From scripts to strategy: how genAI enables agent-driven telecom operations

Commercial context

The growing complexity of modern networks means the case for automation is getting rapidly stronger. Yet most CSPs still struggle to apply AI at scale. Legacy operations rely on static workflows, fragmented tools, and manual decisions. AI pilots remain siloed, disconnected from real-time orchestration or business value. The result: slow resolution, mounting costs, and missed opportunities in fast-moving markets.

CSPs face rising pressure to manage cloud-native, multi-vendor infrastructure while keeping SLAs, customer satisfaction, and sustainability targets on track. That pressure is high, but most current systems can't support the scale or speed needed. Rule-based scripts and static dashboards don't hold up in dynamic situations like outages, weather disruptions, or traffic spikes. The time has come for something smarter — and more coordinated.

The 'GenAI-powered cognitive toolkit for network and service management - Phase III' Catalyst is designed to overcome this issue. It provides a practical, standards-aligned framework to coordinate autonomous agents across domains; link generative AI (genAI) capabilities to business goals; and embed trust and governance into every decision.

The solution

The project team built a genAI-powered operating framework designed to move beyond narrow, disconnected automation. It relies on an AI mesh: a network of autonomous agents, each assigned to a specific operational domain such as fault prediction, service provisioning, dispatch, or customer assurance. These agents communicate through event-driven APIs and use shared data structures to coordinate decisions in real time. Unlike static workflows or domain-limited scripts, the mesh allows agents to reason together toward shared goals, prioritizing actions dynamically based on business outcomes.

Each agent interacts with a temporal knowledge graph (TKG), which serves as a structured, time-aware memory layer. The TKG isn’t just a data lake — it captures relationships between service topologies, historical incidents, customer impact, and policy lineage. Built on a graph-based schema with temporal indexing, it allows agents to detect patterns across time, compare current states to known baselines, and apply reasoning strategies such as root cause correlation or impact forecasting. This gives agents context for every decision, and provides explanations to human operators through traceable decision chains.

Above this sits an intent orchestration layer. Users can express goals in natural language, which are parsed and mapped to technical actions using TM Forum’s Open Digital Architecture (ODA) component model and TMF APIs. The system supports both structured prompts and freeform input, using a natural language interface to translate business intent into policy-compatible commands. This removes the need for deep technical input at every step, allowing commercial and operational users to trigger autonomous actions without touching scripts or command line interface.

To ensure trust and control, the Catalyst includes a policy-governed execution layer. Every agent decision passes through a validation step using the TMF725 Policy Management API. This framework checks each proposed action against operator-defined guardrails — including compliance rules, override permissions, and escalation triggers. Approved actions are then executed via TMF622 and TMF641, ensuring service changes and provisioning tasks run through standard, auditable workflows.

The final layer shifts automation from rule-based triggers to value stream-based prioritization. Agents monitor real-time metrics related to SLA compliance, monetization potential, resource efficiency, and sustainability. As a result, the system can adapt to changing business conditions in real time. For instance, it can prioritize premium service recovery during a regional outage. Additionally, it may reallocate resources to meet sustainability thresholds when the network experiences high demand.

Taken together, these five components create a modular, standards-aligned architecture that makes AI both scalable and operationally usable. The system isn’t just smart — it’s coordinated, explainable, and grounded in business value.

Wider application and value

This Catalyst can act as a blueprint for how CSPs can make AI operational for network automation. The results are already compelling. Agent collaboration and shared context have cut fault detection and triage time by up to 85%. As Irshad Deen, Deputy Chief Innovation Office of Sri Lanka Telecom explained: "Automation of recovery tasks and better resource planning suggest early OPEX savings of more than 20%. These gains also provide new revenue streams by linking automation to B2B2X monetization models, SLA-driven slicing, and dynamic service delivery."

The framework helps CSPs shorten time-to-market as well. By connecting business intent directly to execution, CSPs can launch new offerings faster — without manual rework or downstream errors. At the same time, SLA compliance improves, as the system dynamically adapts to shifting priorities and performance conditions.

One of the Catalyst champions is DNB (Digital Nasional Berhad), a MOCN operating in Malaysia. DNB provides radio access for all six national mobile operators. Its current challenge is to develop an AI mesh service to reduce MTTR and improve support for access seekers.

Today, DNB Service Desk agents use intent-based operations and automation to serve six national carriers. With generative AI integrated into the ServiceOps tool, agents gain faster access to 5G site data and automated resolutions. The AI assistant supports agents through a fully conversational interface and improves problem-solving capabilities. These enhancements significantly reduce resolution times.

The Catalyst follows the Autonomous Networks Maturity Model and TM Forum’s IG1251 reference architecture. It offers a reusable framework for closed-loop, explainable automation. Partners from four diverse markets — du (UAE), PLDT (Philippines), DNB (Malaysia), and BT (UK) — tested and refined the model. These trials confirmed the solution's portability and cross-market relevance.

There are wider benefits to society too. During extreme events like typhoons or heatwaves, this architecture helps networks maintain service — especially for healthcare, education, and emergency response. Smart provisioning also cuts energy use by over 25%, contributing to CSPs’ net-zero commitments. And with a natural language interface, even non-engineers can engage directly with operations, broadening participation and reducing complexity.

What makes this Catalyst stand out is the way it enables AI to operate: transparently, purposefully, and in service of real business and human goals. It moves CSPs away from automation as a bolt-on, and toward a new operating model where intelligence is native, accountable, and always aligned with value.