The rise of agentic AI is fundamentally transforming enterprise network traffic, creating a scale of complexity that legacy routing strategies weren’t designed to handle.

In a recent guest column for SiliconANGLE, Lee Peterson, Vice President, Branch and WAN Network at Cisco, drew a striking parallel between the evolution of enterprise networking and the current state of artificial intelligence. Peterson observes, “Just as SD-WAN uses real-time telemetry, policy, and observability to steer traffic, model routing can direct each request to the best endpoint for its performance, cost, privacy, and accuracy requirements.”

This insight signals a pivotal maturation in enterprise AI. We are no longer managing simple, predictable requests; we are orchestrating autonomous workflows where every decision directly impacts performance, cost, and reliability. While the industry is largely focused on model selection, the real competitive advantage lies in assurance—the ability to synthesize network, cloud, and security context to ensure every AI interaction meets specific business outcomes.

For years, organizations have been constrained by a one-size-fits-all approach, often overpaying for massive frontier models to process routine tasks that smaller, specialized models can execute faster and at a fraction of the cost.

At Cisco, we’re shifting from passive connectivity to proactive, intelligent model routing that treats AI models as flexible endpoints within a wider fabric. By leveraging our unique cross-domain visibility, we can guarantee not just where a query is routed, but also the quality of the result it delivers. This is our blueprint for the new era of enterprise intelligence, where operational rigor meets the agility of the AI age.

From static connections to dynamic intelligence

Dynamic model routing (DMR) is rapidly evolving from a niche optimization technique into a cornerstone of enterprise IT architecture. DMR is an intelligent policy and decision layer that assesses incoming AI requests and directs each query to the optimal execution path—such as a domain-specific small language model (SLM), an on-premises model, a deterministic application programming interface (API), or a frontier large language model (LLM)—based on complexity, cost, privacy, and latency parameters.

In the early cloud era, network engineers recognized that backhauling branch traffic to a centralized data center was inefficient and costly. Today, enterprise AI teams face an identical challenge with enterprise tokenomics: routing every routine query to a monolithic, energy-intensive LLM cannot only create latency bottlenecks but also burn through token budgets with diminishing returns.

By deploying an intelligent model routing approach, organizations can evaluate incoming requests in real time against specific criteria:

• Context and complexity: Differentiating between simple telemetry checks and complex, multi-variable root-cause analysis.
• Cost and latency budgets: Leveraging cross-domain network intelligence to guarantee the end-to-end performance and reliability of every AI request.
• Data privacy and governance: Enforcing security policies so sensitive corporate data remains on private, on-premises, or air-gapped models rather than public APIs.

Much like the transition from rigid, expensive multi-protocol label switching (MPLS) circuits to agile, software-defined wide-area networks (SD-WANs), DMR empowers organizations to eliminate single-vendor lock-in and optimize operational spend.

The power of assured outcomes

Beyond simple model steering, Cisco introduces a critical differentiator: the assurance layer. While many routing solutions focus solely on selecting the most cost-effective endpoint, our approach synthesizes real-time telemetry from across the network, cloud, and security stacks.

This cross-domain visibility allows us to validate the end-to-end path before a request is ever processed, ensuring performance and reliability. Instead of merely finding the cheapest path, we prioritize the assured path—the one that meets the strict latency and security requirements of enterprise-grade operations.

Cross-domain visibility allows us to validate the end-to-end path before a request is ever processed, ensuring performance and reliability.

--

By integrating this intelligence into the model router, we transform AI deployments into predictable, high-performing systems. Ultimately, this enables IT leaders to move beyond managing tokens and toward managing the quality of verified AI outcomes.

Improving enterprise data privacy and governance

A dynamic routing layer serves as an architectural policy enforcement point (PEP). It inspects query metadata and data classifications to ensure that sensitive, proprietary, or regulated payloads stay confined to secure on-premises models while allowing non-sensitive workloads to be processed using public cloud APIs.

Beyond data security, DMR enhances operational governance by enabling IT leaders to move from reactive spending to proactive management. By analyzing average token requirements for specific tasks, organizations can establish automated guardrails. Real-time evaluation-powered checkpoints can stop “runaway agents” before they exceed budget limits, so actual resource consumption remains consistent with original financial forecasts.

Can DMR run in hybrid and on-premises environments?

Yes. Through unified management platforms and edge-enabled SLMs (such as those developed with NeuralFabric), routing layers can execute localized intelligence directly on hardware or within private clouds, ensuring strict regulatory compliance and low latency.

What role does latency play in routing decisions?

Latency is a critical factor in routing decisions, particularly for time-critical operational workflows such as real-time voice over internet protocol (VoIP) triage or interactive incident response. In these scenarios, DMR prioritizes compact, local models to eliminate network round-trip delays, switching to deeper cloud reasoning only when additional analytical depth is required.

The role of latency has evolved as AI transforms enterprise traffic patterns from predictable paths between users, data centers, and applications to far more complex interactions that often initiate dynamic workflows spanning branch locations, multiple clouds, software-as-a-service (SaaS) applications, and AI models.

Observability platforms help manage this complexity by providing IT teams with a consolidated view across branches, clouds, and SaaS environments and allowing for continuous validation of routing decisions, ensuring that the network adapts to changing conditions rather than relying on static, “set-and-forget” configurations.

Reducing token costs and consumption through an ensemble architecture

Real-world enterprise operations demonstrate that model routing is most effective when integrated into a broader ensemble strategy. In production environments, such as Cisco AgenticOps for autonomous network operations, not every step requires an LLM call:

• Deterministic execution first: Foundational operational tasks—continuous monitoring, telemetry ingestion, baselining, and compliance checks—run on deterministic compute, APIs, and runbooks without consuming model tokens.
• Domain-specific optimization: For specialized domains, purpose-built engines like the Cisco Deep Network Model and SLMs can match or exceed frontier-model accuracy on bounded tasks while reducing operational costs by up to 30x.
• Flexible hybrid deployment: With advancements from technologies like NeuralFabric, SLMs can run locally on edge network hardware or within on-premises and air-gapped environments, eliminating cloud egress costs and data-movement latency.
• Conditional escalation: When a workflow encounters conflicting hypotheses or high-risk decision points, the routing harness dynamically escalates the query to deeper reasoning models.
• Sustainable enterprise tokenomics: The true measure of AI efficiency isn’t raw token volume. It’s the cost per verified outcome. By leveraging deep observability, we prioritize routing decisions that deliver reliable, high-value business results rather than simply selecting the lowest-cost endpoint. This approach allows organizations to maximize analytical rigor while keeping infrastructure costs both predictable and performant.

The new blueprint for enterprise intelligence

The transition from static, monolithic AI deployments to a dynamic model routing architecture is more than just a cost-saving measure. It represents a fundamental shift in how enterprises operationalize intelligence. By treating AI models as flexible endpoints rather than rigid destinations, organizations can finally balance the competing demands of performance, privacy, and price.

Learn about our approach to optimizing model performance in this overview: Cisco Deep Network Model: Purpose built intelligence for networking.

Modernize and secure your network for the AI era

Explore AI-ready secure network architecture