Reactive troubleshooting was always a race against time – and increasingly, it is a race operators are losing. At MWC26, a session moderated by Abe Nejad of The Network Media Group (NMG) brought together leaders from Colt, TELUS, Ribbon, and Rakuten Symphony to examine how AIOps is changing the equation. The conversation moved across the full spectrum: from the organizational and data challenges that slow adoption, to the machine learning architectures that power genuine predictive intelligence, to what autonomous network operations will actually demand in the years ahead.
Speakers:
Watch the full interview.
The panel was aligned from the outset: the hardest part of AIOps adoption is not picking the right model or platform; it is the human and organizational layer around it. Governance frameworks, clear value propositions, measurable outcomes, and bottom-up ownership within business units are the prerequisites that determine whether AI initiatives stick or stall. The leaders stressed that a centralized "AI team" responsible for delivering AI to the rest of the organization is a structural dead end. What works is a hub-and-spoke model: centrally governed competency with execution teams embedded inside each domain (RAN, core, OSS/BSS) who own the outcomes and live with the results.
The shift from reactive to predictive network management is well understood in theory. In practice, it requires a fundamentally different approach to data. The panel drew a sharp distinction between traditional MLOps, which relies on structured, deterministic data like stats and alarms, and AIOps, which layers unstructured inputs on top: logs in plain language, RAG-enriched knowledge bases, historical incident databases, and captured tribal knowledge. Getting the right data at the right time, streamed in real time rather than pooled in data lakes that no one actually uses, is what separates operational AI from expensive shelfware.
One of the more provocative contributions to this thread came from Rakuten Symphony's experience building its own data collection layer from scratch. Rather than inheriting vendor-defined KPI formulas and alarm severities, calibrated in a lab, not a live network, the team went back to raw counters and let AI surface its own leading and lagging indicators. The result was a set of AI-defined KPI formulas that revealed correlations and failure signatures that human-defined metrics had never captured. The lesson: sometimes the most important data improvement is questioning what you are measuring in the first place.
A recurring theme across the panel was the danger of building AI models that are technically accurate but operationally incomplete. The leaders pointed to silent failures, such as network faults that customers detect before operators do, precisely because all the monitored KPIs look healthy, as evidence that network data alone is insufficient. Customer experience signals, NPS scores, subscriber-facing data from marketing systems: these are not soft metrics sitting outside the operational domain. They are inputs that, when injected into ML models, reveal the hidden correlations between network events and service impact that have been invisible for years.
The panel converged on a clear view of what the next phase demands. Predictive intelligence alone is not enough; it has to be connected to action orchestration. The difference between automation and autonomy is that an autonomous network learns from its own feedback. Realizing that requires universal API standards for action orchestration, causal ontology graphs that map the impact radius of any network event across services, and a fundamental rethink of how guard rails are defined when AI agents (not humans) own the decision logic. Checklists built for human operators do not transfer cleanly into an agentic world, while guard rails designed for AI reasoning are a different engineering problem entirely.
“In Rakuten, we took a step back and instead of relying on the EMSs and the ready-made MI files and KPI definitions, we just analyzed the counters. This gave us an AI-defined KPI formula instead of a human-defined KPI. This approach really helped us to get the insights that the human-defined KPIs could never give us.”
