Why Agentic AI Alone Won't Deliver Autonomous Networks
Artificial Intelligence has become the defining technology of modern telecommunications. Nearly every operator is investing in AI-powered operations. Vendors are introducing generative AI assistants, autonomous troubleshooting tools, predictive analytics platforms, and intelligent copilots designed to transform Network Operations Centers (NOCs). Conference agendas are dominated by discussions of Agentic AI, Large Language Models (LLMs), Digital Twins, and autonomous networking.
On paper, the industry appears closer than ever to self-operating networks. Yet, inside most operators, the reality looks remarkably familiar. Network engineers still investigate alarms manually. Root cause analysis still requires information from multiple OSS systems.
Configuration changes continue to move through lengthy approval processes before remediation begins. Major incidents still depend on experienced engineers piecing together information from dozens of disconnected operational tools.
Despite rapid advances in AI, network operations remain fundamentally manual.
So why the disconnect between promise and reality?
AI can't operate what it can't understand
If today's AI models can write software, generate complex analysis, and reason across massive datasets, why can't they operate a telecommunications network autonomously?
The answer is surprisingly simple: the biggest limitation isn't artificial intelligence, it's the operational environment AI is being asked to work within.
Most AI initiatives in telecommunications have focused on adding intelligence on top of existing operational systems. This assumes that enough AI layered across alarms, tickets, telemetry, topology databases, and inventory systems will naturally deliver autonomy.
However, every OSS maintains its own view of the network, with different data models, update cycles, and operational scope. No single system represents the complete state of the network. When AI queries these systems, it frequently receives conflicting answers. One platform reports a service as healthy, another reports packet loss, while a third contains outdated topology. Without knowing which version is correct, autonomous action becomes risky.
The reason isn’t the AI model. The real bottleneck is the data architecture beneath it.
The AI overlay problem
Many organizations believe they have an AI challenge; in reality, they have a data architecture challenge.
Today's AI deployments often function as intelligent overlays sitting above fragmented operational environments. Modern service providers typically operate multiple inventory systems, monitoring platforms, EMS, NMS, ticketing systems and configuration databases. Each provides a slightly different version of reality.
While humans can sometimes compensate by correlating information manually, AI cannot. AI can only reason about the information it receives. If operational data is fragmented, inconsistent, incomplete or disconnected across OSS systems, even the most advanced
AI can’t safely make autonomous decisions.
The models themselves may be highly capable, but the data they consume is inconsistent, incomplete, or stale. Engineers remain responsible for validating recommendations because the underlying operational picture cannot be fully trusted.
Why Agentic AI needs more than intelligence
Agentic AI does more than recommend actions – it observes, reasons, decides, executes, and learns. Each capability depends on a consistent view of network reality. But before taking action, an AI agent must understand the current state of the network, what changed, why it changed, which customers are affected, what dependencies exist, and whether previous actions achieved the intended outcome. These questions can’t be answered reliably if every operational system presents a different version of reality.
Five common data problems prevent the move toward autonomous operations:
- Multiple sources of truth - Different operational systems disagree about the current state of the network.
- Temporal data skew - Operational systems observe events at different times, causing incorrect event ordering and poor root-cause analysis
- Spatial context loss - Traditional telemetry identifies devices but not the customers, services or SLAs affected by a fault.
- Missing causal relationships ¾ Knowledge connecting alarms, root causes, services and remediation often exist only in experienced engineers' heads rather than in machine-readable form.
- No verification loop - Many automation systems execute actions but never verify whether the customer experience actually improved.
- Before networks can become autonomous, operators need a trusted operational foundation that solves for these five data problems and continuously reconciles network state across every domain, technology, and workflow. Only then can AI move from recommending actions to executing them confidently.
The Missing Foundation
For autonomous operations to be trustworthy, the challenge isn’t simply adding AI or automation, it’s creating a data foundation where every decision can be understood, traced, and validated. That requires a structured data foundation consisting of four pillars that establish a progression from raw telemetry to closed-loop operational intelligence:
- Spatial and temporal alignment - to establish shared context
- A reconciled Digital Twin - to establish trusted network reality
- A causal knowledge graph - to establish understanding
- Action-outcome observability - to establish learning
Together these capabilities transform telecom data from isolated telemetry streams into a structured intelligence layer where autonomous systems can perceive, reason, act, and continuously improve with evidence. Trust in autonomous operations ultimately doesn’t come from the AI model itself. Rather, it comes from the integrity of the data foundation beneath it.
Looking Ahead
The telecom industry has spent years improving AI models but far less time improving the operational data those models consume. Data consistency—not model capability—is becoming the limiting factor for autonomous networking.
The journey to TM Forum AN-4 autonomy doesn’t require replacing existing OSS systems. Operators can progressively improve visibility, reconcile network state, introduce causal reasoning and automate within policy boundaries. Ultimately, autonomous networks won’t be built simply through larger AI models. They will be built by providing AI with trusted, contextual, continuously reconciled operational data.
When every observation is aligned, every relationship understood, every action verified and every system shares the same operational truth, AI finally has the foundation required to operate networks autonomously.
In the next article in this series, we'll explore the concept of the data contract: the operational framework that allows AI to trust the information it receives and make decisions with confidence. Before AI can become autonomous, it first needs a single version of the truth.
Read our whitepaper to learn more.
