MWC: Laying the ground for agentic AI

“You might have hundreds of agents today, but you'll have 100,000 tomorrow,” said Andy Markus, AT&T, Chief Data and AI Officer, speaking on a panel discussion about agentic AI at MWC.

As Markus’ figures suggest, it is still early days for most telco agentic AI deployments. But stands across this year's MWC reflected an expectation that AI agents are set to play a considerable role in helping telcos automate their network and business operations.

Junlan Feng, Chief Scientist, China Mobile Group, highlighted just what AI agents can mean for telcos, and their workforces.

When ten years ago China Mobile set out to use AI to revolutionize customer care for its one billion plus customers “it took us three years to cut the number of customer representatives by 50%,” said Feng, speaking during the same panel discussion as AT&T’s Markus. “I think with agentic AI now, if we do the same process, probably that will take … a couple of months of work to do it.”

Deutsche Telekom was one of the operators at MWC showing how it is using agentic AI today. It dedicated a considerable portion of its stand at MWC to demonstrating MINDR (Multi-Agentic Intelligent Network Diagnostics & Remediation), an automated network management tool. Co-developed with Google, MINDR is designed to help the operator perform autonomous diagnostics and operations across complex, multi-domain networks.

The system gives an end-to-end view of service performance, supported by root cause analysis of issues and suggestions for remediation, with the aim of moving from reactive trouble-shooting and towards predictive, service-driven automation. MINDR follows the initial success of RAN Guardian Angel, which Deutsche Telekom also co-developed with Google and launched in November. The telco said it has used RAN Guardian Angel to reduce the time needed to manage network performance at major events where more than 10,000 people gather from hours to around a minute.

The German operator was far from being alone in touting the benefits of RAN automation.

AT&T joined forces with Ericsson and Aira to showcase how rApps deliver what Aira describes as “evidence-based reasoning, contextual awareness, and intelligent prioritization” for AI agents. Ericsson describes rApps as "software applications designed to manage and optimize Radio Access Network (RAN) in a non-real time manner". Tasks they perform include consistency checks, failure detection, performance monitoring, or network optimization. And in the run up to MWC Ericsson announced ‘Agentic rApp as a Service on AWS’.

“Moving from operator intent to trusted, deployable automation is the real test of AI-native RAN. We’re demonstrating a repeatable, standards-based path that turns GenAI-generated rApps into production-ready capabilities,” according to Rob Soni, VP RAN Technology, AT&T, in a statement.

Data for AI at scale

AT&T’s AI deployment extends far beyond the RAN. During his presentation Markus explained that “we're doing this at scale. So almost 2 billion production API calls bring this into the systems and the screens that our employees work in every day. We're now at over 27 billion tokens a day on average.”

The US telco has invested heavily is making sure its data is accurate and well-curated. “It's you bringing your own data … that drives the true magic. The hardest thing … is doing really complex data analytics accurately,” said Markus. AT&T has done much of this hard work in-house.

“We can't do it just using large language models out of the box. We have to fine tune those, and we do that with our own data, over 122 RAG or fine-tuning pipelines, because we all know that for an enterprise …large language models or agentic AI out of the box doesn't bring value.”

Especially at scale. “This is really hard, not for one table, but at the scale of AT&T; hundreds of tables with many columns with complex questions. How do you bring that together and do it accurately?”

Like many telcos, Colin Bannon, CTO, BT Business, underscored the critical importance of agent security. “These agents are not going to just sit in the cloud in one area with one perimeter. They're going to be highly distributed in the network. They're going to be in these devices. They'll be in our pocket, they'll be at the edge, they'll be in some of the devices around us,” said Bannon.

In addition, as industries, from finance to telecoms, increasingly deploy agents in decision-making roles, they are likely to face questions about who holds responsibility for the decisions and actions of their agents.

AT&T’s Markus, for example, referred to the US, where “we’re really, really, really close on getting ... Sarbanes-Oxley (SOX) controls. (The Sarbanes-Oxley Act is a US federal law established in 2002 to address weaknesses in corporate governance and financial reporting.)

Markus pointed out that SOX controls “are extremely rigid. They have to be proven. We're getting them certified for agentic processes, non-deterministic processes,” he explained. “This is kind of cutting edge. How do you think about that in an agentic space? And then what's next?”

Whatever is next, Markus is clear that as the number of agents in the ecosystem grow, their management becomes “really, really important”.