The expectations for generative artificial intelligence in the telecom space are high: McKinsey & Co. says gen AI could revitalize telco profitability; Ericsson says that it is already transforming the industry. The applications of gen AI are manifesting in customer care chatbots and more personalized offers, with more opportunities being eyed in network-specific areas ranging from wireless channel modeling to configuration and threat detection for network security.
AI and ML aren’t newcomers in the telecom network; they have been part of testing, monitoring and assurance tools for years now, quietly evolving. But Raj Savoor, VP of network analytics and automation at AT&T, says that the new compute power that GPUs bring “really is an inflection up. It’s a massive step-function up.” While AI in telecom is still being developed, “The long-term potential, I think, is understated. I think there’s huge potential,” he said.
Savoor sees three broad areas for the use gen AI in the network context, particularly related to network operations, testing and assurance.
Co-pilots. Not the actual Microsoft Copilot AI engine; Savoor was referring generally to the use of gen AI as an assistant to humans to provide answers to questions, summarize various types of information and generating code, text, images and multimedia—being trained with network manuals from vendors, for instance, or composing network incident reports. “Almost any place where you have human engagement is a candidate for using a bot or AI,” Savoor said. There are many potential scenarios in which gen AI co-pilots could augment the telecom network from planning to decommissioning; Savoor noted that the question of whether or not to involve AI in a given scenario comes down to: What is the value involved? And what are the risks involved?
Because there are also limitations to the use of gen AI as a co-pilot in a network context. In the webinar conversation on this topic, Savoor said that “generalized bias” is at work in gen AI models, meaning that they often will produce a “safe” answer—one that is unlikely to be wrong, but may also be generic and unsatisfactory to a human expert like a network engineer. In some cases, he added, an AI’s answer may actually be wrong because it responds with generalized or “diluted” information that does not accurately reflect specific telecom parameters or implementation details in a particular network.
“That is a challenge, and one of the reasons why we are looking at network foundation models,” Savoor said. He differentiated such models from both existing LLMs or Small Language Models (SLMs). LLMs, he noted, are based largely on fluency in language and data in the public domain. Telecom networks, however, operate on a set of very specific and complex data types: network configurations, for example, and various types of numerical data. “The LLMs are not quite as mature there,” he said, so telcos have to invest in the potential of the models, and transform their multi-model data so that the models can be trained and apply it. AT&T, he continued, is taking a bottom-up approach and not just training models only with documents, but also working on fine-tuning by associating context and labeling in the training data. It takes work and investment up-front, he said, to get a model that ultimately provides more precision for a telecom network context. “It’s not something you can broadly do and have one model for the entire network,” Savoor warned. “You have to be surgical: Pick that specific domain and understand the chain of thought for that audience that you’re targeting the capability for, and work that.”
Insights for closed-loops. While closed-loop automation in the network has been something the telecom industry has been working toward for a long time, Savoor says that there is a new, emerging role for gen AI.
While machine learning does not inherently make recommendations or decisions, Savoor said that artificial intelligence engines can look at a history of cause and effect and make recommendations. “It is very similar to how humans act, saying that we’ve learned from these past tests, and we know when there is a cause and effect; this is a good outcome, this is a bad outcome, this is a false positive, this is a false negative; and it can go take that to work,” Savoor said. “The levels at which human thinking goes, how we synthesize data and then apply it—that is baked into that engine.”
Because of this capability, “There is real power in the decision-making function to summarize, find patterns, draw insight and then take it to something actionable,” he explained. Gen AI, then, could trigger testing functionality or measurements, take the resulting data, extract insights and then make a recommendation for closed-loop action to be taken.
Synthetic data. With artificial intelligence comes concerns about data privacy, regulation and permissions for the use of actual network and user data. Generative AI is seen as potentially being able to produce realistic but synthetic data as needed, to refine and streamline testing. “The generative synthetic data use cases is a pretty unique area that we are focused on,” Savoor said. “We see a huge need for synthetic data to drive both predictions and digital twin use cases. This is particularly relevant for network and field testing where field measurements can then be used to drive tests cases, especially around impairments and overload, in the lab.
This is an emerging field, he says, which has come out of machine learning. Machine learning has enabled statistical analysis to identify trends based on patterns in time, and has even been able to make some predictions. “The advantage of gen AI is, it establishes more relationships than this traditional ML model,” Savoor says. “So when it sees different bodies of data, it finds better patterns.”
He offers an example of interference measurements taken in the field, across different RF environments—dense urban, suburban and rural—with and the characteristics such as reflections in each environment. Traditional ML models would look at trends within that data without having insight into, say, the morphology or geospatial information. Generative AI, however, could—with the right contextual labeling—make associations across those relationships and provide synthetic data accordingly. That, Savoor said, results in a much more powerful way to, for example, test a new massive MIMO radio for a particular environment and get a sense of what kind of traffic patterns and interference should be expected. “To be able to go do that, with that context, is better than traditional ML models that have existed,” he added.
Looking for more insights on the use of artificial intelligence in network testing and assurance? Check out the available webinar on-demand, featuring AT&T and Spirent Communications and look for our in-depth editorial report on the topic, coming later this week.