The rapid development of standard foundational models has led the JV formed by the founding members of the Global Telco AI Alliance to change its original delivery targets. Now the JV known as Syntelligence AI, formed with $37.5 million of funding from its five founding operator partners, employs nine people and for its first product is targeting commercial development of a network-based, AI-powered service to combat fraudulent voice calls.
CEO Prateek Chowdhary says further use cases will follow, but confirmed that the company has shifted its focus away from the original intention of the Alliance to create a customised, multi-language telco LLM.
When it launched in the summer of 2023, the Global Telco AI Alliance’s original four members had grand plans to “accelerate AI transformation of the existing telco business and create new business opportunities with AI services”. A statement said, “The Telco AI Platform is expected to serve as the core foundation for new AI services, including those designed to improve the existing telco services, digital assistants, and super apps that offer a wide range of services.”
By MWC 2024 the now five member Alliance (SK Telecom, Deutsche Telekom, e& Group, Singtel and SoftBank) said they planned to develop Large Language Models (LLMs) specifically tailored to the needs of telecommunications companies. The LLMs would be designed to help telcos improve their customer interactions via digital assistants and chatbots.
“We as telcos need to develop tailored LLM for the telco industry to make telco operations more efficient, which is a low-hanging fruit. Our ultimate goal is to discover new business models by redefining relationships with customers,” said SKT CEO Ryu Young-sang.
In June 2024 at the TM Forum’s DTW event, the Alliance’s founding members outlined plans for the telco LLM, signing a joint venture agreement to develop and launch multilingual LKMs specifically tailored to the needs of telco companies. “Singtel finds that a customised telco LLM definitely outperforms a general LLM for telco-related tasks,” said William Woo, Group Chief Information Officer / Chief Digital Officer, Singtel, at the launch.
We decided to defer spending time on telco-domain specific LLMs just because the standard LLMs are becoming very good
Syntelligence is formed
Although the intention to form the JV was signed in mid-2024 it took until late 2025 for the JV to actually come into being, under the Syntelligence name. By then, things had changed, and Syntelligence is a perhaps a different-looking company compared to the expectations formed by the Alliance in previous announcements.
With a staff of nine, it is developing one main product to provide network-based, cross-platform protection from fraudulent and spam voice calls. The product will use network and subscriber data such as traffic patterns and roaming status to determine likely fraudulent activity. During the call, if it is from an unknown number and if there is consent from the user, it can even monitor the call using an AI model and then alert the user in real time to make them aware of a possible scam. Syntelligence CEO Prateek Choudhary says the company is building AI models that can have “decent conversations that don’t go into loops and loops that would gather good information from you, why you’re calling, who are you? How urgent it is. And after that call is ended, we’ll give the summary to people.”
Choudhary is a man with spells at building foundational models at Meta and Amazon on his CV. He says the ambition still remains to “solve the biggest problems within the telco domain using AI technologies”.
“Essentially, the goal of Alliance was to build a company, a joint venture, which was fulfilled. It took longer than we thought because, you know, there were so many companies involved, but that was the ultimate goal and we are very happy that it’s live now. What we are doing is working with our founding members to see what are the biggest problems they have today, and we will tackle those one by one using the AI technologies that we have.”
But what has changed is how Syntelligence AI will do that.
No more telco-specific LLMs
Despite the repeated aims of the Alliance members through 2024 and 2025, there has been a shift away from developing telco-specific LLMs. That is explained, Choudhary says, by the advances in standard foundational models.
“It was always planned that the Alliance would build a startup company that could move faster without the bureaucracy layers, and where data could be shared to handle AI opportunities that, indirectly or directly, are not being handled by the internal AI teams. That was always the plan.
“What has shifted is the initial plan of actually building and training telco-based LLMs. Two years ago standard foundational LLMs were not as good, but they are becoming better every week. So now the question that arises is, ‘Is a generic standard LLM good enough – because they are very, very good – or do we still need pre-trained models based on telco data and network information?’
“The answer probably depends on the application where it’s being used, but we decided to defer spending time on telco-domain specific LLMs just because the standard LLMs are becoming very good, and for the application that we are building right now – we realised the advantages that we would have from being pre-trained with telco data would not give us a huge leg-up.”
Choudhary said there may yet be a need for those telco-specific LLMs, for instance in customer care or contact centres, “where there’s very specific telco-level information that customers ask for”.
“But even there it’s a little bit questionable, because, these days, you can actually have a very good model that can read and consume all the knowledge base that telcos have and produce information. So you don’t really need to train a foundational model on that data, you can actually get the data later on.”
Aside from the change of target from producing telco-specific LLMs to relying on standard foundational models, there has also been a narrowing of focus. Choudhary said the plan is to stay small and agile, and prove that the team can deliver a commercial, production grade anti-fraud model. Then other use cases may arrive.
We want to make sure we are building one product really well, and it’s not a very easy problem to solve
“We are a four month old company. We want to make sure we are building one product really well, and it’s not a very easy problem to solve. There have been other products on the market, so we are trying to make sure we are building a very robust one. Right now, our focus is on this. Once we do this, and we do this well, we will go on looking at other products.”
The JV partners may well buy the technology, but there are no guarantees, Choudhary says.
“Essentially, all the founding members evaluate us against the solutions that are in the market. So it’s not that we are getting favoured status. What we are definitely getting is insights on the biggest problems that telco have, and expertise on privacy and regulation. And we are, of course, getting data to train our models from these partners. So they definitely have skin in the game which is really, really good. But the idea for Syntelligence AI is to be an independent entity, generating revenue.”
