{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreiddvcikl5so4anw2wo2jng7ur2hc2hny5tntfd55kozoaujwyzuc4",
    "uri": "at://did:plc:qzjwstutqk2cy7df7jbzd2hx/app.bsky.feed.post/3mmilb3hyjrw2"
  },
  "coverImage": {
    "$type": "blob",
    "ref": {
      "$link": "bafkreicl6qjqvau3lq4jhoae7offhfeqi7tld4fbfbzapu7sg5dllatqda"
    },
    "mimeType": "image/jpeg",
    "size": 9458988
  },
  "path": "/article/4175890/cisco-ai-traffic-is-radically-reshaping-wans.html",
  "publishedAt": "2026-05-22T14:43:17.000Z",
  "site": "https://www.networkworld.com",
  "tags": [
    "Artificial Intelligence, Industry, Markets, Networking, Telecommunications Industry, WAN, WAN Optimization",
    "AI Impact on Wide Area Networks 2026",
    "Crosswork Assurance User Experience",
    "blog post"
  ],
  "textContent": "AI agents generate up to 450% more network traffic than humans can, and they’re beginning to reshape network traffic patterns in measurable ways.\n\nEnterprise network traffic without agentic AI is expected to grow 2.5x in the next decade — agentic AI adoption will push that growth to 9x compared to current traffic levels, driven by autonomous task execution and inference-heavy workflows, according to new research from Cisco.\n\nCisco’s new study, AI Impact on Wide Area Networks 2026, finds AI and agentic AI will not only increase traffic volume but also, “they will change traffic shape, symmetry, duration, and criticality,” the study reports. “AI inference paths will become strategic network assets, requiring high levels of resilience, observability, and differentiated treatment, for example, Quality of Service (QoS) and path security.”\n\nThe report combines real-world traffic analysis (using Cisco’s Crosswork Assurance User Experience service), third-party industry data, and Cisco-controlled lab tests of AI agents. The research examined a number of parameters, including direct measurement of live AI inference traffic across service provider networks as well as tests of AI traffic characteristics, to train models that identify and precisely track AI flows across the network.\n\n“For service providers, network architects, and digital infrastructure leaders, the real risk is not that AI traffic will appear overnight. The real risk is assuming it behaves like everything else when it doesn’t,” wrote Javier Antich, principal product management engineer, and Guru Shenoy, senior vice president of Cisco provider connectivity, in a blog post about the study.\n\nAgents operate at machine speed instead of human speed, and that changes everything, the authors stated.\n\n“If AI models are the ‘brains’ of this new era, then networks are the nervous system, and when autonomous agents begin to act, decide, and transact on behalf of humans at scale and machine speed, that nervous system of connectivity must be ready,” Antich and Shenoy wrote. “If you are planning capacity, designing architectures, or defining strategy for the next decade, this conversation isn’t optional—it’s foundational. While AI inference is perceived as mostly a compute or GPU problem, the insights in the report indicate that as inference evolves, the networking part is becoming more relevant.”\n\nWhile the industry has spent decades optimizing the network for human-paced, bursty video streams, the rise of agentic AI is changing network traffic profiles and behavior, the study reports.\n\n“By 2035, one-quarter of network traffic is projected to be AI inference (Source: Cisco model). These flows don’t behave like the web. They live longer, demand more upstream capacity, and operate at software speed, not human speed,” the report states. “The connectivity between agent logic and AI models effectively becomes the agent’s ‘spinal cord’—a critical dependency whereby any network degradation directly impairs agent functionality.\n\n“While AI inference traffic remains negligible compared to dominant categories like video streaming, observed growth rates are exceptional,” the reports states. “Token-consumption data shows nearly 10x year-over-year growth, while in some service provider measurements we are seeing around 4x growth in just eight months. Sustained growth at these rates means AI traffic will become a meaningful component of overall network traffic by 2035,” the report states.\n\nCisco’s report cites research from other sources that shows how quickly the agentic AI world will be upon enterprise networks. For example, the Cisco study notes that Gartner is forecasting that 40% of enterprise applications will include integrated, task-specific AI agents by 2026, up from less than 5% in 2025. Gartner also projects that by 2035, agentic AI will drive approximately 30% of all enterprise application software revenue, exceeding $450 billion globally (up from just 2% of software revenue in 2025).