- Agentic AI may help operators automate fault management, service assurance and RAN optimization, but live networks will require tighter controls than POCs
- NGMN says AI agent identity becomes critical in multi-agent systems, where operators need auditability, least-privilege access and traceable operational history
- Interoperability risk may show up less in basic roaming and more in service assurance
AI will likely become a key enabler in creating a credible pathway towards trusted Level 4 autonomous mobile networks, but achieving this will require significant transformation across architecture, standards, security, operations and governance, a new report from the Next Generation Mobile Networks Alliance (NGMN) argued.
The report investigated the motivation, opportunities, requirements and ecosystem developments associated with applying agentic AI for network automation in autonomous mobile networks. It offered three main takeaways: a reference framework for agentic AI-driven network automation; an assessment of standards and open-source gaps; and guidance for safer, interoperable deployment.
According to NGMN, commercial adoption of agentic AI will depend less on individual AI models and more on reliable support systems, ecosystem alignment and avoiding fragmentation.
What will operators need to adopt agentic AI?
Operators will need a myriad of guardrails, including governance, trust, policy controls, observability, explainability, operational safety, determinism, assurance, security and cost controls. The report also expects work on agentic AI to move beyond isolated proofs of concept toward controlled, high-value deployments where agents manage complex workflows across network domains.
Ultimately, broader telco collaboration will be needed to reduce fragmentation and align on architectures, interfaces and responsibilities, according to the Alliance’s report.
Fierce Network asked NGMN how the industry can reduce fragmentation and address interoperability. Johann Reindl, project lead and senior program manager, and Guenter Klas, project co-lead, both from the NGMN Alliance, responded by email.
Fierce Network: Your paper notes: "Every AI agent must have a provable identity, a tightly scoped role and act within the principle of least privilege. The identity should remain stable across the agent lifecycle so that verified operational history and reputation information can be consistently associated with the correct agent." Is this table stakes before any of the other use cases (agentic fault mgmt., service assurance, radio access network (RAN) optimization) can be implemented? Are there any real-world examples of an operator implementing agentic identities on a live network?
Johann Reindl and Guenter Klas: We would not necessarily consider this table stakes before any high-value use case can be implemented as a proof of concept, particularly where a single AI agent is involved. However, these aspects become table stakes in multi-agent systems for live networks. Principles of least privilege and narrowly scoped roles are carried over from today’s human-centric and cloud-oriented operations into the world of collaborating AI agents. Without agent identities in a multi-agent system, operators risk uncertified, unapproved, untraceable agents operating outside a governance model. Agent identities also enable trust among collaborating agents and are important for auditability and forensic analysis upon incidents caused by AI agents.
NGMN does not have visibility into the extent operators already implement agent identities, but the topic is getting explored, for example, within TM Forum, where operators and service providers consider some form of agent identity as a must-have security feature. For example, see this Self-Optimizing Autonomous Networks: An implementation guide from Google, TM Forum and Vodafone.
Fierce Network: Please explain the importance of graph neural networks (GNN) for telecom operators. How do the tokenomics of GNNs differ from LLMs? Have any operators used these in trials or production, and if so which GNNs have they used?
Reindl and Klas: While there are different types of GNNs, the key point is that they are complementary to LLMs and knowledge graphs rather than substitutes for them. Due to the different roles of GNNs and LLMs, natural language tokens and thus tokenomics do not apply to the GNNs we have come across. Here is why:
Regarding role and importance of such GNNs: Assume a network topology with associated states of network components describes a Radio Access Network (RAN) and is captured in a knowledge graph. This provides a source of ground truth for an LLM. A GNN can then be used to frequently evaluate such a graph and predict new graph node property values and new links (relationships) between graph nodes. Technically this is enabled by a GNN's main job to compress a multi-dimensional neighborhood shape (e.g. of cell towers in a RAN) and all its properties into a single row of numbers called an embedding vector.
Now, remember: Today, a chatbot with LLM can answer a user’s query by performing retrieval augmented generation (RAG) on a body of curated documents that are represented in a vector database as a set of embedding vectors. The user’s prompt is turned into an embedding vector, and a similarity search is performed to find the chunk of text in the database that likely answers the user’s query. This helps to ground the LLM and avoids hallucination.
