Verizon aims for autonomous network: Takeaways on architecture, AI agent returns, managing costs

Verizon is automating its network and its closed-loop automation platforms have executed more than 70 million network configurations with AI agents and frontier models such as Anthropic's Claude.

The effort, outlined in a recent blog post, highlights how Verizon is looking to automate physical networks, continually optimize and ultimately reach Level 4 network autonomy, which will approach high-level cognitive automation.

Verizon's primary return on its software-defined network automation is time and labor hours. AI agents have been able to make changes that would have required thousands of labor hours from technicians.

According to Verizon, there's also a customer experience return because rich physical telemetry can resolve coverage and capacity issues before a customer is impacted.

In a recent webinar with analysts, Verizon CTO Yago Tenorio outlined the project and the following takeaways.

Verizon is moving from rules-based automation to autonomous networks that can reason. "How to embed AI, and in essence get the network ready to respond to situations for which it was never programmed, and that's the difference between automation and autonomy," said Tenorio.

The company has rolled out Anthropic's Claude Code to about 33,000 technology employees to increase internal development capacity. "We're demonstrating the ability to shrink an eight or five to eight hour Manhattan network outage down to just mere seconds," said Tenorio.

• How Verizon manages AI agent sprawl (PDF)

Two architectural pillars drive network autonomy: A common data layer and shared skills and agents repository. Tenorio said:

"You need a common data layer of really good input data. That's the first thing. The second thing is a common repository of skills, tools, and agents. Common repository means it is shared across the corporation. It also means that it creates line of sight for you to train the agents and to write the skills that collect and gather incorporate global expertise."

He added that Verizon's approach is to have an open interface that can leverage multiple models and vendors so it can deploy its own agents. "This is a system that operators should do themselves," said Tenorio.

Shift from network KPIs to customer experience data. Tenorio said Verizon is using device-level telemetry to capture real customer experiences that network level data can't see. "We used a lot of data from our network with logs. That is great but everything is from a network perspective," he said. "We are now using input data that reflects actual customer experience."

Use an AI agent training camp. Verizon is using a training environment and visual tools to teach agents how to diagnose and fix complex radio issues. "We write the skills to train them to understand the radio conditions, the situations to spot patterns and to find solutions. It's an iterative process, so we train them until they are proficient, and then maybe they graduate from school to production," said Tenorio.

Keep humans in the loop. Verizon is deliberately constraining AI agent scope and enforcing security review as well as observability, traceability and reversibility to prevent rogue behavior.

Tightly manage AI costs. Tenorio said Verizon is tightly managing AI costs via token allowances, model selection, and ROI tracking, while seeing strong displacement of external vendor spend.

He said:

"We have a framework where depending on your job you have a certain allowance for tokens before you have to request permission to use more.

And you don't need Opus for everything. In many cases, you don't use Opus at all. In some cases, Sonnet is enough. Matching the complexity with the model is the first thing you have to do to control costs.

We are measuring returns on everything now. What we are seeing so far is a significant displacement of external work. The reliance and spend on third parties are going down dramatically."