Can your enterprise network keep up with its agents?

In only three years, AI has become embedded in workflows across businesses of all types and sizes. An NBER survey of 6,000 executives in the US, UK, Germany, and Australia found that 70% of all companies were actively using at least one AI technology, primarily around content generation and data processing. The challenges to long-established business practices, security, and enterprise infrastructure are enormous but CIOs need to ready for the next wave of AI-driven innovation, which will bring even greater pressures to bear.

In addition, IDC predict that by 2029, over 26% of global IT spending will be on agentic AI, totalling $1.3 trillion. Much of this will be on building out enabling infrastructure including data centers, agent construction, and applications. But a significant portion will go to upgrading corporate networks to cope with the demands of AI workloads.

Agentic AI comes with a set of new challenges as emerging tech moves from responding to human prompts to acting more independently and communicating directly with other systems.

More than bandwidth

The last 30 years has seen networks of all types focused on building out capacity from Tier 1 networks down to LANs. Throughput and speed were crucial as the internet and mass computing created huge demands for data sharing. AI, particularly agentic AI, adds new priorities with latency and predictable, lossless networks key to success in this rapidly evolving landscape. Real-time AI chatbots become unusable when packets are lost or delayed, impacting customer service and, ultimately, corporate reputation.

The nature of the emerging agentic AI environment will amplify these problems as a distributed ecosystem of data sources, LLMs, and fragmented agents requires real-time coordination. “In professional services, the value of AI won’t come from how many agents you deploy, but from how well they’re orchestrated,” says Steve Chase, US vice chair and global head of AI and digital innovation at KPMG.

The weakest link

We’re seeing the AI Velocity Paradox play out here with AI helping businesses generate content including text, images, video, and code at massively greater speeds, but slowing down business delivery at the same time. A study from cloud provider Viavi Solutions highlights the new demands being placed on networks by AI workloads. A 1% packet loss, not significant for most business practices, resulted in a 33% drop in GPU utilization as the network experienced “incast congestion” from too many GPUs sending data to the same network switch.

Agentic systems are particularly sensitive to latency issues as well. For example, in a customer service environment, an agent may need to verify a user’s identity, check inventory, and process a payment or a refund in a single reasoning loop. Even a 100 millisecond delay in data retrieval can cause these loops to time out or hallucinate, breaking the chain of action. As agents become more autonomous and need to communicate with an even more dispersed set of systems, these problems will become amplified. Network traffic is moving from the traditional north-south client server architecture to a more east-west topology that introduces more points of delay and failure. “Applications that rely increasingly on real-time analytics and decision-making, such as industrial automation or video analytics, depend on low-latency data paths, more local data break-out, and more integrated management of distributed compute resources,” says Matt Hatton, founding partner of Transforma Insights.

Network intelligence

Organizing these new demands on networks will require ever greater embedded intelligence to prioritize mission critical AI traffic over less urgent data flows. The potential impact of agentic AI on an enterprise’s cost base will make effective deployments an essential element of competitive advantage. “Networks being AI-ready isn’t just about bandwidth,” says Julian Skeels, CDO at managed network services provider, Expereo. “It’s about deterministic performance, deep telemetry, and API-level programmability. If the network can’t be consumed and controlled by software, agentic systems will hit a ceiling.”

Key requirements for AI-native networks include such observability and telemetry that move beyond traditional dashboards and reporting systems to provide deep insights into the paths taken by agents to find data. Embedded security is vital as agents move data autonomously, and a zero trust architecture needs to be at the core that facilitates agent verification before allowing access to sensitive information. Overlaying this, the network needs to be self-healing via a system of automated remediation that can detect issues and reroute traffic before customer experience is impacted.

In an agentic AI world, the network is the nervous system enabling new ways to do business. It should be treated as a strategic asset in the same way as your company’s intellectual property, data, employees and customers. “CIOs need to treat network intelligence and orchestration as a key strategic requirement,” says Skeels. “The observability, control plane, and automation fabric around them will determine whether your network accelerates or constrains AI over the coming years.”