CSPs and vendors collaborate to address the challenges of scaling agentic AI

Scaling agentic AI raises multiple challenges, from containing the cost of using language models to avoiding the creation of new technical debt and learning how to manage agents as employees. Network operators and suppliers got together to discuss these challenges at this year’s DTW Ignite.

A glimpse of how AI agents are already working as humans in network operations was provided by Jamie Gaudette who, as Vice President, Networking and Infrastructure, Microsoft, runs one of the world’s biggest optical networks.

Speaking during a panel at DTW Ignite entitled “AI-native ODA: The path to open digital autonomy”, Gaudette explained how in Microsoft’s network “our digital coworkers have email, they have Teams, they have the org knowledge plane, and they do real work. They show up in the org chart, they report to people. We kind of treat them like human beings in the network, and we onboard them the same way we'd onboard a new grad into the network. And these have had a profound impact on our ability to actually get work done with agents.”

Microsoft started with copilots that supported engineers before building out AI ‘co-workers’. “The copilots were fantastic for bootstrapping the knowledge plane, but they didn't move the bottom line significantly,” said Gaudette. “In other words, I wasn't getting much more output per human, although the humans were less cognitively tired.” (Read more about knowledge planes here.)

The company therefore built guardrails and workflows within which AI agents could operate more independently.

“Once we felt that we were confident in our overall framework, we switched to these digital co-workers that actually do work, pull tickets out of a ticket queue, solve problems, update everyone, chase internal dependencies and stakeholders, chase external dependencies, and do work on behalf of team members,” said Gaudette.

Keeping a lid on costs

A major consideration for many communications service providers (CSPs) as they develop agentic AI is cost, as other speakers on the panel explained.

Jio, which has more than 530 million customers and an annual data consumption rate of about 240 exabytes of data, is “architecting in many of our platforms the use of LLMs … in pursuit of autonomous network or autonomous operations in [the] customer side or business operations or enterprise operations,” said Sudhir Mittal, Executive Vice President & Director of Architecture at Jio Platforms.

He pointed out that “as platforms make greater use of agentic AI, then the cost also goes up.”

Mark Austin, VP of Data Science at AT&T, also referred to the importance of cost management, an issue that TM Forum Members are tackling within Model-as-a-Service (MoDaaS), a Catalyst in which CSPs and vendors are working together to develop a unified gateway for AI model consumption.

“Right now, we're doing about 40 billion tokens a day,” said Austin. “My new favorite project is model-as-a-service.You’re trying to pick which model is suitable for the job,” he explained, adding that it is “the most appropriate model that's cheapest, but still solves the problem.”

Cost is not the only benefit of MoDaaS, said Austin. “You save costs, but [you also] get the management, you get the finops, you get security and governance,” he added.

Both Austin and Mittal also underscored the important role that smaller language models tuned to telco data can play when building accurate and cost-efficient agents.

“The real need is to have SLMs, perhaps, or more tuned agents, as per the telco data… We need to have the trained and small language model acting with precision,” said Mittal.

Architecting AI-native operations

A crucial lesson from the early days of generative AI deployment is the need to adapt architectures so that they can support the secure, accurate use of AI at scale.

Deutsche Telekom is one CSP turning to AI to help weed out legacy software and to simplify its IT landscape, as Laurent Donnay, SVP, Chief Engineering Officer, Deutsche Telekom, explained during the panel discussion.

“[Initially] we realized that AI is just accelerating technical debt or mess, because every AI generates something that was very different than the other one. We had no consistency,” he said. But since then, the company has designed an AI-native architecture using TM Forum’s Open Digital Architecture (ODA).

“Today we were able to not only get to an acceptable price point, depending on the topic we target, but also we're able to have the integrated view… because we have underlying architecture consistency,” said Donnay. “We're really infusing ODA at the various steps of the software delivery lifecycle through… the agents.”

He went on to explain: “We're now clearer on what to use where… what are the various software delivery lifecycle steps to perform and depending on… whether building a new feature on a cloud-native application or whether re-platforming an old Siebel system.”

Deutsche Telekom’s work on AI-native architecture fits within the wider ongoing body of work by TM Forum Members to make ODA AI-native, as Andy Tiller explained during the panel.

“In order to manage the governance and control… you need a common framework for the industry,” said Tiller. “This is particularly true where you try to join these scenarios up into end-to-end autonomous flows” across network domains, or from the business layer to the service layer and down to the network and back.

This is because, as Tiller pointed out, “agents working in different domains need to understand each other and to coordinate their response”.

“You really need not just the robust IT foundations that are provided by ODA. You also need common semantic understanding across the different layers and across the domains, and a framework for unified governance, and this needs to operate in a multi-vendor environment.”

TM Forum Members have already extended ODA to enable the lifecycle management of agents, for example. Now, said Tiller, they are working to ensure ODA can support event-driven architecture; to develop ontologies to provide agents with operational knowledge; and to present ODA as a data product for consumption by agents in developer environments.