Cisco: 36 months to modernise networks before AI overwhelms capacity

More evidence of the immense pressure that growing artificial intelligence (AI) workloads are placing on networks has been revealed in research from Cisco, which has fundamentally confirmed that large language models (LLMs) and the emerging wave of agentic AI are placing unprecedented strain on enterprise campus and branch networks, while security surfaces are already expanding beyond what defences can manage.

The networking giant surveyed 3,472 IT leaders in Asia-Pacific, Europe, the Middle East, Latin America and North America between March and April 2026 about how AI is impacting their campus and branch networks. The sample comprised CIOs, as well as networking, end user computing and technology leaders, at organisations with more than 500 employees and an average of 3,292 campus/branch locations.

The topline finding, and indeed call to action, for businesses was that network resilience, observability and adaptive security are essential in the AI era.

The study recognised that the network has survived decades of transformation, from dot com to the cloud, by adapting and evolving to meet the moment. Yet it stressed that going forward, those organisations that treat network modernisation as a prerequisite to their AI strategy, rather than a parallel workstream, will define the next decade of enterprise AI.

On a quantitative basis, the research data predicted that three years from now, AI will triple network traffic – representing a 235% increase – and said AI workloads are changing traffic patterns across enterprise environments in ways many existing workplace networks were never designed to support.

The survey said growth was attributable to the fact that, unlike human users, AI agents operate at machine speed, triggering dozens of application programming interface (API) calls, database lookups and model inferences in seconds. This generates dense east-west traffic – lateral device-to-device or server-to-server communication required for AI agents to exchange data – that legacy workplace networks were not designed to handle.

For example, 67% of the participating respondents reported increases in east-west traffic tied to these workloads. Additionally, 61% noted growth in continuous automated traffic generated by AI systems. Most enterprises believe that the likely net result of this is that they will hit campus and branch network capacity limits in two years.

These changes are expected to become even more significant as organisations move beyond generative AI (GenAI) experimentation and deeper into agentic AI capable of autonomous action. A third of firms surveyed said they already have broad enterprise-wide agentic AI deployments, and 97% overall expect an expansion in agentic AI use within 24 months.

In addition, the study observed that the same agentic AI workloads that have the potential to transform enterprises are also uniquely fragile. Mature AI adopters globally – those that are ahead in AI deployments – reported that AI workloads are acutely vulnerable to networking issues, making them more sensitive to reliability and uptime (80%), bandwidth (75%), latency (71%) and packet loss (62%) than traditional applications.

In all, less than a third of mature AI adopters say their networks are fully prepared for projected AI growth. Overall, 76% of respondents admitted they need upgrades, and 73% said that they have hit, or will hit, campus and branch capacity limits within 24 months. Crucially, almost ubiquitous Wi-Fi is emerging as a major bottleneck for AI, with more than half listing it as the area driving the greatest increase in capacity requirements.

Worryingly, the study also revealed a disconnect between ambition and reality, with three-quarters of IT leaders agreeing that they are more confident in their organisation’s AI strategy than in the network’s ability to deliver it. Yet even though 91% cited budget constraints as a barrier, almost all enterprises were planning to modernise their workplace networks.

The explosion in AI workloads and general usage was also causing increased security headaches. The overwhelming majority of firms conceded that they were struggling to keep up with an increasingly challenging security environment (92%) and that AI has already caused some damage (90%).

Over two-thirds also believed AI-related threats are evolving faster than their ability to adapt, and that failing to adapt networks over the next two years will only increase security risks. At the same time, an observability gap is widening as traditional monitoring tools struggle with bursty, east-west agentic flows.