Verizon is rapidly ramping its use of AI to power internal operations, a process that CTO Yago Tenorio said is also quickly gaining steam toward helping the carrier start to embed autonomous control of its radio access network (RAN).

Tenorio touted Verizon’s efforts in a recent blog post, which lays out the carrier’s reasoning and path toward “moving away from traditional hardware constraints to turn our infrastructure into a software-defined, self-organizing reasoning engine that optimizes itself in real time intentionally.”

That post explains that Verizon is “pushing hard toward [level-four] autonomy in critical segments of our core network,” an effort that builds on past network disaggregation and virtualization work and is taking advantage of the latest generative AI models.

Tenorio in an interview with SDxCentral explained that Verizon’s steps toward this control began with its initial push on virtualization network assets a decade ago. This included virtualizing core and network assets, with the former running on the carrier’s own private cloud and the latter leading to the deployment of virtualized (vRAN) open RAN components.

Tenorio noted that this was essential as “one aspect of open RAN is open interfaces, and open interfaces is something that is an essential component for you to throw automation on top, particularly if you're going to do it yourself.”

This work fed into those automation capabilities that allowed Verizon to close 70 million “loops” last year.

“That's without any human intervention,” Tenorio explained. “So the human may be supervising and for sure the human has written the script, but that's the first step into autonomy. If it wasn't for the infrastructure, our own private cloud, and the skills, and the first steps that we took on automation, I don't think we would be speaking about autonomy, because I don't think there is a way that you can skip that phase.”

Tenorio explained that autonomy path accelerated this year with Verizon “rolling out Claude Code to literally everyone in technology and that made everyone a software developer.” This has since expanded to greater use of platforms like Anthropic and Google’s Gemini.

Taking on autonomy

As with its virtualization and automation efforts, Verizon also took its own lead on developing an autonomy platform. Tenorio noted that one path could have been to buy that platform from a vendor and then stay within that vendor environment, “but that’s not what we’re doing.”

“What we're doing is to train an AI model ourselves on our own architecture and our own platform, so if you're going to do that, the ability to interface to your own systems and your own network through open and standard interface become essential,” Tenorio explained, added that this is also helped by the disaggregated architecture that allows for different models to be inserted in the process, including Claude code, Anthropic, and Gemini.

“It's important that your platform is modular, so we can actually flag any LLM [large-language model],” Tenorio said. “We think, today, Gemini and Anthropic are particularly useful for this, but watching this space and the speed at which it’s changing, maybe like the computer model, in six-month’s time there will be something else, so we need to be the platform that we own, and we control, and then we can plug a different language model without changing everything else.”

Tenorio also explained that Verizon’s “private cloud” approach toward training these models is an important distinction, noting it’s essential for Verizon to have the right data to feed into its AI models.

“I think that's far more important, in my view, than the brutal, like, how intelligent the algorithm is going to be,” Tenorio said. “Frontier models are really good at it. It's not about how clever they are, it's about the training that you put in and it's about the input data even more than the training that you deliver.”

Training AI to work

That training is accelerating, with Tenorio looking toward having more specialized AI agents tackling more complex network automation tasks. This includes the ability to come up with root-cause analysis and explanations as to why a specific part of the network is under-performing and then providing a solution.

“We're using that to iterate and train the AI to get better at our job using that,” Tenorio said. The next step is to graduate the agents from university and have them specialized in different tools that we operate in a system that we currently have in prototype, so I want to put that in production in the next three months.”

Tenorio downplayed the “layer” classification of such a move, noting, “the important point is it'll be like a complex network of agents and sub-agents that are monitoring the network 24/7 spotting patterns and anomalies, and then calling a network of sub-agents specialized on different tasks to come up with a root-cause analysis and a solution autonomously.”

This early work has already seen “real-life tasks” that would have taken hours to resolve “done in 90 seconds.”

“I want that in production before the end of the year,” Tenorio said of that work. “We'll probably start with the radio, but very quickly will take on transport and core, so hopefully within the next six months my agents will graduate from university, and they will find a job in operations.”