Matilda Bailey stands at the forefront of the modern networking revolution, serving as a seasoned specialist with a particular focus on how cellular, wireless, and next-generation cloud solutions are converging with artificial intelligence. With a career spent watching network engineers navigate the exhausting “drill” of firing alerts and opening endless dashboards, she has become a leading voice in the shift toward agentic AI that prioritizes immediate answers over mere data visualization. In this deep dive, we explore how the industry is finally moving past the failed promises of first-generation chatbots to create truly ambient, context-aware systems that behave more like expert teammates than static software. We discuss the critical failures of early AI implementations, the psychological barriers to automation, and the technical necessity of a “context layer” to solve the “mean time to context” problem that plagues IT operations.
The traditional approach to networking data has long been criticized for being siloed, where a simple term like “client ID” might mean something entirely different depending on whether you are looking at switching, Wi-Fi, or SD-WAN. How does the implementation of a purpose-built context layer and a knowledge graph fundamentally change the troubleshooting experience for an engineer on the ground?
If you look at how we’ve operated for the last decade, we were essentially asking engineers to be the human glue between disconnected databases. Within a single vendor environment, you might have fabric, switching, and software-defined wide-area networking all using the same terminology for different data points, which creates a massive “unnormalized” mess when you try to plug it into a standard large language model. By building a dedicated context layer, we are finally mapping the actual relationships between users, devices, applications, and the environmental conditions of the network. This means that instead of a bot giving you a generic definition of a protocol, it can actually tell you why the third-floor conference room experienced a service degradation at exactly 2:00 p.m. last Tuesday. It transforms the AI from a search engine into a diagnostic tool that understands the specific nuances of your local architecture.
First-generation AI in networking was often criticized for being little more than a chatbot built on top of unconsolidated data, leading to a situation where engineers simply ignored the tool during real crises. What specific behavioral shifts are necessary to move AI from a “demo-friendly” feature to an ambient tool that engineers actually trust when things go sideways?
The hard truth we learned from watching focus groups is that when a real incident hits, the last thing an engineer wants to do is start a conversation with a chatbot. They have a muscle memory built on decades of pulling logs and pings, and they will revert to those old habits every single time if the AI feels like an extra step in their workflow. To fix this, the industry is moving toward “ambient” AI, which is always running in the background and reaches out to the user with a “Nudge” rather than waiting to be asked. We saw a staggering 900% increase in interaction once we moved the AI from a separate window into a proactive command bar at the bottom of the platform. This approach meets the engineer where they are, providing a pre-investigation before they even have to ask, which bridges that gap between a flashy demo and a production-ready tool.
Honesty in AI is a major concern, as general-purpose models are often programmed to be helpful to a fault, sometimes leading them to “hallucinate” or guess when they don’t have the data. How does an awareness scale or a built-in honesty mechanism change the relationship between the operator and the AI agent?
A bluffing AI is significantly more dangerous to a network than no AI at all, because a single wrong configuration change can take down an entire enterprise. To combat this, we’ve developed an awareness scale that forces the agent to determine its own confidence level before it ever speaks to a human. If the system realizes it has a gap in its data or capabilities—such as not being able to remediate a specific wired device yet—it is designed to say so plainly and then offer a constructive alternative, like packaging the evidence and opening a support ticket. This honesty builds a foundation of trust because the engineer knows the agent isn’t just trying to be “helpful” by making things up. By being transparent about its boundaries, the AI becomes a reliable partner rather than a unpredictable black box.
The shift toward autonomous networking is often met with skepticism, yet we are seeing a clear path from “Coworker” modes to “Operator” modes. How do you manage the gradient of trust required to let an AI handle complex tasks while keeping a human in the loop?
Trust in enterprise networking isn’t a binary switch that you just flip one day; it’s a gradual progression that has to be earned through consistent, predictable performance. We’ve structured this by ensuring that “Coworker” mode is always ambient but never autonomous, keeping a human as the final decision-maker for every action the system proposes. As we move toward more autonomous “Operator” modes, we are introducing features like a “recap” that briefs a returning engineer just like a human teammate would, explaining exactly what was handled while they were away. This allows an organization to say, “I trust the AI to handle these specific low-risk tasks, but I’m keeping the keys for these high-impact configurations.” It’s about creating a governance layer that respects the expertise of the staff while offloading the mundane, repetitive tasks that cause burnout.
We’ve seen recent data suggesting that subscription bookings for advanced AI platforms are driving nearly half of all new business for major players in the space. How does this shift toward agentic AI specifically address the critical “skills gap” and the time it takes to onboard new engineers in a complex environment?
The onboarding bottleneck is one of the most expensive problems in IT today, often taking months for a new hire to become truly proficient with a specific network’s quirks. By using AI to detect when a user is lingering on a screen or struggling with a task they knew how to do on a legacy system, the agent can surface the top three most common actions and offer to walk them through the process. We’ve seen this reduce new-engineer onboarding time by as much as half because the product documentation is essentially being fed to them in real-time, contextually. In a world where we’ve seen sequential growth for nine consecutive quarters and the number of million-dollar customers is rising to nearly 200, the ability to scale your human talent at the same rate as your infrastructure is the only way to stay competitive.
Many organizations track Mean Time to Recovery (MTTR), but you’ve highlighted a different metric: Mean Time to Context. Why is this specific measurement so vital for understanding the true value of an AI-driven network?
If an outage lasts sixty minutes, usually forty-five of those minutes are spent just trying to figure out what is happening, where it’s happening, and who is affected. That window is what I call the “Mean Time to Context,” and it is the most frustrating part of a network engineer’s life because it’s spent in a state of high-stress discovery rather than actual repair. Agentic AI is designed to eliminate that discovery phase by performing the pre-investigation the moment an anomaly is detected, cutting the resolution time from hours down to just a few minutes. When you measure the success of an AI implementation, you shouldn’t just look at the final fix; you should look at how quickly the engineer was handed a complete package of evidence and a suggested path forward. That is where the real ROI lives, as it allows your most expensive and talented people to focus on strategy and complex problem-solving rather than hunting through log files.
What is your forecast for the evolution of AI-driven network operations over the next few years?
I expect we will see a rapid departure from the “manager of managers” dashboard philosophy as enterprises realize that they can no longer hire their way out of complexity. By 2028, the standard operating procedure for a Tier-1 incident will likely be entirely handled by an autonomous “Operator” mode, with humans only stepping in to audit the decisions and refine the underlying blueprints. We will see a massive push to codify tribal knowledge into digital skills that the AI can execute, making the network far more predictable and less dependent on the specific expertise of a single individual. Ultimately, the dashboard as we know it will become a secondary reporting tool, replaced by a conversational and proactive agentic layer that manages the “noise” so that the humans can focus on the “signal” of business growth.
