IDC survey flags networking bottlenecks for agentic AI

IDC has published findings from its 2026 AI in Networking Special Report Survey on enterprise concerns about networking infrastructure for agentic AI. The survey was sponsored by Google Cloud.

The research found that security, automation and staffing limits are among the main obstacles as companies try to move AI projects from pilot stages into production. The main bottleneck is infrastructure rather than the AI models themselves, with networking emerging as a significant cause of project delays and abandonment.

According to the survey, 32.6% of respondents cited security concerns, 26.8% pointed to automation challenges, and 24.7% identified limits on staff time and talent. Those issues become more acute with agentic AI because such systems create more distributed, dynamic interactions across applications, services, APIs, tools and data sources.

In production settings, those interactions can span agent frameworks, model providers, clouds, open-source tools, software-as-a-service APIs and internal applications. This broadens the operational scope and increases the security and governance burden for companies trying to run AI systems at scale.

Network role

IDC argued that networking now does more than connect systems. It described networking as part of the infrastructure control layer, applying policy-based controls, supporting observability, and helping maintain security and governance across AI agents and related services.

Framework-level controls are not enough when agents and services operate across different runtimes, clouds, deployment models and operating domains. In that environment, an infrastructure-level approach gives companies a broader, more consistent way to apply policy across fragmented architectures.

Agentic AI also puts pressure on network design by increasing east-west traffic and creating shifting interactions that require closer policy enforcement, visibility and control near the application workflow. Cloud network services are becoming more strategically important as organisations try to reduce fragmented observability, uneven policy application and unmanaged shadow agent activity.

Platform debate

The survey also highlighted a split among organisations over whether to rely on platform approaches or best-of-breed point products for AI workloads. IDC said the division reflects the complexity of building agentic AI systems, which often involve fast-changing business requirements, open-source components, evolving protocols and new architecture patterns.

Among respondents who preferred platform approaches, 32.9% cited stronger security as the main reason, 27.7% pointed to reduced complexity, and 24.2% chose faster deployment. Point solutions may still be needed for specific technical requirements, but using too many separate tools across distributed AI environments can lead to inconsistent policies, operational complexity and governance gaps.

Rather than treating the choice as ideological, the research said companies should take a strategic view. IDC argued that platforms can provide a common operational and policy base for AI deployments, as long as they remain modular and extensible enough to incorporate external tools when needed.

Open approach

The preferred platform for agentic AI should support integration with third-party and open-source tools and allow security and observability functions to be added without a complete architectural overhaul. That flexibility is necessary because businesses are trying to meet AI goals while managing systems that are distributed, autonomous and still evolving.

The findings come as enterprises continue to test AI applications beyond initial pilots and face the harder task of operating them reliably in live environments. IDC's position is that networking has become central to how organisations establish operational control, apply policy consistently and maintain trust across agentic workflows.

The report concluded that the demands created by agentic AI are unlikely to be met by point solutions alone and that organisations will need networking approaches supported by infrastructure platforms that remain open, flexible and extensible.