Cisco executive Bob Everson told a U.S. Senate panel that artificial intelligence is profoundly changing enterprise and service provider networks, forcing operators to rethink capacity, architecture, security and the very way traffic moves across the internet. Everson, chief architect of provider mobility at Cisco, appeared before the Senate Subcommittee on Telecommunications and Media on July 30 to explain how AI is reshaping network demand and how networks themselves can use AI to become more reliable, efficient and secure.

The hearing, titled Intelligent Networks: Powering Artificial Intelligence and Transforming Communications, was planned to explore how rapid AI adoption has changed network infrastructure. In her opening statement, subcommittee chair Sen. Deb Fischer (R-Neb.) said widespread AI use has forced networks to evolve, requiring more capacity and more complex designs so that AI can run efficiently. Fischer noted that private companies have invested hundreds of billions of dollars in network deployment in recent years and pointed to federal broadband programs that have provided billions for network expansion and maintenance. Other witnesses at the hearing represented organizations including U.S. Telecom, Vanderbilt University and the Nebraska Public Service Commission.

AI is changing the volume and behavior of network traffic

Everson said Cisco is seeing a shift that goes far beyond simple traffic growth. AI workloads generate different traffic patterns than typical web browsing, video streaming or file downloads, and that difference matters for carriers, enterprises and regulators alike.

“AI is changing not only the volume of network traffic, but the behavior,” Everson told senators. He cited Cisco data showing a fourfold increase in AI inference traffic over eight months. Inference, the process by which a trained AI model uses new data to produce answers or decisions, is increasingly distributed across everything from data centers to factories, hospitals and public safety systems.

Traditional networks were often optimized for content flowing downstream to users. Videos, web pages and software updates all travel from the core of the network out toward the consumer. AI, by contrast, is far more two-way and uplink-intensive. Users and devices constantly send prompts, context, sensor readings and agent activity back toward AI models. Those sessions also stay active longer than conventional web transactions, making sustained bidirectional capacity more important than ever.

Everson also highlighted the rise of AI agents as a source of traffic complexity. Agents, unlike humans, can operate at software speed and continuously communicate with AI services. In Cisco testing, an agent generated 450 percent more traffic than a person performing the same task, and roughly 70 percent of that additional traffic was inference-related. That means agentic applications could place sudden and persistent demands on networks, especially when many agents are running at once across an organization.

Campus and branch networks are feeling the pressure first

For many businesses, the first evidence of AI-driven strain is showing up in campus and branch networks. These are the networks that support office buildings, research facilities, retail locations, branch bank operations, hospitals and even the Senate office buildings where policymakers work.

Everson said customers have already reported a 34 percent increase in traffic tied to AI workloads over the past 12 months on campus and branch networks. Those same customers expect to see a 96 percent increase in the coming year. The jump is not limited to wired connections either. Half of enterprise customers say AI demand is concentrated on Wi-Fi networks, and 73 percent of organizations already face or expect to face campus and branch capacity limitations within the next 24 months.

Several technical patterns are driving those limits. Organizations report increases in east-west traffic, which moves between servers and devices within the same network rather than into a centralized data center. They also report more latency-sensitive traffic, meaning delays that were once acceptable for web browsing can break real-time AI features. And they are seeing more continuous, automated AI traffic, often generated by software agents, monitoring tools and always-on analytics applications.

Everson pointed to distributed AI models as another reason network design must change. Much of the public conversation about AI has centered on huge foundation models running in central clouds. But enterprises are increasingly deploying small language models, open-source models and specialized models for voice, vision and other tasks. Those models can be distributed across the network, closer to users, devices and sensors, which creates a more complex topology and requires more intelligence in the network itself.

The network edge becomes central to AI performance

In his prepared remarks, Everson described several areas where AI is driving the shift toward edge computing. The edge is no longer just a place to cache content; it is becoming a place where AI models run, where data is processed locally and where real-time decisions happen.

