Artificial intelligence is reshaping the digital landscape, and most of the attention has focused on compute capacity, GPUs, and data center scale. But connectivity is just as critical. As AI moves from experimental pilots to production workloads, the networks that carry data between users, applications, and AI agents become the foundation on which success is built.

Agentic AI, in particular, changes the traffic patterns that network architects have relied on for decades. Instead of human users initiating requests during predictable business hours, autonomous software agents continuously interact with APIs, databases, and other AI systems. They trigger actions, retrieve information, and collaborate across distributed multi-cloud environments. That creates a level of network demand that cannot be met by a static, manually configured IP infrastructure.

The End of the Busy Hour

Traditional network planning often used a "busy hour" model. Operators analyzed peak usage times and sized links, routers, and switches accordingly. Traffic was mostly predictable because people followed regular work schedules. Voice calls, video conferences, web browsing, and email generated relatively consistent flows that could be managed with static routing and a modest amount of overprovisioning.

Agentic AI destroys that model. AI agents operate around the clock. They do not take weekends off, nor do they follow the nine-to-five calendar. A customer service agent may be processing requests at 3 a.m. while a supply-chain agent negotiates with supplier systems in another time zone. These agents make decisions in microseconds and require immediate responses from the underlying network. The traffic profile is continuous, spiky, and far less predictable than in the past.

That reality brings a new dynamic for enterprises and service providers planning their next phase of AI deployment. Those who modernize their IP networks can unlock new revenue from AI-driven services. Those who delay risk losing competitive relevance.

Legacy Networks vs. AI Workloads

Legacy IP networks were built for a different era. They delivered voice, video, and general internet traffic with reasonable reliability. They supported VPNs and enterprise connectivity through multiprotocol label switching and other established protocols. But these technologies were designed for relatively static connections and predictable service-level agreements.

AI workloads are different. They need low-latency access to distributed GPUs, high-throughput data pipelines, and the ability to reroute traffic instantly when a particular path becomes congested or fails. AI agents need to communicate with each other across sites and clouds, and they need those communications to be secured down to the packet level. Traditional networks simply cannot provide that level of agility.

One of the biggest challenges is visibility. AI networks require real-time telemetry to help operators understand traffic patterns and support automated intervention. Without real-time information, operators are left trying to support AI workloads through reactive manual troubleshooting. They rely on static reports that are often outdated by the time they are read. In an agentic AI environment, that is not just inefficient; it is dangerous. A network congestion event can cascade into delayed AI decisions, lost revenue, and poor user experiences.

Operators need streaming telemetry that continuously captures metrics such as latency, jitter, packet loss, and utilization at every network point. This data must be fed into automation systems that can act automatically. If an AI agent is attempting to move a large dataset between two data centers and the primary path becomes congested, the network should recognize the issue in milliseconds and route traffic around the problem without human intervention.

Segment Routing and EVPN as a Foundation

Evolving from a bloated, rigid, and complex IP architecture to more modern ones based on segment routing and Ethernet VPN is essential. Segment routing, which can be implemented over MPLS or IPv6, enables the network to steer traffic along explicit paths without relying on complex signaling protocols. It leverages existing network investments while creating an evolutionary path to the flexibility needed for AI workloads.

Segment routing allows operators to encode the path a packet should take into the packet itself. That simplifies network operation and provides precise path control. With segment routing, operators can direct latency-sensitive AI traffic away from congested links while sending less critical traffic along alternative paths. They can also engineer traffic around failures in real time.

Ethernet VPN, or EVPN, is equally important. EVPN provides a modern, standards-based way to deliver Ethernet services over IP and MPLS networks. It overcomes many of the limitations of older Layer 2 VPN technologies, offering better scalability, faster convergence, and support for advanced multihoming. Combined with segment routing, EVPN gives operators the flexibility to create paths with specific performance properties across an automated, scalable infrastructure.

In the past, network architects often had weeks to make changes to support new demands. Today, network conditions must change within seconds to meet the requirements of AI agents. Segment routing and EVPN make that possible because they are designed for programmability and integration with software-defined networking controllers.

FlexAlgo: Matching Traffic to Performance

Another critical capability for agentic AI networks is FlexAlgo, short for flexible algorithm. FlexAlgo lets the network calculate optimal paths for different traffic types based on operator-defined constraints. Unlike a best-effort routing table, FlexAlgo can create multiple logical topologies over a single physical network.

For example, one class of traffic might be optimized for low latency and jitter. Another might be optimized for maximum available bandwidth. A third might be engineered for resiliency, perhaps to meet disaster-recovery requirements. And still another might need to satisfy data sovereignty rules by restricted to paths that stay within certain geographic boundaries. Depending on the needs of each AI workload, the network can automatically select the correct virtual topology.

