Why Recent Generative AI Solutions for the Network Leverage Graphs

A key similarity is emerging among breakthrough applications of generative Artificial Intelligence (AI) for network orchestration: the use of AI to manipulate network graphs. Two key players, EnterpriseWeb (an Information Technology (IT) automation platform vendor) and China Telecom (a Communication Service Provider (CSP)), exemplify this trend. While their roles differ, both share a successful strategy, underscoring the broader implications for vendors and CSPs alike.

• EnterpriseWeb uses graph Domain-Specific Language (DSL) to enhance network visibility and drive automation. The integration of generative AI enables automation based on Natural Language (NL), without the need for code. For example, the platform can compose, deploy, and manage network services (optimizing for low latency, energy efficiency, or resource consumption) through NL-based intent. EnterpriseWeb overcomes generative AI’s reliability challenges by channeling probabilistic Large Language Models (LLMs) through a logical, deterministic environment, triggering rule-based commands for the graph DSL back end. The versatile automation platform can be applied to any network segment represented through graph language.
• China Telecom uses knowledge graphs to enhance network visibility, enabling automation and preventing data separation across various CSP units. Without generative AI, queries of the knowledge graph may only retrieve data that are predefined to address that specific query. Generative AI expands this to a wider range of data relevant for the scenario, according to learned patterns, while remaining within the constraints of graph language and structures.

Despite differences in market role, firm size, and regional context, China Telecom and the award-winning EnterpriseWeb illustrate breakthrough graph-based solutions for generative AI-based network orchestration. China Telecom credits generative AI with advancing its network automation, contributing to the type of closed-loop, Level 4 network automation that is also offered by EnterpriseWeb. This offers convincing support for their common claims: 1) integrating generative AI with network graphs can provide sufficient structure for telco-grade reliability, 2) using generative AI for NL-based graph queries and commands, rather than for extracting human-readable output can advance network automation.

These solutions may not be within the immediate purview or pathway of CSPs. For instance, the EnterpriseWeb solution is fully cloud-native and serverless for agile exchanges among public LLMs, an intermediate AI Operations (AIOps) layer, and back end servers; telco workloads still awaiting cloudification will not be ready. Likewise, China Telecom uses a proprietary LLM for knowledge graph queries, producing a unique architecture that will not be followed by most CSPs. However, the impact of these market developments is in revealing opportunities for generative AI in network orchestration down the road, drawing attention to an underrepresented area of application. Another impact is in advancing from strategy to tactic. The strategic focus involves upholding telco-grade reliability and security standards with network generative AI, facing unique challenges. This is why CSPs are proceeding so slowly and cautiously in this area, yielding to customer and business use cases. But it is becoming clear tactically what this means—that network generative AI requires channeling through deterministic environments or use with graphs, taking a problems-first and AIOps-forward approach.

The broadest recommendation offered by these two preceding cases is for CSPs to reconsider generative AI for long-term network strategies: Network-graph approaches reveal the potential of generative AI for network orchestration and offer tactics for approaching it. Although CSPs may not be prepared to implement such solutions in the short term, CSPs can begin working out the requirements and plan internal development of a Proof of Concept (POC). In doing so, two additional factors common to the EnterpriseWeb and China Telecom cases stand out as crucial:

• Prioritize Network Visibility and Data Operations: In both cases, generative AI network orchestration solutions grew out of plans to increase network visibility through graphs. This provides a starting point for investing in generative AI-based networks. It also places generative AI in continuity with existing graph-based automation solutions, tempering claims that it is revolutionary (it is not so for network orchestration), and CSPs would benefit to approach it in this way. In augmenting existing AI solutions, CSPs will need to double-down on existing investments for network visibility through high-quality data and robust DataOps practices. CSPs already invested in graph-based solutions, such as those by Dynatrace or VMware, are well-positioned to begin researching how generative AI can support graph queries and graph network manipulations to enhance automation.
• Establish Strategic Guardrails: CSPs should erect strategic guardrails preventing generative AI from interacting with network areas that are off-limits. The built-in graph languages and network ontologies create their own structures within which generative AI procedures are confined. Cautiously building this framework and ensuring high-quality meta-data helps to create effective guardrails for generative AI. Security guardrails are also necessary. Most CSPs cannot adopt the China Telecom approach to using a proprietary LLM, so they might consider the EnterpriseWeb model, separating public models and telco data using an intermediate AIOps layer. This layer adds rules for generative AI-based modification of back end graphs, adding security, while enhancing reliability.