- Observability and simulation are major gaps in network AI readiness
- Agentic AI's machine-speed traffic patterns are fundamentally different from human traffic
- Physical AI requires deterministic latency, instant failover and edge intelligence that leaves no room for network error
Operator networks need a lot of work to be ready for upcoming demand.
Leaders at Bell Canada, MetTel, Zayo, Rakuten Symphony and Blue Planet pointed to three major gaps in telco network readiness for the future of AI: observability, simulation and the foundational shift from human- to AI-operated operations.
The deeper requirement: AI needs to go beyond a bolt-on network addition and instead become fundamental to the network.
"We need to stop treating AI as a feature and start treating it more as an operational model," said Ahmed Abdelaziz, VP of automation and transformation at Rakuten Symphony. Abdelaziz and other leaders spoke on a panel at the recent AI and the Automated Network Virtual Summit, hosted by Fierce Network.
The readiness gaps
Network engineers are already rethinking their roles as AI traffic grows more complex and observability tooling is struggling to keep pace. "Observability is where we're lacking — where we're not ready yet," MetTel CTO Ed Fox said. "We don't know how to give it good quality-of-experience for the end users."
To improve resilience, simulation needs to be beefed up as well, Abdelaziz said. "Networks are heartbeats — you cannot just stop a heartbeat and expect the human to survive," he said. However, network digital twins, which could fill the simulation requirement, remain an emerging capability as vendors, including Nvidia, slowly push digital twin tooling into the market.
Then there's physical AI. Though not yet at scale, when it emerges it will generate different network requirements than generative AI and agentic AI, Abdelaziz said. Generative AI consumes data and compute; agentic consumes workflows and APIs; and physical AI requires deterministic network performance.
The deepest gap is cultural and organizational, not technical, said Kailem Anderson, VP of global products and delivery for Blue Planet, a division of Ciena. Blue Planet, along with Viavi Solutions, sponsored the summit.
"As an industry, we're not fully prepared for a world where AI isn't just assisting operations but actively participating in operations — and this isn't a technology issue, it's a people issue," Anderson said. "Most networks are still designed around human decision-making." The shift to humans supervising AI rather than approving each step requires new operational processes and new tooling "and we haven't been good at that."
Agentic AI is already straining networks
Operators are already feeling readiness gaps. MetTel has had to upgrade edge routers in the past two months to handle the new traffic profile created by enterprise agent deployments. "It's a lot of micro transmissions. The fingerprint of traffic has a different impact than we're used to," Fox said.
AI inference, unlike streaming video, generates personalized responses that cannot be cached, said Joda Schaumberg, Zayo SVP of digital infrastructure. With video, everyone is consuming the same content, "but now with AI and inference, that consumption effectively becomes personalized content. The caching, the storage model really doesn't work, and that puts a stress on the distribution networks," Schaumberg said.
Zayo has begun exploring architecture changes to address the problem, which he predicted will accelerate without industry action.
The architecture demands of physical AI
Physical AI, including robotics, autonomous manufacturing and real-time industrial systems, raises the stakes. Delayed responses in physical AI contexts are potentially "catastrophic. You're potentially putting people's lives in danger if you don't have that real-time feedback loop," Anderson said.
Networks must be "designed for peak performance and not averages, because you need an instant response within a distributed network," said Célia Lamarche, Bell Canada VP of applied AI and operations support. Bell projects its total network traffic will grow sevenfold by 2035, driven primarily by the shift from consumer video to enterprise AI workloads — a transition that telcos must clear significant structural barriers to monetize.
"The conversation moves from how do I build a fast network to how do I operate an intelligent network that continuously adapts to the AI workloads," Anderson said.
Success looks like operating across silos
Success over the next two to three years hinges less on raw capacity than it does on coordination. "The operators who can automate, simplify and gain end-to-end visibility across that AI lifecycle will be the ones who extract the most value from their AI assets," Anderson said.
That echoes findings from Telstra's network automation program, which has shown that automation is something telcos must earn through operational discipline — not something deployed in a single step.
For Zayo, success means having enough fiber in enough places proactively so AI factory deployments don't stall on connectivity constraints, Schaumberg said. For Rakuten Symphony, the benchmark is commercial.
"Nobody buys AI," Abdelaziz said. "They're buying better uptime, lower opex, a better customer experience."
For deeper analysis on the infrastructure, automation and operational changes operators need to make for AI, download the Fierce Network research report "The telco guide to AI-driven networks" And watch the full AI and the Automated Network Virtual Summit on demand.