• Technical requirements: Physical AI applications such as robotics, autonomous vehicles and industrial automation may need sub-millisecond decision-making. If an autonomous robot sends data to a central cloud and must wait for a response, the round-trip latency could be too high for safe real-time operation.
• Cost: AI operations generate enormous quantities of data. High-definition video analytics used for public safety can produce terabytes of data every day. Moving all of that data to a central cloud is expensive and can create severe network congestion. Processing it at the edge is often more practical and economical.
• Data sovereignty and security: Enterprises and governments are increasingly concerned about data sovereignty and security. Many customers do not want to send sensitive information across the public internet to a third-party cloud provider, either because of regulatory requirements or because of company security policies. Edge compute can keep that data within a controlled environment.

Everson said network operators are moving compute toward the network edge, including cell sites and other remote facilities, and will increasingly run applications directly from the network. That shift will enable a new generation of services that combine connectivity with awareness of the physical world.

He called particular attention to Integrated Sensing and Communication, sometimes called ISAC. This approach combines wireless communications with radio-frequency sensing so the network can detect an object’s position and path using reflected radio waves. Unlike optical sensors, RF sensing can detect intrusion in low-light conditions, through smoke or around obstructions where traditional video analytics may fail. Everson said the technology has already been prototyped and holds promise for autonomous systems, robotics, AI-driven smart facilities and public safety.

AI can make networks more resilient

Everson emphasized that AI is not only a challenge for network operators; it is also a powerful tool. Networks can use AI to improve performance, detect security threats and manage the growing complexity that AI itself creates.

“While AI workloads present several challenges for network operators seeking to ensure seamless performance, reliability and security, there is a tremendous opportunity to leverage AI to deliver new applications and better performance, infuse security into the fabric of the network, and manage the increased complexity,” Everson said.

He pointed to Cisco’s AI native tools as examples of how network automation can improve uptime. These tools can monitor the network for signs of performance degradation or impending hardware failure and can take action automatically by rerouting traffic, adjusting capacity or reconfiguring network nodes. That self-healing ability is especially important for mission-critical services that cannot tolerate prolonged downtime.

Everson also said AI can help close the networking industry’s talent gap. Managing software-defined and AI-powered networks requires advanced skills, and many organizations are struggling to hire and retain enough engineers. AgenticOps, a term Cisco uses to describe AI-driven operations, can automate repetitive, low-value tasks such as ticket resolution, configuration updates and routine maintenance. These tools can lower the barrier to entry for junior analysts and allow experienced network engineers and cybersecurity specialists to focus on higher-value architectural strategy, threat hunting and innovation.

Policy recommendations for an AI native future

Everson closed his testimony by offering three policy suggestions for lawmakers and regulators to consider as AI continues to transform communications infrastructure.

• Accelerate the U.S. AI native stack: Cisco is investing across multiple dimensions of AI native networking and is bringing new capabilities to 5G-Advanced while laying groundwork for 6G. Everson highlighted AI-WIN, a collaboration involving Cisco, NVIDIA, MITRE, Orion Development Company, Booz Allen and T-Mobile. The effort combines AI, compute and wireless technologies to create a secure, American-led path from 5G-Advanced to AI native 6G. He encouraged Congress to focus on areas where the United States has strategic leadership, including compute, core networking and applications.
• Modernize permitting and infrastructure: As computing becomes more distributed, permitting processes must allow rapid and responsible deployment of network infrastructure. Everson also said that if the committee considers the future of the Universal Service Fund, it should account for the evolving costs of AI-ready networks so that both rural and urban communities can share in the benefits of improved connectivity.
• Maintain balanced spectrum policy: The 800 megahertz of licensed spectrum recently made available by Congress is essential for high-capacity, high-uplink connectivity that AI applications require. At the same time, the FCC’s decision to authorize the full 6 gigahertz band for unlicensed Wi-Fi use is important for meeting enterprise demand inside buildings and campuses. Everson said a dependable pipeline of both licensed and unlicensed spectrum is foundational to American leadership in AI and networking.

Everson’s testimony underscored the emerging consensus among network architects that AI will not simply ride on top of existing infrastructure. Instead, AI is forcing a re-examination of how networks are designed, where compute gets placed, how spectrum is used and how operators ensure reliability when traffic patterns become more interactive and agent-driven. The challenge, he suggested, is not just to make room for AI traffic but to build networks that are intelligent enough to anticipate and adapt to a new era of distributed, real-time and increasingly autonomous communication.

Source: Network World News