Summary Bullets:

• Multiple small language models (SLMs) are likely to be more affective than one big large language model (LLM).
• AI can offer proactive root cause analysis informed by multiple data sources.

Agentic AI has dominated the tech headlines in recent months and was the hot topic at this year’s MWC event in Barcelona (Spain). The technology will play a significant role in revolutionizing indirect sales and customer support processes across pretty much all verticals. It can also play a major part in one of the most important challenges facing telcos at the moment: automating their networks.

Much automation work has focused on more traditional AI technologies and machine learning techniques supported by adjacent features such as APIs. All of these will continue to be critically important. Agentic AI is perhaps a less obvious fit. It isn’t necessary for most automated decision-making processes for core network platforms. Similarly, most network buying processes will either be as part of a managed service process, or through a GUI- or API-enabled self-service marketplace.

There is some potential for agentic AI support in this latter scenario: e.g., asking for recommended settings for certain products – and further down the line even asking AI for recommendations on, for instance, the best access service to use in a given scenario. This sort of support is helpful, but not necessarily game changing.Where agentic AI can really come to the fore is in ongoing network monitoring and management. AI is already in play in network portals to generate alerts and to automatically create trouble tickets. Orange Business Manage Service Watch proposition guarantees that a minimum of 10% of trouble tickets will be issued by automated AI – and it intends to increase this percentage over time.

However, at present, AI hits challenges when it comes to more proactive scenarios involving customer WAN networks. Telcos can automate their own network platforms according to their own internal policy frameworks. Programming AIs to be able to able to operate inside a customer’s network and to be trusted to make decisions automatically is a much bigger and more difficult step.

Agentic AI can bridge the gap and offer a concierge like service even to customers favoring a self-manage model. An agentic AI agent with access to network data can alert enterprises to problems that are impacting network performance, and therefore the end user experience. Furthermore, an agentic AI agent will be able to suggest the likely cause of the problem from multiple data sources including the customer’s network, the provider’s own network platform, and data feeds (e.g., increased DDoS attack activity) as well as from external feeds such as social media (e.g., Is there a large-scale online event in a specific area causing more latency on internet access services than normal?).

The agentic agent will be able to offer temporary and/or permanent fixes to the problem. Ideally, it will also be able to use digital twinning techniques to demonstrate what the likely impact of each potential fix will be on network and application performance: e.g., what will happen if bandwidth on a given circuit is increased by 50%, or if traffic is rerouted. The AI agent should also be able to reverse temporary fixes (such as increased bandwidth) once the scenario requiring that increase has ended.

Programming LLMs to be able to operate across all of a network is not an easy challenge, and the cost of running AI agents may well be more expensive that using human support staff. Creating a situation where multiple SLMs, each concerned with different parts of the network, are able to communicate may well be both more efficacious and more cost effective.

This kind of proactive network monitoring is what enterprises want and offers a way to offer genuine business value and competitive differentiation.