In April, Bain & Co published a report about the potential effect of Generative AI (GenAI) on the world’s telecom companies. Given the title: Telcos, Stop Debating Generative AI and Just Get Going – the urgency of the message was clear.
In the fast paced world of GenAI, the debate has certainly moved on. According to TM Forum’s latest GenAI telco survey, 53 percent of respondents say they have already set up a GenAI center, and 59 percent have identified families of use cases and specific use cases in each family. AWS also conducted its own survey focused on the telecom industry and revealed that half of telcos expect to adopt GenAI within two years.
However, the question remains – How can GenAI be safely introduced into the telecom domain? And equally important – what will it take to tap into the precious telecom data and knowledge systems so that it adds real value to the business?
Of the many GenAI telco surveys available, safety usually ranks as the number one challenge to overcome.
Understanding the problem
GenAI models, such as Open AI’s GPT-4 that powers ChatGPT, are only as good as the data they are trained on. With over 1.7 trillion parameters of training data, the model has superb communication and content creation skills. But it doesn’t know the world of telecom. And as we know, telecom is extremely complex.
Training public GenAI models on the telecom domain is simply not an option. The most obvious reason being security. Telco networks host highly sensitive customer and network data in its BSS/OSS systems, and this data cannot be given open access to public models.
There’s also the real-time challenge. A significant amount of telco data is constantly changing. For example, mobile device data usage or an inventory system reflecting real-time topology are in a constant state of flux. This makes a large chunk of telco data unsuitable for the technique of ‘fine tuning’, which is used to refine pre-trained models with domain-specific data.
Accuracy is another issue. GenAI models are only as good as the data they ingest. Since they don’t ‘know’ the telco business and its processes, they may make assumptions that can produce erroneous results.
Then we get to the cost. The most advanced LLMs in the market today are good at what they do because of the enormous data sets they have been trained on. According to Open AI, it cost over $100 million to train GPT-4, and an incredible $700k/day to run it. For communication service providers (CSPs) who want to build their own custom LLM, and achieve the quality found in GPT-4, this level of investment may be out of reach for many.
How to get telecom domain knowledge into GenAI
Tapping into the telco’s BSS/OSS and knowledge base systems is a must to obtain any real value from GenAI technology. An effective method to access this domain-specific information is to augment user prompts with additional context – in the form of real-time data and instructions – giving the GenAI model everything it needs to create the optimal response.
This context-prompt enrichment technique, leveraging technologies including prompt engineering and retrieval augmented generation (RAG), has the benefit of:
- Returning the highest quality responses as the model is grounded with relevant data and instructions. In fact, this helps to eliminate the issues of hallucinations that have appeared in public models.
- Working securely with real-time data through API calls to BSS/OSS systems, which ensures the model has the most current information.
- Accessing up to date knowledge base systems, which have details about the telco business – networks, services, business processes, systems, and more.
- Working with any mix of public and custom-built GenAI models.
- Reducing the time to resolve a request and subsequent costs as the model has more accurate data requiring less ‘too and fro’ with the user.
By combining this enrichment process with model fine tuning for more static data, CSPs can unleash their extensive domain knowledge to powerful GenAI models (public or private) and begin creating an assortment of valuable use cases.
How do we make it all secure?
The telecom business has very strict confidentially rules due to its vast customer data, as well as regional laws including GDPR. Strict security measures are needed when using public models or hosting private models on a public cloud to avoid privacy violations.
The GenAI model needs telco data to do its job and resolve a query. However, the key is to make sure the model never gets access to sensitive customer or company data. This requires the use of sophisticated anonymization techniques – such as just-in-time anonymization or obfuscation – to keep customer data confidential and protect it from exposure to GenAI public models.
For example, a customer’s mobile number, address, or name is substituted with pseudo data before being sent to the GenAI model. When the model returns the response, the real data is reintroduced for communication with the user.
This forms part of a robust security framework, including strict data access control measures that must be put in place across the entire GenAI ecosystem.
Fusing GenAI models and telecom
Across the globe, CSPs are forming teams and bringing in the skills needed to apply GenAI technology to their telco business. However, the fact remains – telcos face unique challenges in bringing GenAI to fruition. From fluctuating data and erroneous data sets to costs that can prove to be prohibitive, CSPs need a way to tap into their BSS/OSS and knowledge base systems without risking the confidentiality of the sensitive customer and network data in their possession.
This necessitates a new and essential component of the GenAI ecosystem that sits at the intersection between GenAI models, the users and the proprietary telco data sets. With this new capability, CSPs can extract real time data as and when needed, enrich models with context-based prompts, fine tune models and execute sophisticated data anonymization techniques to keep sensitive data secure. The same techniques can be used for any public or private GenAI model, enabling CSPs to choose the mix of models that bring the best returns for their business.
With this fusion of GenAI and the telco domain, CSPs can unlock the tremendous potential GenAI has to offer and begin to safely execute the first use cases.
