AI is no longer just another workload riding on enterprise networks. It is changing the assumptions behind traffic engineering, capacity planning, and network security. That was the message from Bob Everson, chief architect of provider mobility at Cisco, who testified before the U.S. Senate Subcommittee on Telecommunications and Media on July 30. The hearing, titled “Intelligent Networks: Powering Artificial Intelligence and Transforming Communications,” examined how rapid AI adoption is forcing networks to grow more complex, more responsive, and more distributed.

Everson centered his testimony on two questions: how AI is reshaping networks, and how networks can use AI to become more intelligent. Senator Deb Fischer of Nebraska, chairwoman of the subcommittee, set the stage by noting that widespread AI use now requires more capacity and more complex designs. She also acknowledged the scale of investment already under way, with private companies spending hundreds of billions of dollars on network deployment and multiple federal broadband programs helping to fund targeted buildout and maintenance.

AI traffic behavior is shifting

According to Everson, AI is not just producing more traffic; it is changing network behavior. Cisco measured a fourfold increase in AI inference traffic over eight months, he said. Traditional internet traffic was designed primarily for downstream content delivery, but AI workloads are far more interactive and uplink-heavy. Prompts, contextual data, sensor feeds, and agent activity are constantly moving back toward AI models. Connections also remain active longer than ordinary web transactions, making session persistence and reliability more important.

The rise of AI agents makes this effect even stronger. In Cisco’s testing, an agent completed a task while generating 450 percent more traffic than a person doing the same job, and roughly 70 percent of that extra traffic was attributed to inference. That pattern has immediate consequences for local networks. Cisco customers in campus and branch environments have reported a 34 percent increase in AI-related traffic over the past 12 months and expect traffic to jump another 96 percent in the coming year.

Capacity pressure reaches the edge

Capacity concerns are already visible close to end users. Half of enterprise customers say AI demand is concentrated on their Wi-Fi networks, Everson said. In addition, 73 percent of organizations say they face or expect to face campus and branch capacity limits within 24 months. Many are seeing larger volumes of east-west traffic, latency-sensitive applications, and continuous automated AI workloads. AI, Everson explained, is no longer confined to centralized data centers or foundation models in hyperscale clouds. Enterprises are deploying small language models, open-source models, and specialized vision and voice models at the edge, and those distributed systems change where and how traffic enters the network.

That shift is driving some of the most important architecture changes in years. AI is pushing compute toward the edge of the network, which means service providers have to design for AI-native traffic profiles rather than treating AI as another form of cloud traffic. Everson identified technical, cost, and sovereignty considerations behind this change.

• Technical demands: Physical AI examples such as robotics, autonomous vehicles, and industrial automation need decision-making in under a millisecond. If an autonomous robot depends on a round trip to a central cloud, the delay can be too dangerous for real-time operation. That makes local processing and ultra-low latency networking essential.
• Cost efficiency: AI can produce massive quantities of data. High-definition video analytics used for public safety, for example, can generate terabytes every day. Sending all of that information to a centralized cloud for processing creates expensive backhaul bills and massive congestion.
• Data sovereignty and security: Governments and enterprises are increasingly reluctant to send sensitive information over the public internet into a third-party cloud. Regulated industries need controlled paths for data, and in many cases local processing is the only way to satisfy compliance requirements.

How AI can improve network operations

Although AI workloads create complexity, Everson said they also create an opportunity to make networks safer and more efficient. Cisco is applying AI to network operations in ways that let operators automate repetitive tasks and detect problems before they affect users. Agentic AI, as he described it, is able to operate at machine speed and provide greater performance, efficiency, and security. Rather than simply making the network faster, AI helps the network become self-healing. Cisco’s AI-native tools can reroute traffic, adjust capacity, or reconfigure network nodes when they detect performance degradation or impending hardware failure.

This self-healing capability is especially valuable for mission-critical services, because it reduces the need for manual intervention during brownouts or hardware anomalies. Network engineers can focus on architecture and innovation rather than routinely responding to alerts. Cybersecurity teams can shift their time away from low-level ticket resolution and toward strategic threat hunting and detection engineering. At the same time, AI lowers the barrier to entry for less experienced staff, allowing them to ramp up faster.

From network pipe to intelligent fabric

Everson argued that modern networks are moving away from being simple pipes that carry bytes from one point to another. AI-native platforms are becoming the fabric of intelligent connectivity. As service providers move compute to the edge, including cell sites, applications can run directly from the network infrastructure. That trend enables new capabilities such as Integrated Sensing and Communication, or ISAC.

ISAC combines wireless communications with radio-frequency sensing. By using radio waves that reflect off objects, a network can detect an object’s position and path. This works better than optical sensors in smoke, darkness, or spaces with obstructions, making it attractive for autonomous systems, robotics, smart facilities, and public safety. Everson noted that the technology has been prototyped and demonstrated already.

Policy recommendations for an AI-ready America

In his prepared remarks, Everson offered the committee a roadmap for supporting AI-driven network evolution. His first recommendation was to accelerate what he called the U.S. AI-native stack. Cisco is working with companies including NVIDIA, MITRE, and T-Mobile on AI-WIN, an effort designed to combine AI, computing, and wireless into a secure path from 5G-Advanced to AI-native 6G. Everson encouraged Congress to invest in areas where the United States already has strategic leadership, such as computing, core networking, and applications.

His second recommendation was to modernize permitting and infrastructure rules. As computing becomes more distributed, the process for deploying new network facilities must keep pace. Everson also called on the committee to consider the Universal Service Fund and evolving costs of AI-ready networks, ensuring that rural and urban communities both benefit.

His third recommendation was to preserve a balanced spectrum policy. Congress recently made 800 MHz of licensed spectrum available, which Everson said is essential for high-capacity and high-uplink connectivity. The FCC’s decision in 2020 to authorize the full 6 GHz band for unlicensed Wi-Fi is equally important for enterprise demand. He thanked the committee for working to rebuild a dependable pipeline of both licensed and unlicensed spectrum, calling it foundational to American leadership.

Source: Network World News