The road to Level 4 autonomous networks runs through Level 3: Lessons from building an AI-driven NOC
To embed our video on your website copy and paste the code below:
<iframe src="https://www.youtube.com/embed/HUtnlFHMUc4?modestbranding=1&rel=0" width="970" height="546" frameborder="0" scrolling="auto" allowfullscreen></iframe>
By Gaurav Jain, Vice President - AI and Data Products, Rakuten Symphony
Operators increasingly want Level 4 autonomous networks that can diagnose issues and repair them automatically. We have been building toward this inside a live operator network operations center (NOC). Immediate challenges were related to cost – AI-based agents quickly consume tokens – and whether the AI was actually solving the problem. I discussed both in detail at TelecomTV's AI-Native Telco Forum recently.
Level 4 autonomy relies on Level 3
Before you permit a network to heal itself (i.e., Level 4), you must have evidence that the network will successfully self-heal. This evidence is provided by Level 3.
At Level 3, you run closed-loop automation. You have identified your root cause(s), confirmed the effectiveness and accuracy of remedial activities, and validated it before executing the activity. If successful, Level 4 then largely comes down to executing these trusted processes, with tollgates (described below) catching anything that falls outside them. In contrast, if you bypass these necessary validation steps, then every time your network self-heals, it relies on probability associated with a model that you did not train. The principle is quite simple. Complete Level 3 and Level 4 becomes a manageable task.
A note on terms: in TM Forum’s autonomous network framework, Level 3 is conditional autonomy, where the system senses, analyzes, decides and acts within defined domains, and Level 4 is high autonomy, extending that to more complex, cross-domain conditions with predictive decision-making. Here, “Level 3” also describes an operating discipline: every remediation is validated, and the evidence recorded, before it is trusted to run unattended.
General-purpose AI in network operations
General public models continue to improve substantially; however, in telecommunications, they typically leave a meaningful accuracy gap. In addition to this, we experienced difficulties related to tool sprawl, data gravity and data sovereignty. Due to data sovereignty requirements, related to data privacy and security, our data cannot be sent outside our environment in any format. Anonymizing the data so that it could be processed by publicly available models would have removed any contextual information required for processing. We couldn’t reach operational closure without training the model ourselves.
As such, we developed an ecosystem of curated pre-trained models tailored specifically to our use case, along with our own decision-making framework to determine which models should be used where and when. Each model is trained as needed. All models operate within a single environment, thus minimizing token usage and maintaining data sovereignty.
A multi-layered approach
Keeping everything in one box was about speed as well as sovereignty. Each additional platform increases latency and pulls us from the immediate problem space. All components sit atop a single platform. Basic analytics occur before AI-based processing occurs. Governance happens before everything else, since telcos follow strict regulatory requirements.
- Sovereignty: The base layer is air-gapped from the internet. All operational data remains local.
- Data and memory: Any issue or resolution entered into the knowledge base includes details regarding the work performed by engineers during each issue-resolution cycle and their reasoning.
- Orchestration: All forms of machine learning (e.g., classical ML, Small Language Models, Large Language Models) remain viable options; we utilize Directed Acyclic Graphs (DAGs) with tollgates (validation checkpoints, described below) to direct processing flow and ensure proper sequence of processing. We developed this framework from scratch due to our reliance on having confidence in our decision making process.
- Agent plane: Various levels of work require various types of agents. After orchestration determines the correct path, agents perform the appropriate tasks.