Widely available small language models for telco applications now within reach

Commercial context

One of the clearest trends to emerge in the industry over the last 18 months is the use of generative AI (genAI) and large language models (LLMs) to improve business operations, to provide better service and value. Yet reliance on LLMs presents significant infrastructural demands and security concerns – private hosting can be costly, and public hosting risky, meaning that CSPs must compromise business efficiency with the need for data security.

The Mighty Minions: Unleashing domain-specific genAI via SLMs Catalyst is exploring alternative means to achieve desired business use cases without compromising on security. The answer lies in using small language models (SLMs) finely tuned with domain-specific data to addresses key requirements within the telecoms industry. These industry-specific applications include network anomaly detection, customer support chatbots, revenue assurance analytics, order management optimization, social media comment analysis, and as a conversational assistant for edge devices.

The solution

Project stakeholders identified a number ways in which genAI and its natural language processing and content management abilities could improve upon common pain points. The resulting SLM was then put to use in a variety of scenarios. Perhaps its most significant success has been in the real-time monitoring and analysis of data to identify network irregularities, where it has been used to proactively respond to potential issues and minimize network downtime. Although in an early phase of development, SLMs are showing promise in supporting edge devices, such as set-top boxes and network routers, where they can provide personalized troubleshooting interfaces. These pre-trained models provide a user-friendly interface for resolving issues locally on the device, improving customer experience and reducing the need for external support.

In customer-facing areas of the business, SLMs have shown themselves to be just as successful. For example, AI-driven chatbots, powered by SLMs, provide personalized assistance to customers and support agents by understanding and responding to inquiries effectively, enhancing customer satisfaction and reducing support costs. This application is also being applied to social media as well, where SLMs are able to analyze customer sentiment in comments and integrate feedback into existing customer data platforms. This capability helps provide meaningful feedback to customers, enhancing engagement and satisfaction by addressing their queries promptly and effectively.

SLMs have also demonstrated their value in common CSP business operations, such as analyzing revenue streams to detect billing discrepancies and address revenue leakages. By using data analytics and predictive modelling techniques, these AI-driven solutions ensure accurate billing and optimize financial performance. SLMs have also demonstrated their ability to streamline order management processes by automating tasks and optimizing resource allocation, improving order fulfilment workflows – all of which can reduce operational costs.

Commercially, all of this amounts to multiple routes for CSPs to unlock new revenue streams and enhance customer satisfaction. Crucially, these SLM-based applications are capable of being deployed within existing telecom ecosystems, making adoption far easier. For companies seeking to deploy such solutions, this approach means that opportunities for growth and customer engagement come without the need to invest in developing their own SLMs. Adoption is simplified by using TM Forum best practices and standards as laid out in TM Forum's Ecosystem Modelling Framework (GB1032), Information Framework (GB922), and AI Maturity Model (GB1003A), for example. Using these in conjunction with Open APIs and the AI Governance Toolkit (IG1238, IG1239, IG1232), makes it easier to integrate SLMs while also supporting ethical and responsible AI practices throughout the project lifecycle.

Because these models can be run within CSP-owned environments, data remains secure and eliminating the need to share valuable datasets with external hyperscalers. This approach enhances data privacy and reduces the risk of data breaches. Furthermore, the use of SLMs contributes to sustainable AI practices by reducing carbon and water footprints due to their lower compute power consumption. This makes SLMs a more environmentally friendly alternative compare with traditional LLMs.

Wider application and value

This Catalyst has illustrated the advantages of using SLMs over traditional LLMs, particularly in terms of cost reduction and operational efficiency. Due to their smaller size, SLMs consume fewer computing resources, resulting in a 30% reduction in the total cost of ownership for genAI solutions. Moreover, their streamlined model tuning process reduces the time and effort required to fine-tune domain-specific models, making it easier for CSPs to deploy AI capabilities in applications where LLMs would otherwise be impractical.

According to Ayoma Wickramaarachchi, Deputy General Manager – Product Development and Management at Sri Lanka Telecom PLC (SLT-Mobitel), “the widespread adoption of small language models across industries has broader societal implications. By enabling more efficient and personalized customer support, these models contribute to a positive user experience and foster trust between consumers and businesses. This can result in increased digital literacy and confidence among users, empowering them to embrace and utilize technology more effectively in their daily lives.”

Part of the project’s success has been to make significant inroads in demonstrating that SLMs can increase CSP revenue growth by 15-20%. This huge growth is realized in part by the provision of domain-specific capabilities as APIs to external parties, which will help create new revenue opportunities. This will make it far easier for small and medium-scale enterprises to incorporate AI capabilities into their applications with low cost and effort. SLMs can be deployed across a wide range of devices and applications, simplifying the model tuning process and making AI more accessible. By enabling the integration of SLMs within existing telecom ecosystems, the industry is can be confident that it can remain competitive with big tech in the era of AI.

To learn more about how this initiative will support development of SLMs) for industry-specific genAI applications, please click here to view the project space on the TM Forum website.