The Most Important Question in Operational AI: Show Me Where It Actually Works

The moment I left the whiteboard and entered the network operations center, I understood just how much of the shiny marketing fluff of our industry melts when exposed to the brutal truth of an actual outage. We’ve all been there: the storm of notifications that blinds the SREs in charge of the situation, the real problem buried behind a dozen different SRE tools, and, of course, the clock ticking away the minutes until the problem can be solved. It’s stressful. And honestly? Exhausting.

Our operational environments used to be limited in the early years. They are currently spread across multi-cloud infrastructures, edge computing components, and vast services. The level of complexity has exponentially increased to the point where the method of adding more ops tools to address the issue has become not only ineffective but also harmful. In our case, there’s just too much information available.

The past few years of hype about AI being the “solution” are behind us. Almost every platform provider in the world of infrastructure, cloud computing, and networking today says they’re “powered” by AI. However, from my experience as a veteran leader who’s also spent plenty of time in the trenches, the question being asked today hasn’t changed. Instead of asking “Are you using AI?” today, we’re asking a far more pragmatic, or even brutal question: “Where’s the AI actually being used, and what’s it allowing my team to do that they hadn’t been able to do before?”

This is essential. This is the difference between investing in an intangible buzzword of a capability and investing in understanding the business. The value of operational intelligence isn’t in the scale of the model and the marketing budget; it’s the way this system thinks about your messy world.

See also: Scaling Agentic AI: The Emerging Role of the Model Context Protocol

The Foundation of Trust: Why Context Trumps Code

It’s time to talk about data. Far too often, teams attempt to tack a massive complex model of machine learning onto a messy heap of siloed data from their monitoring. They hope the model will somehow see the patterns themselves because the data available has been messy and illogical, and doesn’t necessarily provide the context. It’s the difference between being lost in a new city and possessing a box of road signs ripped from the earth. No matter the speed of reading, without the information of their placement, they do nothing.

This is the dilemma of what our modern hybrid cloud environment looks like: the network state is always shifting. The familiarity of the network interface changes, the devices shift roles, and the services move up and down through the clouds. If the AI/ML system is looking at the raw data points in a singular manner – without context – the best the AI/ML will be able to do is provide an intelligent guess. This results in the system churning out unnecessary white noise while also exacerbating the existing problem of fatigue regarding the already existing notifications.

A better model of intelligence would begin not with the model itself but with the data it’s based upon. Even before the first model has been executed, there must be data enrichment happening in this system. The system must understand not only that a value has occurred but also where the device exists, what it connects to, and its role in the overall application delivery process. To understand first and interpret second—the context above the data—is of the utmost importance. By this point, the model has a solid platform upon which it can truly learn through the machine learning algorithms.

See also: Amplifying Agentic AI’s Benefits with Collaborative AI Agents

From Symptom Management to Surgical Diagnosis

A smart engineer will recognize that problems seldom occur in a vacuum. Instead, there’s almost always a series of reactions to consider. What might begin as a basic Border Gateway Protocol event concerning the core router can cause flickers of performance directly impacting the content delivery networks, which in turn generates notifications of performance issues within numerous cloud applications. The initial cause becomes twenty additional problems scattered about without apparent connections. Our teams are stuck right smack in the middle of the war room situation, trying to work the various discrepancies in real-time. This explains why the problem resolution time remains so irritatingly long—it’s costing the company real-world dollars.

Good operational intelligence has to emulate the thinking process of an experienced engineer: observing, hypothesizing, correlating.

It must be able to see not only the signals but also understand the patterns of co-occurrence of various factors across the domains of the network, infrastructure, and application levels over time. When a change occurs, the system must point not only at the signal but also provide a clear understanding of cause and effect: This routing change caused the connectivity degradation because the application there is experiencing errors.

The results should not be presented as a simple list of alarms but rather as a clear story of what occurred and in what order. This allows the operation to shift from reactions based on symptoms to a surgical and precise diagnostic approach that can enable engineers to move from the diagnosis to the solution.

Trust and the Human Interface

The final piece of this puzzle that can perhaps be the most overlooked has been the role of the human operator. There has been a misconception about the end goal of AI development: removing the human from the decision-making loop. Having been in the trenches, I can say definitively that this is not the case. Clearly, the role of the human has been the decision-making role. The role of the human has been to understand the context of the business situation and decide the level of aggressiveness.

The value-for-value exchange is in collaboration. Machine intelligence can handle volume: analyzing billions of data points, building baselines, and making connections across events. Expertise has the needful interpretive and prioritized components of the decision.

This can only happen if there’s no need to understand the reason behind the system’s decision. What the system thinks is the cause of an event or the connection between two seemingly unconnected behaviors has to be explained to the operator. A black box would immediately undermine trust. The engineers would be forced to manually confirm everything just to be sure, rendering the value of the intelligence worthless.

This explanation thing isn’t a nicety – it’s the vital conduit of the partnership between the human and the machine. The man can then check and even correct this understanding of the workings of the system, and this will also be used to improve the intelligence the next time around. This intelligence becomes a living and evolving asset to the man rather than an unmanageable technological platform.

Leaders must understand that high-quality contextual and explainable operational intelligence is not only about the tech—it’s about the choices that empower the most important tech teams in the company. By pivoting from models to context and through insights together, we can at last get beyond the hype and unlock the real value of ML to keep our complex digital world in harmony. This is the only way to achieve operational resilience through the investment of understanding.