You Don't Have an AI Problem. You Have a Data-Integration Problem.

TM Forum has validated autonomous-network levels, and Telefónica closed 2025 with twelve fully operational Level-4 use cases out of more than four hundred in production, across fixed, mobile, transport, IP, core, and telco-cloud domains. The conversation moved from “will networks self-heal?” to “which domains have you certified?” Agentic AI, agents that diagnose and act on the live network, is the inflection operators are betting on.

But validating Level 4 in named domains revealed the binding constraint, and it is not the model. It is whether the data those agents need, across OSS, BSS, network, and subscriber systems, is unified, governed, and real-time. Most of it is not.

McKinsey finds 57% of telcos scale AI across multiple domains, yet only 16% call it the “new normal.” That gap is fundamentally a data-foundation and operating-model gap. Agents can only diagnose and act on a network if the underlying data is integrated across the silos that decades of OSS/BSS sprawl created. As the chart shows, fragmented data and legacy operating models, not model capability, are what stall telco AI.

This is the consultancy's wedge made literal: the constraint on telco AI is no longer the AI; it is the data substrate underneath it. Whoever unifies that substrate first turns Level-4 ambition into production, and turns the network's own telemetry into a data product, network APIs already reach roughly 80% of mobile connections, with Network-as-a-Service revenue projected to grow at around 42% a year.

A telco's view of who connects to whom, where, and how is its most defensible asset and its most sensitive data, and it is what churn prediction, next-best-offer, and personalised service all read from. NIS2 classes telecom and data-centre infrastructure as high criticality, mandating supply-chain vetting of providers for jurisdictional risk. Feeding the subscriber graph to a frontier API to save on infrastructure erodes the one asset that makes the business more than a pipe, and creates a dependency the operator does not control.

European operators are turning this into product: Deutsche Telekom is building an EU AI factory with sovereign LLMs (Teuken, SOOFI), and Orange and Mistral offer GenAI-as-a-service in a sovereign framework. Sovereignty here is monetizable, not merely defensive. The figures are Rindogatan models; the Level-4 and adoption data are Telefónica's and McKinsey's.

Unify the OSS/BSS/network/subscriber data substrate across domains, that integration, not the model, is what separates the 16% from everyone else. Classify subscriber-data workloads as sovereign-by-default.

Stand up sovereign inference behind the existing platform, keep the frontier to the non-identifying edge, and build network telemetry into governed data products. The moat is the subscriber graph; the strategy is keeping it home while still shipping the AI it enables.