- Rakuten Mobile has already removed humans from the loop for network energy savings at scale
- Data accessibility, not AI models, proved to be the biggest blocker to automation
- OSS systems keeping pace with AI remain major challenges
FUTURENET WORLD 2026, LONDON – For Petrit Nahi, Rakuten Mobile’s journey toward Level 4 network automation has been a long time coming — more than two decades.
Nahi completed a Ph.D. from Queen Mary University of London in the early 2000s, researching distributed multi‑agent systems that could dynamically adjust network coverage based on traffic patterns. “It was basically using a distributed, multi-agent system to control antennas for coverage,” he said, describing research that would not be practically deployable for years. The ideas worked in theory, but networks and network operators weren’t ready.
Now, some 24 years later, data systems have finally caught up – and his research is much more applicable.
“I think the networks needed to evolve data accessibility,” said Nahi, whose title is senior member of the technical staff, CTO Office, for the mobile operator.
Without consistent, centralized data, agentic AI-based automation simply wasn’t possible, he told Fierce at FutureNet World in London. Many traditional telecom environments were (and still are) fragmented across OSS, BSS and domain‑specific teams, each “running its own kingdom,” noted Rakuten Symphony CMO Geoff Hollingworth, who also met with us at the show today.
Rakuten Mobile’s greenfield build offered Nahi a rare chance to design around that problem from day one. When he joined the operator in 2019, he insisted that critical telemetry — traces, probes and performance data — sit with the data science team, he said. “That’s the foundation for you to build AI models,” he told Fierce.
Open RAN plays a key role
Open RAN has helped, but not in the way the industry often thinks of it.
“Open RAN in itself might not be interesting," Hollingworth said. "But, the second-order effect of having open networks in RAN is very interesting, because suddenly you get all of the data that you can work with in a normalized way.”
That data, flowing continuously from roughly 150,000 cells, now feeds both predictive and reactive AI models operating across the live network, according to Nahi. In at least one domain — energy savings — Rakuten Mobile has already achieved L4 autonomy; humans are no longer in the loop for those decisions. TM Forum certified Rakuten Mobile’s efforts in February.
Other autonomous workflows are in progress, but the operator isn’t ready to announce them yet, according to Hollingworth.
Moving faster exposes frailties. “The OSS systems were not necessarily designed originally for this speed and volume,” Nahi said, noting that AI‑driven actions can stress-test every system upstream and downstream.
What Rakuten Mobile has in common with Tesla
Hollingworth likened the challenge to building autonomous vehicles. “We’ve got more in common with how Tesla builds their cars than we probably have with the traditional [telecoms] network,” he said.
In making sure everything works the way it should, the AI “[stress-tests] everything… if there is any weakness, it hits it really hard,” said Nahi. “The benefit of it is that we’ve been able to make the systems more robust.”
Rakuten Mobile’s experience suggests that Level 4 automation is less about AI claims and more about unglamorous foundations: data ownership, system design and organizational change.
While Nahi’s research is decades old, the industry is finally catching up to what it requires to make agentic AI work in the real world.