How SLMs help CSPs achieve more with less – while paving the way to reimagine BSS, OSS and network autonomy

Why are CSPs looking to SLMs instead of LLMs?

Small language models (SLMs) can now improve a wide range of complex processes by allowing them to operate efficiently on devices with limited computational power, such as smartphones, IoT devices and edge servers. This allows CSPs to deliver real-time, context-aware services, improving both user experience and operational efficiency, opening new revenue streams and securing customer loyalty. This development helps to mitigate the limitations of large language models (LLMs) – which, while impressive in natural language understanding and generation, face significant barriers to delivering effective end-to-end business solutions in the telecommunications industry.

CSP operations rely on complex OSS and BSS systems - requiring solutions that are secure, customizable, scalable, cost-effective, and performant at scale. Privacy and governance concerns arise when using LLMs in these systems, especially when handling sensitive data, as centralized infrastructures hinder compliance with security standards and regulations. LLMs also incur high operational costs, requiring substantial infrastructural and management resources, as well as significant computing power. LLMs further lack the flexibility to effectively address the specialized use cases unique to CSPs.

The Mighty Minions: Unleashing domain specific GenAI via SLMs Phase II Moonshot Catalyst, which won the Best Moonshot Catalyst – AI Challenge Catalyst Award at Innovate Asia 24, shows how SLMs could co-exist with LLMs for specific downstream tasks, offering lower overheads and better data governance. The project also demonstrates how SLMs offer greater customizability, enabling CSPs to tailor AI solutions flexibly to specific needs like network management and customer support, leading to faster deployment times. Phase 1 proved the viability of SLMs through proofs of concept – Phase 2 has developed production pilots for three key CSP business functions, and a scoring framework to help CSPs decide whether to deploy an SLM or LLM based on security, cost, performance, and scalability criteria. The Catalyst draws on contributions from participants at Axiata Digital Labs, TCS, and AWS, with the support of Project Champions at Ezecom (Cambodia), Omantel, TPG Telecom (Australia), SLT Mobitel (Sri Lanka), Verizon, Econet (Zimbabwe), and Dialog (Sri Lanka).

How SLMs can be used - from infrastructure to invoicing

Domain I of the Catalyst dealt with infrastructure management – for which a copilot was developed using SLMs to automate virtual machine provisioning in an OpenStack cluster to streamline resource planning, monitoring performance and utilization. Manual intervention was thereby reduced, ensuring optimal use of cluster resources, while automated root cause analysis reports supported troubleshooting and reduced resolution time. SLM use here enabled precise, task-specific automation, clearly illustrating the benefits of focused Gen AI applications in driving efficiency and scalability in critical infrastructure management processes.

Domain II focused on knowledge plane management, by integrating SLMs into knowledge plane components, using fine-tuned domain-specific language models and knowledge bases to improve intent-based network anomaly management. The SLMs incurred lower recurrent costs than LLMs, supporting a more efficient, autonomous operations pathway. This approach therefore proves foundational in realizing experiential intelligence use cases in intent-driven service optimization and customer engagement, helping CSPs to achieve higher autonomous network maturity (L4/5).

With Domain III, the project addressed billing management by seeking new automations in anomaly detection and invoice explanation. Billing operations teams need real-time anomaly detection to identify suspicious usage and proactive alerts for billing issues, while customers require clearer bill explanations, with line-item breakdowns and spend optimization suggestions. By using TM Forum APIs TMF 635 and TMF 678, the solution flagged billing discrepancies via sudden usage spikes, unusual call patterns, or unexpected locations, and introduced a proactive bill explanation section in customer bills and dashboards – helping to improve customer experience, reduce churn, and minimize billing queries.

Pilot proof points: efficiency, cost and customer experience

These pilots demonstrate the effectiveness of SLMs in addressing key challenges while showing that smaller, targeted models can achieve superior performance with fewer resources than LLMs - and business benefits are already clearly measurable.

The infrastructure management copilot yielded an 86.67% increase in resource provisioning efficiency, with a similar cost saving in overheads on engineering hours. The knowledge plane management solution improved correlation between customer experience and network experience, with rich knowledge-based inferencing, leading to accuracy and efficiency improvements of 15%. Meanwhile the billing management solution achieved a reduction in helpdesk calls of 20%, cost savings of $2.5K per 1000 discrepancies, higher CSAT and NPS scores, and clear revenue leakage prevention.

The Catalyst also illustrates that SLMs can deliver high performance while remaining energy-efficient, and the new scoring framework helps CSPs avoid unnecessary investment in LLMs where SLMs can suffice – while allowing LLMs to be used when their broader capabilities are needed. “As a leading ISP, Ezecom plans to implement the SLM-based OpenStack copilot to manage our internal infrastructure and data centers,” explains Prasad Jayasinghe, Head of Software Engineering at Ezecom. “We aim to improve operational efficiency and productivity through exceptional automation, real-time monitoring, smart hardware provisioning, and detailed RCA reports, ensuring seamless operations and quicker issue resolution process.”

SLMs offer the telecommunications industry major opportunities to improve operational efficiency and innovation, enabling CSPs to navigate new complexities and meet evolving business needs. To find out more about how this technology can support your organization’s needs, or to support this Catalyst, please visit the project space here.