Verizon Puts AI Agents Into The Network Automation Guardrail Test

Agentic AI Moves Into Network Operations

Verizon is testing how far agentic AI can extend telecom network automation beyond fixed scripts and closed-loop workflows.

The carrier already uses the on-premise Verizon Cloud Platform to host virtual and cloud-native network functions, including its 5G standalone core and virtualized radio access network.

The same platform also carries GPUs for on-prem network AI workloads.

That architecture gives the agentic AI effort a real operating base rather than a detached software trial.

Verizon says its vRAN footprint is one of the largest in the world, with roughly 60,000 vRAN sites in service.

The carrier had 22,900 vRAN sites in service in early 2025, when 40% of its network was running on the vRAN platform.

The scale changes the question for operators.

The issue is not whether AI can write a recommendation for a single site.

It is whether an agent can help generate, validate and execute actions across a distributed network while staying inside security and compliance boundaries.

Brown said Verizon is at the front of the pack globally on network cloud automation and argued that agents could increase both the speed and depth of automation.

That claim is still tied to customer experience rather than a standalone AI showcase: the operational value has to appear in faster network changes, clearer fault handling or better optimization.

Three Automation Buckets Shape The Use Case

Umashankar Velusamy, Verizon's senior director for technology development and network automation, divided radio access network automation into planned changes, unplanned changes and optimization.

Planned work includes deployments and upgrades.

Unplanned work covers network degradations.

Optimization is where rApps and the RAN intelligent controller continuously watch conditions and make closed-loop adjustments.

Verizon already treats vRAN deployment as highly automated once hardware is installed and connected at the far edge.

The carrier can also upgrade thousands of sites at a time.

That matters here because agentic AI is being evaluated on top of an existing automation estate, not as a replacement for every operational system.

For planned changes, agents can help generate and validate configurations for new deployments or large upgrades.

For service assurance, they can connect signals from different systems and narrow the path toward a fix.

For optimization, Verizon sees room for autonomous monitoring that turns an identified network opportunity into an action and then checks whether the action produced the intended result.

Guardrails Become The Deployment Constraint

The carrier is not treating network agents as unrestricted decision makers.

Velusamy said current implementations are showing good results, but he also emphasized the need for security guardrails.

A telecom network contains domain knowledge, deterministic automations and operational rules that cannot be handed to a model without context, access controls and traceability.

That is why transparency is becoming a product requirement.

If an agent takes a decision, Verizon wants to understand the reasoning behind it and the exact action taken.

Integration is the other requirement: agentic tools have to connect with the carrier's ecosystem while conforming to internal security guidelines and compliance requirements.

Standards May Lag The AI Cycle

Verizon also wants a common framework for agentic AI interoperability, especially around security, trust, handoffs, negotiation and context sharing.

The constraint is timing.

Velusamy said the AI ecosystem is evolving faster than telco standards, so carriers still need enough flexibility to build and adjust independently.

The near-term signal is therefore cautious but material.

Verizon's existing cloud, vRAN and automation base gives agentic AI a realistic path into network operations.

The unresolved test is whether suppliers and operators can make those agents explainable, integrated and secure enough for production network decisions.