Similarly, an AI agent’s LLM can perform GraphRAG to fetch a portion of a knowledge graph that matters for the user’s query. As a result, the AI agent has now access not only to the ground truth placed into the knowledge graph (e.g. a network topology) but in addition it has access to what this type of GNN has predictively added to the knowledge graph as a forecast (e.g. router A -> will cause failure with a 90% chance -> base station B). Again, this helps the LLM to reason more carefully, not just guess what might happen.
Regarding tokenomics: For the GNNs we have seen, the billing is rooted in traditional cloud compute metrics, not token consumption. For LLMs, a token is a discrete linguistic unit (roughly 4 characters): an LLM is ingesting tokens and creating new ones. The GNNs we refer to do not process text, and they do not generate character sequences. Instead, they process matrices, create embedding vectors and predict graph node property values and relationships between graph nodes. Traditional ‘tokens’ do not apply here. Finally, NGMN has not yet surveyed operators on the use of GNNs for agentic AI-driven network automation.
Fierce Network: Your paper highlights a lack of "organizational readiness" and "cost management support frameworks" within operator organizations. Are there specific steps you recommend, and/or have you seen any operators implement anything that prepares them to better leverage AI in the network?
Reindl and Klas: Many operators are already preparing their organizations for the adoption of GenAI, for the use of copilots that assist employees (from mundane tasks to specification-driven software engineering) and for the use of modern AI in products, platforms and services. Significant progress is already being made in workforce skills development and AI governance, including privacy and security. However, designing, leveraging, lifecycle-managing, supervising and governing autonomous multi-agent systems clearly brings new challenges, including a change to various job roles e.g. in operations. Steps for systematically educating the relevant workforce and strategically upskilling people should start now.
Regarding cost management, operators are already recognizing that widespread, organization-wide usage of GenAI tools and AI models can lead to unexpected budget overruns, prompting initiatives to improve cost control around token consumption. A cost management framework that helps control billable LLM token consumption becomes even more relevant when that consumption is generated by autonomous AI agents operating a live network.
While in the former, case management can simply tell a software engineer, “You have consumed all your LLM credits for this month, so please stop consuming for now,” a similar message to an AI agent in charge of service assurance could lead to impairment of the agent and disruption to live services. Thus, in the latter, case the challenge is to find an optimum balance that enables an AI agent to perform its role without blowing a budget, through various smart technical means.
Fierce Network: Global telecommunications depend on network interoperability/seamless handovers between networks. Are there any issues operators need to consider here as different companies implement agentic AI at different speeds and with different vendor partners?
Reindl and Klas: This is an important question, as networks are connected, and users roam into foreign networks. However, this is also an area that warrants further analysis.
Within a given operator’s network, interoperability and well-defined open APIs remain critical to enabling plug and play multi-vendor networks in the context of agentic AI. Where two different networks exchange information e.g. in case of international roaming or for phone calls between two users on different mobile networks in the same country, the data exchange is happening on the network level using interfaces already well standardized by 3GPP. Use of agentic AI for network automation purposes is not expected to remove those 3GPP interfaces.
Thus, differences in autonomy maturity levels between operators are not expected to break services like roaming. However, it could lead to different issue resolution experiences for a scenario involving two networks. In case of a mobile-to-mobile phone call scenario, end-to-end service assurance could be limited by the slowest network's manual repair in the event of a network fault.
While network issues in the network that performs fault management at machine speed are expected to get resolved very quickly, the same network issue occurring in the other network might take longer to resolve, if the use case operates at a lower maturity of autonomy and involves traditional workflows and operations engineers. Another scenario might be cross-operator network slices for enterprises, where asymmetric maturity might impact the consistent assurance of an end-to-end service level agreement. Overall, this question deserves some more study to identify and confirm potential issues.
Read more about agentic AI on Fierce Network
Lightyear launches agentic AI platform for enterprise telecom procurement
ROI is not enough: Telcos seek unique value from AI
Calix bets agentic AI can offset broadband's aging workforce
AI is creating a new wave of technical debt for enterprise IT
Martha DeGrasse contributed to this article.