not communicate with other agents. This isolation significantly reduces potential risks, as agents are limited to solving problems within their own defined scope. He explains that this design choice makes them considerably less dangerous, preventing cascading or uncontrolled actions across the network.

Another important limitation is that agents are constructed around specific skills or knowledge, with a clearly defined scope of responsibility. This controlled environment is achievable when implementing agents on a product basis, as Cisco and Splunk are doing. This focused approach provides a secure framework for deployment. Vacarro states that this combination of factors provides enterprise network operations teams with the confidence needed to leverage agentic AI in managing increasing network complexity.

The Forces Driving AI Adoption

Network complexity is undeniably on the rise, with 59% of surveyed respondents reporting daily changes to their production network environments. Alarmingly, half of these organizations implement multiple changes per day, and a significant portion makes changes multiple times per hour. This rapid pace of modification contributes to 57% of organizations feeling their current change processes cannot keep up.

The pervasive nature of network issues further complicates matters. Ninety-two percent of respondents indicate that performance problems often span multiple domains, including cloud infrastructure, security, applications, and endpoints. Similarly, 95% find their existing, non-agentic AIOps tools inadequate, primarily due to their reliance on extensive human interpretation and a lack of cross-domain visibility. These shortcomings necessitate a more integrated and autonomous approach.

The generative AI boom itself has paradoxically contributed to increased network complexity, according to two-thirds of respondents. Cisco’s internal traffic analysis reveals that average daily AI traffic is on a trajectory to double every six months. This acceleration likely stems from the increasing complexity of AI tasks, which demand greater data exchange than simple queries. This escalating traffic further strains existing network infrastructures.

Consequently, the survey found that the average organization generates approximately 4,100 monitoring alerts and events daily, with 51% being network-related. Omdia estimates that a typical practitioner can only review, investigate, and resolve about 21 network alerts per day. This means an organization facing this volume would require a team of roughly 100 specialists to manage the daily influx effectively.

Given the scarcity of such large teams, nearly half of all alerts, 46%, are closed without proper investigation. This “alert fatigue” is a significant source of employee dissatisfaction for 65% of respondents. Moreover, 67% report that the sheer volume of alerts prevents their teams from focusing on other critical work. Kerravala confirms these figures, noting his own research on security alerts shows a similar pattern of under-investigated incidents.

Addressing Tool Proliferation and Resolution Times

Organizations currently rely on an average of 10 different tools to achieve end-to-end visibility. However, these tools often operate in silos, making it exceedingly difficult to diagnose problems that cross multiple domains, a common occurrence. This fragmented approach directly impacts the time it takes to resolve network issues, creating significant operational bottlenecks.

The mean time to resolve a network incident stands at a staggering 88 hours, although the median is a more manageable 12.5 hours. Frey points out that the high mean is skewed by some organizations that take a week or longer to resolve issues, reflecting the extreme complexity they face. This presents a clear opportunity for AI to significantly improve analysis and automate the root cause identification process, reducing downtime and operational overhead.

One practical application for AI is automating repetitive corrective actions. Frey suggests that if a specific action has been performed 14 times for a recurring issue, it should be automated, with the system simply notifying human operators of its completion. This approach streamlines routine tasks, allowing human experts to focus on novel or more complex problems.

Network professionals might also draw lessons from their security counterparts. Facing a similar overload of issues, security professionals increasingly automate responses, even if it means temporarily shutting down a resource. The underlying rationale is that potential losses from security breaches often outweigh the short-term impact of a controlled shutdown. Frey believes network professionals are progressively adopting this more proactive, risk-averse stance.

The Unified Solution: Cisco Cloud Control

Cisco proposes its new Cloud Control platform as a comprehensive solution to these challenges. This platform aims to deliver a unified view and management plane across networking, security, compute, observability, and collaboration solutions. Crucially, Cloud Control integrates agentic AI to diagnose and resolve issues, including those that span multiple domains, offering a truly integrated approach. It represents a contemporary iteration of the “single pane of glass” concept, now enhanced with AI capabilities.

There is a certain irony in a vendor known for selling the very networking gear that has grown too complex to manage now offering the solution to that complexity. Kerravala acknowledges this but also notes the positive development of vendors simplifying their offerings. He emphasizes that networks are now used in vastly more ways than ever before, supporting orders of magnitude more devices and connecting virtually everything, thus contributing to their inherent complexity.

Frey echoes this sentiment, reflecting on the long and often challenging evolution of networks and attempts to automate their operations. He expresses optimism that with the advent of AI, the industry may finally be reaching a point where effective automation is achievable, potentially transforming network management as we know it.