David Trossell features in this article from The Stack about the latest trends in AI networking.
February 11, 2025
Cisco says the use of AI and machine learning in enterprise networking is “increasingly evident,” and yet vendors often don’t provide much detail about what they mean by network management. Does it involve managing the flow of encrypted data or is it just about helping network managers to find issues that may be causing latency and package loss to enable fix them manually?
Ashley Poundall, Director, Networking Specialists at Cisco, says that AI networking is increasingly discussed because it represents a significant shift in how networks are managed and optimised. Cisco says that “AI/ML improves troubleshooting, quickens issue resolution, and provides remediation guidance. It brings about critical insights to improve user and application experience. AL/ML can be used to respond to problems in real-time, as well as predict problems before they occur. It also augments security insights by improving threat response and mitigation.”
Poundall adds that AI networking enables networks to “learn from data, predict potential issues, and offer solutions before problems escalate.” He admits there is often still a need for manual intervention.
However, he confirms that the goal is to move towards more automated and self-optimising networks. In his view this shift is crucial because of the need to handle the growing complexity of and demands of modern enterprise networks.
Opportunity for network automation
To a certain extent that complexity, with increasingly voluminous amounts of data being handled daily and the increasing number of cyber-security threats, are already present. Arguably this should mean that it’s time for AI to be used to automatically fix network performance issues automatically without human intervention, which could make network teams more productive and strategic in their use of their own time and resources. The appropriate use of AI and ML could even extend the life of legacy networks, negating the need to buy the latest technology on a whim.
David Trossell, CEO and CTO of Bridgeworks explains that some large vendors use AI to “manage their machinery, taking a natural language to consider how to implement it.” He says most devices use a specific AI language to create low-level commands but adds: “It seems that everyone has jumped onto the AI bandwagon. I’ve not seen anyone doing anything automatically. I don’t think it takes control of the network switches, it’s more advisory, such as how engineers can change the network settings to improve performance, whereas we allow AI to manage the flow of data.”
“It’s down to organisations to find out how they are going to manage the network, but it won’t be self-configuring. Engineers will work on the feedback from the AI, and then they will need to find a solution. They have bolted things on rather than designed what they do from scratch. It’s 50 years of the network working, so how do you ditch it to become graphically orientated, so it’s just a matter of click, click, click and it’s done.”
Trossell, meanwhile, believes there is still a need for WAN Acceleration. While some vendors have WAN optimisation built into their products, he argues they have a limited ceiling of performance – especially WAN optimisation, which he says limits itself early on because the workload becomes too high once organisations get into the bigger data transfers and high bandwidths. However, for 5-10Gb data transfers – or greater – WAN Acceleration comes into its own.
He also suggests that AI networking is a part of SD-WANs, but what he sees is AI managing the network switches and controllers. As for other vendors, he thinks they are often using AI as an excuse to upgrade their products, while suggesting that there is no “intrinsic AI in their AI networking products.” For it to be true AI, he argues that vendors need to embed it and machine learning into their products.
AI networking – growing demand
Praful Bhaidasna, Director, Product Management for Arista Networks, nevertheless says AI networking is growing in demand. This is because of the complexity of hybrid cloud, IoT and 5G network environments, the need for real-time insights, the prerequisite of stronger security, and the requirement for greater operational efficiency “to stay ahead of evolving challenges.” There is a need to use predictive analytics to detect potential failures and security threats before they impact the network too.
Bhaidasna also suggests AI networking is about anomaly detection and issue inference to identify deviations from normal behaviour and to pinpoint underlying issues; automate troubleshooting assistance to suggest fixes based on past data and patterns; enable traffic optimisation to analyse traffic patterns and to recommend optimisations; to monitor network security to identify new threats and to automate threat investigations; and it provides network configuration assistance to recommend best practices for provisioning and policy management.
Future vision: Full automation
At present though it’s down to a human-being to make the final decisions whenever it comes to ‘AI-augmented networking’, and Bhaidasna states that fully autonomous networking is part of the future vision of AI networking. With AI, networks would dynamically adapt and optimise networks without any involvement from any engineer. As to why most of the industry hasn’t reached that stage yet, they say – for example – that IT teams are hesitant to let AI make changes without human oversight, fearing that it might lead to unintended disruptions.
The other factors that the company says that may be preventing full automation include:
- Data Quality & Bias – AI models need high-quality, diverse datasets to make accurate decisions.
- Vendor-Specific Challenges – Different vendors have proprietary systems, making end-to-end AI-driven automation difficult.
- Regulatory & Compliance Constraints – Automated network changes must align with industry standards and security policies.
Scot Schultz, senior director, HPC and technical computing at NVIDIA admits that network management teams are constantly looking for ways to spend less time “keeping the lights on.” They want to spend more time innovating. “This has driven the evolution from manual command line configurations to automated solutions for rolling out updates to hundreds or thousands of switches,” he says before adding that AI solutions are predominantly focused on making troubleshooting and predictive maintenance available at scale. “Solutions must deliver this AI for networking functionality through an API-native, automation-driven approach,” he adds.
Poundall concludes that improvements in network performance can be achieved by deploying SD-WANs, traffic shaping, and load balancing to optimise data flow and reduce latency. While Trossell argues that SD-WANs need a WAN Acceleration overlay to boost performance, using AI, ML, and data parallelisation to fully automate data flows, to mitigate the negative impact of latency and packet loss and eliminate any human involvement in the process.
AI networking – more to do
For now, most AI networking doesn’t go far enough. There is both a need and a demand for more automation with artificial intelligence and machine learning. As Schulz explains, there are still more opportunities to make “practitioners’ lives easier using AI.” Over time manual intervention will decrease while automation will increase, minimising or eliminating human errors that can cause downtime. This won’t just be about predictive maintenance, better anomaly detection, regulatory compliance, and fewer outages, as it will be about fully automating ever larger amounts of encrypted data over WANs in ways that most ‘AI networking’ just doesn’t do today.
