Safely applying rich telco datasets is key to GenAI gains

Q: How are telcos overcoming GenAI challenges to realise its true value?

A: The emergence of Generative AI (GenAI) offers clear potential to transform the telecom industry, ushering in disruptive innovation across every aspect of the business. From enhancing customer support resolution and accelerating onboarding of new partners to optimising planning efficiency for new network builds, the technology promises a revolution in productivity and monetisation.

However, Large Language Models (LLMs), such as Open AI’s GPT-4 – the engine of ChatGPT – are only as accurate as the data they are trained on. For operators to fully realise the promise of GenAI, a deep understanding of the telco domain is needed, specific needs of the Communication Service Providers (CSPs) need to be understood, and – crucially – how to securely harness valuable telecom data needs to be determined.

The majority of telco GenAI use cases under discussion today require the use of highly sensitive BSS/OSS and knowledge base data. The ability to access this information while safeguarding privacy and security will therefore be pivotal.

To overcome these challenges and maximise the opportunity, CSPs across the world are now forming dedicated teams and bringing in the expert skills needed to accelerate the successful adoption of GenAI technology in their business.

Q: What are the challenges CSPs face in fusing GenAI models and telecom?

A:  Choosing the right model strategy is the first challenge in bringing GenAI technology to telecoms. 

On-premise custom models, specifically trained on the telco business, appeal to some operators to keep control of their data and alleviate the security concerns of public models. However, custom models require a considerable amount of training data, consume vast compute resources, and need significant AI expertise. OpenAI recently disclosed that its GPT-4 models cost over $100m to train, with daily running costs in the order of $700k. For CSPs with the ambition to build a custom LLM that performs to the current standard of GPT-4, this level of investment may be cost prohibitive.

The majority of operators are keen to work with readily available advanced public LLMs that have been pre-trained on vast quantities of publicly available data but have no knowledge of the telecom business and its complex processes. A recent TM Forum report cited that 59 per cent of CSPs are opting to use an open-source, cloud model and off-the-shelf LLMs for quick deployment and scalability. Fine tuning techniques can bring some domain knowledge to these pre-trained models, however it’s still a complex process requiring the right skillset. And these operators are rightly concerned about the leakage of proprietary customer sensitive data and need robust security solutions in place to take advantage of this powerful technology.

Further, CSPs will need to consider the real-time challenge. Many telco datasets are constantly changing and in a state of flux – for instance, mobile device data usage statistics or inventory systems reflecting real-time topology. This makes a sizeable proportion of telco data unsuitable for the technique of fine tuning GenAI models, which is typically used for refining pre-trained LLMs (public or on-premise) with domain-specific data.

For CSPs to overcome these challenges, fundamentally new approaches are needed to mediate between different types of GenAI models, users, and proprietary telco data and business intelligence.

Q: What are the steps and techniques operators are taking to overcome these issues?

A: Operators are investing in a platform approach that isolates the GenAI models from the users and telco data. This abstraction layer performs three fundamental roles: 1) it avoids being locked into any specific LLM/FM 2) it provides an essential security layer that protects sensitive customer data from public models, and 3) it ensures the highest quality of responses to maximise GenAI’s value across the business.

A single model, public or on-premise, will not be suitable for all use cases at the right price points. Moving forward, telcos will require multiple GenAI foundational models that are tailored for specific business needs – including complex or simple tasks, or those focused on images, designs, or code. Operators will need a flexible choice of models and the ability to adapt their model strategy over time without major changes to their GenAI implementation strategy. This GenAI platform approach will enable model flexibility.

Public GenAI models need telco data to perform their role and successfully resolve a query. The key is to provide the required information but ensure that the model never gets access to sensitive customer or company data. In practice, this demands the use of sophisticated anonymisation techniques – such as just-in-time anonymisation or obfuscation – to prevent confidential customer data from being exposed to public models. As an example, before being sent to a GenAI model, a customer’s mobile number, address, or name will be substituted with pseudo data. This forms part of a robust security framework in the GenAI platform, including strict data access control measures, that must be instituted across the entire GenAI ecosystem.

To ensure the highest quality responses in the shortest time, user prompts must be augmented with additional information – in the form of real-time data, context, and instructions – giving the GenAI model everything it needs to create the optimal response. This context-prompt enrichment technique is executed in the GenAI platform, leveraging technologies including prompt engineering and retrieval augmented generation (RAG), working securely with real-time data through API calls to BSS/OSS systems and ensuring that the model has the most current information.

By combining this enrichment process with model fine-tuning of more static data, CSPs can confidently and compliantly open up their extensive domain knowledge to powerful GenAI models – public or private – for a wealth of high-value use cases.

Q: Can you provide some user cases for GenAI in the telcos market?

A:  Having successfully bridged the value of GenAI and BSS/OSS data, CSPs can focus on the immediate GenAI use cases, which are anticipated across all areas of the business.

In customer care, LLM-based digital assistants could exponentially improve call centre efficiency and the overall customer experience – helping support agents to provide more accurate responses, respond to customer queries in multiple languages, and resolve complex customer issues more quickly. According to McKinsey, a telco in Latin America has boosted call centre agent productivity by 25 per cent and improved customer experience quality by enhancing its support agent skills and knowledge with GenAI-driven recommendations. Likewise, operators like T-Mobile US will use GenAI to improve their wholesale partner business by solving customer problems faster and driving more efficiency into its business.

We’re also beginning to see GenAI’s impact in areas such as complex network operations – with advanced GenAI capabilities being leveraged to further automate certain aspects of network planning, installation, configuration, and closed-loop assurance. Significant benefits are predicted from these pioneering implementations, which include reducing critical incidents by 35 per cent, decreasing performance problems by 60 per cent, and boosting workforce productivity.

As GenAI becomes more and more integrated in the telco space, operators will be empowered by significant productivity gains spanning all areas of the business, including sales, marketing, business operations, and network operations. To help drive the marketing function, GenAI can quickly generate text and images to support the roll-out of personalised promotions and campaigns. By leveraging GenAI to personalise content, a European telco recently increased marketing campaign conversion rates by 40 per cent, while reducing costs. Sales will benefit from GenAI by using it for sales intelligence and providing advice on customer requests while, on the operations side, GenAI can assist field technicians by providing instant access to information about networks, topology, and planning data – enabling issues to be fixed and new systems installed faster. 

The ability to safely apply rich BSS/OSS datasets to the latest GenAI technology will be transformational for global operators. CSPs that lead this innovation will create a cascade of competitive advantages to include reducing costs through improved customer support and increased staff efficiency; enriching provisioning and troubleshooting support; boosting revenue through business agility; improving predictive maintenance and fraud pattern modelling and – ultimately – delivering exceptional, personalised customer experiences based on an unprecedented understanding of individual needs.

Sue leads strategy and portfolio marketing at Netcracker responsible for defining the marketing strategy and executing marketing initiatives across Netcracker’s BSS/OSS and Orchestration portfolio. She brings over 20 years of experience in the telecoms industry, spanning a variety of leadership roles including product management, strategic planning, product marketing and technical sales. Her expertise encompasses a wide range of technologies including cloud, 5G, SDN/NFV and BSS/OSS with a strong focus on generating business growth. Sue has a Bachelor of Engineering honours degree in Electronic Engineering with Communications from the University of Sheffield in the U.K.