Enterprise networks are entering a new era where artificial intelligence isn’t just consuming bandwidth—it’s rewriting the fundamental rules of network traffic behavior. According to new research from Cisco Systems, the adoption of agentic AI will push global network traffic growth from a projected 2.5x increase over the next decade to a staggering 9x growth, fundamentally altering how organizations must design, secure, and scale their wide area networks (WANs).

The findings, published in Cisco’s AI Impact on Wide Area Networks 2026 report, reveal that AI inference traffic—generated by autonomous agents performing tasks at machine speed—will account for nearly 25% of all network traffic by 2035. What makes this particularly challenging is that AI traffic behaves nothing like traditional web traffic, with dramatically different patterns in duration, symmetry, and criticality that current network architectures weren’t designed to handle.

For network architects and IT leaders, the implications are profound. “The real risk isn’t that AI traffic will appear overnight,” explains Javier Antich, principal product management engineer at Cisco. “It’s assuming it behaves like everything else when it doesn’t.” The report’s findings suggest that by treating AI traffic as just another HTTP flow, organizations risk creating critical bottlenecks that could undermine the very AI systems they’re deploying.

Key Insight: While AI inference currently represents a small fraction of total network traffic, its growth rate is exceptional—with some service providers observing 4x increases in AI-related traffic over just eight months. This explosive growth is being driven by the rapid integration of AI agents into enterprise applications, with Gartner projecting that 40% of enterprise applications will include task-specific AI agents by 2026, up from less than 5% in 2025.

Why AI Traffic Is Different: The Four Key Behavioral Shifts

The report identifies four fundamental ways AI inference traffic differs from traditional web traffic, each requiring different network treatment:

• Duration: AI inference flows last approximately 2x longer than regular web transactions, with the main driver being token-by-token content generation rather than bulk data transfer.
• Flow Rate: Median flow rates for AI inference are 10x lower than regular web transactions, creating different peak-to-average traffic patterns that may require specialized Quality of Service (QoS) settings.
• Asymmetry: 9% of AI inference flows show more upstream traffic than downstream—compared to just 0.5% of non-AI HTTP transactions—a pattern that will significantly impact radio capacity planning for mobile networks.
• Latency Sensitivity: While traditional web APIs aim for sub-100ms response times, even short LLM queries begin producing output after hundreds of milliseconds, with full responses often taking seconds.

These differences aren’t just academic—they have immediate operational implications. For example, network systems that maintain flow state tables will need to accommodate growing tables as AI inference flows proliferate. The report warns that “security and flow-aware network systems are likely to become more distributed to cope with forwarding state growth” as AI adoption accelerates.

Enterprise Adoption: The Coming AI Agent Revolution

The network traffic changes represent just one dimension of a broader transformation in enterprise software. According to Cisco’s analysis of third-party research:

• IBM’s 2025 Global Executive Survey found that 24% of business leaders already have AI agents making independent operational decisions, with 67% expecting this capability by 2027.
• Gartner projects that by 2035, agentic AI will drive 30% of all enterprise application software revenue—exceeding $450 billion globally—up from just 2% in 2025.
• The compound annual growth rate (CAGR) for AI inference traffic is projected at around 25% during the 2029-2032 period, when adoption will experience its most pronounced increase.

This represents a seismic shift in enterprise computing. “nearly one-third of the enterprise software market might be attributable to AI agent capabilities by 2035—a radical transformation in a single decade,” notes the report. The implications extend beyond technical infrastructure to business models, with new software offerings increasingly centered around autonomous functionality.

Network Architecture Implications: What Changes Are Needed?

The report outlines several critical areas where network architectures will need to evolve:

• Resilience Requirements: AI inference paths are becoming “strategic network assets” that require high levels of resilience, similar to mission-critical applications.
• Observability: Networks will need enhanced visibility to precisely track AI flows, which often span multiple systems and services.
• Differentiated Treatment: Specialized Quality of Service (QoS) settings will be required to manage the unique traffic patterns of AI inference.
• Latency Monitoring: Service providers will need to monitor AI inference latency as a key performance metric, as it directly impacts perceived user experience.

Perhaps most significantly, the report challenges the assumption that AI inference is primarily a compute or GPU problem. “As inference evolves, the networking part is becoming more relevant,” states Guru Shenoy, senior vice president of Cisco provider connectivity. “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.”

Practical Steps for Organizations

For organizations preparing for this AI-driven network transformation, Cisco recommends several immediate actions:

• Traffic Analysis: Begin analyzing current AI traffic patterns using network monitoring tools to understand baseline characteristics.
• Capacity Planning: Model future traffic growth with AI scenarios, particularly focusing on the 25% CAGR projection for inference traffic.
• Architecture Reviews: Evaluate whether current network architectures can support the longer flow durations and different traffic symmetry patterns of AI inference.
• Security Posture: Implement differentiated security treatments for AI inference paths given their growing criticality.
• Vendor Engagement: Work with network equipment providers to ensure solutions can handle the evolving traffic characteristics of AI workloads.

One particularly challenging aspect is the variability in AI inference latency. While traditional web APIs aim for sub-100ms response times, AI systems often require hundreds of milliseconds just to begin producing output, with full responses taking seconds. This creates new requirements for network design that prioritize both low latency and high reliability for inference traffic.

Looking Ahead: The 2035 Network

By 2035, the report projects that:

• AI inference will account for 25% of all network traffic
• Nearly one-third of enterprise software revenue will be driven by AI agent capabilities
• Most enterprise software will have AI agents deeply embedded in their functionality
• New software business models will revolve around autonomous functionality

This represents a fundamental shift from today’s network architectures, which were optimized for human-paced, bursty video streams. The challenge for network architects is to design systems that can handle both traditional traffic patterns and the emerging AI-driven workloads—without creating performance bottlenecks that could undermine the very AI systems organizations are deploying.

The good news is that this transformation is already underway. Cisco’s research combines real-world traffic analysis from its Crosswork Assurance User Experience service with controlled lab tests of AI agents, providing enterprises with concrete data to begin planning their network evolution.

The next major checkpoint for this transformation will be the release of updated network infrastructure standards in late 2026, which are expected to incorporate these new AI traffic patterns. In the meantime, organizations would be wise to begin benchmarking their current AI traffic characteristics and stress-testing their networks with simulated AI workloads.

As Cisco’s researchers note, “If you are planning capacity, designing architectures, or defining strategy for the next decade, this conversation isn’t optional—it’s foundational.” The network of the future isn’t just about moving more data faster; it’s about creating a digital nervous system capable of supporting the autonomous decision-making that will define enterprise computing in the 2030s.