\n\nCisco also cited IBM’s 2025 global executive survey, in which 24% of business leaders said they already have AI agents taking independent action in their operations—and 67% expect to have AI agents autonomously making decisions in workflows by 2027.\n\n“In essence, nearly one-third of the enterprise software market might be attributable to AI agent capabilities by 2035 – a radical shift in a decade. At that stage, most enterprise software is expected to have AI agents deeply embedded, and new software business models will revolve around autonomous functionality,” Cisco stated.\n\nOther key findings include:\n\n  * **Traffic growth** : “Total network traffic will experience significant growth over the period 2029-2032, when adoption of agentic AI will experience a more pronounced increase, with a compound annual growth rate (CAGR) of around 25% in AI inference traffic,” Cisco reports.\n  * **Data flow length:** “AI inference flows present different characteristics compared to non-AI web traffic. While not dramatic, these differences may impact capacity planning. Analysis of data flow length shows differences between AI inference flows versus regular web transactions. Statistically, AI inference flows last 2x longer than regular web transactions,” Cisco reports. “The main driver is the way AI inference traffic generates content, one token at a time, resulting in longer and lower rate flows compared to other flows.”\n  * **Implications of longer flow rates** : “For ‘flow-aware’ network systems that must keep state for flows in tables, the proliferation of AI inference flows that last longer means growing flow tables will need to be effectively planned,” Cisco reports. “Over time, security and flow-aware network systems are likely to become more distributed to cope with forwarding state growth.”\n  * **Data flow rate** : “Related directly to the flow length, the data flow rate for AI inference flows shows a different pattern compared to regular web transactions. Median flow rate is 10x larger for regular web transactions compared to AI Inference flows,” Cisco reports. “The main reason for this difference is the process that generates the data in AI inference. Regular web traffic flows peak much higher, as the content can be retrieved from wherever it is stored and delivered to the user.”\n  * **Implications of flow rate differences** : “Different median average-traffic and peak-to-average rates may require different QoS settings in the network to manage AI inference traffic versus non-AI inference web transactions,” Cisco reports.\n  * **Traffic asymmetry** : “Network traffic asymmetry varies depending on access type (mobile vs. wireline networks) and services. Mobile network traffic tends to be more upstream heavy due to social networking, with more content sent upstream compared to wireline networks,”Cisco reports. “Analysis of AI inference flows versus non-AI web transactions shows clear differences in traffic symmetry. In AI inference traffic, 9% of flows have more upstream traffic than downstream traffic compared to other HTTP transactions, where this occurs in only about 0.5% of the flows.\n  * **Implications of traffic asymmetry** : As adoption of agentic AI grows, more significant changes to network traffic symmetry patterns can be expected. “Particularly for radio capacity planning, traffic symmetry assumptions are a relevant factor. As adoption of AI inference and agentic AI grows, it will be important to track the evolution of traffic symmetry, as it will continue decreasing over time,” Cisco reports.\n  * **Network latency impact** : “Large Language Model (LLM) inference requests tend to have significantly higher latency than typical web API calls, and the response times can be more variable. Traditional web application REST APIs often strive for subsecond or even sub-100-millisecond (ms) response times,” Cisco reports. “By contrast, even short LLM queries incur response times of hundreds of milliseconds just to begin producing an output, and full responses often take seconds.”\n  * **Implications of AI inference latency** : “AI inference latency varies widely from a few hundred milliseconds to multiple seconds,” Cisco reports. “Network latency will become a key factor for inference distribution, combined with others like scale, data sovereignty, and security. Additionally Service Providers will need to monitor the actual AI inference latency that the customer is experiencing, as it’s a key factor in perceived user experience.”\n\n",
  "title": "Cisco: AI traffic is radically reshaping WANs"
}