FlexAlgo delivers the traffic-engineering benefits that operators once sought with RSVP-TE, but without the massive complexity. RSVP-TE relied on manually engineered tunnels and extensive state management. It required significant configuration at every router along a path, making it difficult to scale and operate. FlexAlgo, by contrast, allows operators to define performance objectives and constraints centrally, then lets the network automatically compute and maintain the appropriate paths. As networks increasingly support different SLAs for different AI agents and workloads, FlexAlgo ensures traffic is matched to performance requirements rather than constrained by static, one-size-fits-all rules.

This becomes especially important as enterprises move more deterministic workloads into production. A generative AI application that needs a response time of fewer than 100 milliseconds will suffer if its requests are routed over a congested path optimized for bulk file transfer. FlexAlgo lets that application receive a dedicated low-latency path without requiring a physically separate network.

Security and MACsec

Network modernization for AI is not just about performance. It is also about security. AI agents are an attractive target for attackers, and the data flowing between distributed AI systems can be highly sensitive. Organizations in healthcare and finance, for example, are subject to strict data privacy and compliance requirements. They need to ensure that all traffic is encrypted and authenticated across the entire network.

One key technology is MACsec, or Media Access Control Security. MACsec provides link-layer encryption and integrity, protecting data as it travels across Ethernet links. It is especially valuable in an AI environment where many potential attacks happen at Layer 2. With MACsec, an attacker cannot eavesdrop on traffic between data centers, cloud gateways, and AI processing clusters, nor can they inject malicious packets into the stream.

MACsec is a complement to encryption at higher layers, such as TLS or IPsec. By securing the underlying links, organizations can build a defense-in-depth strategy that protects AI workloads from insider threats as well as external attackers. Combined with real-time telemetry, segment routing, and FlexAlgo, MACsec helps create a network that is both capable and secure.

Real-World Adoption in Critical Sectors

Leaders in healthcare and finance are beginning to incorporate all of these capabilities into their network architecture as part of broader digital transformation initiatives. They need to support a mix of AI and traditional workloads while ensuring that traffic adheres to strict policy, sovereignty, and SLA requirements. The network must automatically enforce business rules, prioritize the right workloads, and protect sensitive data at all times.

For example, a healthcare organization may be deploying AI models that assist in medical imaging analysis. Those models need access to large electronic health records and imaging archives. Some of that data may reside in on-premises storage, some in a private cloud, and some in a public cloud. AI agents must retrieve data from all of those locations, process it in real time, and return results to doctors and clinical systems. The network must be able to segment traffic, encrypt every data movement, and deliver the required latency even as clinicians from multiple facilities access the system simultaneously.

Similarly, a financial institution may use AI agents to detect fraudulent transactions, analyze market risk, and provide personalized customer recommendations. These agents need access to account databases, market data feeds, and machine-learning models. Because financial services are heavily regulated, the network must ensure that customer data is not silently routed to unauthorized locations, even if a particular link fails. Data sovereignty constraints may require that certain traffic never leaves the country. FlexAlgo can enforce those boundaries automatically, while MACsec protects customer account numbers and transaction records.

These organizations have a choice in how they consume such capabilities. Depending on their operational model, they can deploy and manage their own IP networks over leased optical services from telecommunications providers. This gives them direct control over network provisioning, policy, and performance, which can be important for very large enterprises with specialized requirements. Alternatively, they can consume the same capabilities through a fully managed network service. A managed service can reduce the burden on internal IT teams, accelerate time to deployment, and create new opportunities for service providers to deliver differentiated, value-added services.

A Competitive Imperative

The rise of agentic AI creates both an opportunity and a challenge for service providers and large enterprises. If they modernize their IP networks, they can meet the demands of AI agents, offer new services, and monetize the next wave of AI innovation. They can differentiate themselves by delivering guaranteed performance, strict security, and flexible traffic engineering. Their customers will benefit from faster AI applications with the confidence that data is protected and compliant.

If they stand still, however, they risk being run over by competitors who embrace network evolution. AI does not wait for network teams to slowly reconfigure routers and switches. It expects the network to be a seamless, intelligent part of the broader infrastructure. This expectation does not change the laws of physics, but it does require a serious commitment to modern architecture.

Operators need to start now. First, they should assess which parts of the network are most likely to support AI agents and evaluate whether existing paths can handle dynamic, latency-sensitive traffic. Second, they should introduce real-time telemetry and analytics to gain a live understanding of network behavior. Third, they should consider a phased migration to segment routing and EVPN, beginning with the most critical connections. Fourth, they should deploy FlexAlgo to create differentiated paths for AI workloads. Finally, they should layer MACsec or similar encryption technology into the physical and virtual infrastructure.

None of this requires a complete overnight replacement of every router and switch. Segment routing leverages existing investments. And a carefully managed overlay approach can deliver many of the benefits while operators gain experience. The key is to move deliberately and with an eye toward the near future. Agentic AI is already here, and its networking requirements are more demanding than anything enterprises have previously experienced. The organizations that act now will be the ones best positioned to thrive in the AI era.

Source:Network World News