Building an AI-Ready Network: Why Your Network Observability Practice Has to Come First

This article, which builds on insights from a Solutions Spotlight with Broadcom, outlines the importance of network observability when building an AI-ready network.

Every major enterprise IT conversation right now circles back to AI. Whether the discussion is about autonomous operations, intelligent triage, or predictive remediation, the assumption underlying all of it is that the infrastructure that powers those capabilities is solid. For most network teams, that assumption is doing a lot of heavy lifting.

According to Broadcom Chief Technical Evangelist Jeremy Rossbach, 95 percent of network teams are still operating reactively when performance issues arise. That number tells you something important: the industry conversation about AI readiness has moved considerably faster than the operational reality on the ground.

What It Takes to Build an AI-Ready Network Observability Practice

The core argument Ross makes in a recent Solutions Spotlight is straightforward: AI-driven network operations tools need data. Specifically, they need clean, normalized, correlated data collected across every layer of your network environment. If that foundation is not in place, the AI layer on top of it produces outputs that are, in his words, flawed, slow, or unreliable.

The analogy he reaches for is a good one. An AI copilot for your NOC is the cherry on top, but without the cake underneath it, the cherry is decorative at best. The cake is a mature network observability practice, and organizations that skip ahead to AI tooling without building that practice first are setting themselves up for disappointment.

The Five-Stage Maturity Model Network Teams Should Measure Themselves Against

Broadcom has developed a network observability maturity model that maps the journey from reactive, siloed monitoring to AI-enabled autonomous operations. The model breaks into five stages, and Ross is candid that most organizations he speaks with at industry events fall somewhere between stage two and stage three.

The early stages involve moving from manual, reactive operations inside the data center to full visibility across both traditional and software-defined infrastructure. Stage three is where many teams stall: they have solid visibility within the four walls of their environment, but have not extended that coverage outward to public ISPs, cloud providers, and the unmanaged network paths their users actually traverse every day.

That gap matters more than it might seem. When a user anywhere in the world calls the helpdesk about a degraded experience, and the path their traffic takes runs across infrastructure the organization does not own, a NOC team without external visibility is left to guess. The data simply is not there.

Stages four and five involve doing something meaningful with all that collected data: normalizing it across vendors, automating workflows, enabling predictive remediation, and eventually creating the feedback loops that AI systems depend on to function reliably.

Why the Data You Have Not Collected Is the Bigger Problem

When Ross surveys audiences about their data readiness, the issue they surface most often is not that their data is bad. The problem is that there are entire categories of data they have never collected.

A network that monitors only within the data center perimeter misses the majority of the paths users actually experience. Modern application delivery runs across managed and unmanaged infrastructure simultaneously. Cloud providers, third-party SaaS platforms, residential ISPs, and mobile networks: all contribute to the end-user experience, and none are visible without deliberate instrumentation.

Multi-vendor environments compound this. A Cisco device and an HPE device speak different languages. Without a normalization layer, that data surfaces in separate consoles, requires specialized expertise to interpret, and cannot cleanly feed into a unified AI analysis pipeline. Good data in, good recommendations out. Missing or fragmented data in, unreliable outputs that erode trust in the tooling.

What AI-Assisted NOC Operations Actually Look Like in Practice

In the Solutions Spotlight, Ross walked through a demo of Broadcom’s network observability remediation manager to illustrate the end state of a mature AI-assisted NOC practice. The scenario is worth understanding in detail, not as a product pitch but as a model for the kind of capability a well-built observability foundation makes possible.

A NOC operator can ask Broadcom’s tool which alarms need attention, and then receive a response that identifies the critical device, pulls correlated data across alarms, performance metrics, and flow information, and surfaces not just the symptom but the probable root cause. In the demo scenario, an older router is being overwhelmed by a backup job routed through it incorrectly, resulting in CPU and memory spikes and packet loss. The tool traces the pattern, correlates the timing, identifies the traffic source, and provides ranked remediation options.

The point is not that AI replaces the operator, though. The point is that the operator spends their time making decisions rather than hunting for context. That shift in the work has real consequences for the mean time to resolution and for how effectively L1 staff can handle issues without escalating to senior engineers.

Why Network Automation Has Been Slow to Mature

Ross addresses the automation question directly and without hedging. Network automation has not taken off at the scale the industry anticipated, and the reasons are partly cultural and partly practical.

Network engineers are responsible for the infrastructure, where a single configuration error can bring down operations across an entire organization. Handing that responsibility to automated processes requires a level of trust that has to be earned incrementally. The use cases that are gaining traction are low-risk and high-value: rolling back configuration changes, enriching tickets before escalation, and handling routine off-hours maintenance tasks with human review built into the workflow.

The organizations pushing toward full automation without that incremental trust-building are, in Ross’s view, taking on more risk than the efficiency gains justify. The more productive path is using AI to accelerate human decision-making rather than replace it.

FAQ: AI Ready Network Observability

What does AI-ready network observability mean?

An AI-ready network observability practice collects comprehensive, normalized data across all network layers, including unmanaged paths such as public ISPs and cloud providers, and applies correlation and analysis that AI operations tools can reliably use.

Why do most networks fail to support AI-driven operations today?

Most teams have strong visibility inside their own data centers but limited coverage outside those boundaries. AI operations tools require complete, correlated data to produce reliable outputs; gaps in coverage produce gaps in AI performance.

What is the Broadcom network observability maturity model?

A five-stage framework that maps the progression from reactive, siloed monitoring through proactive visibility, multi-domain coverage, automation, and AI-assisted operations. Most organizations currently fall between stages two and three.

How does AI assist NOC operations without replacing engineers?

AI tools like Broadcom’s Norm surface root cause analysis, correlate data across alarm types and flow metrics, and recommend remediation options. Engineers retain decision authority and execute actions, reducing time spent on manual data gathering.

What is the biggest data gap holding network teams back from AI readiness?

External visibility. Teams that monitor within the data center but lack instrumentation across public network paths, cloud providers, and third-party infrastructure miss the majority of modern application delivery paths.