At a recent roundtable discussion, Veon’s Lasha Tabidze described how the group is positioning itself as a service delivery company with LLMs at the core of its AI strategy.
The group has just launched AI Tutor, a Kazakh-language AI-powered learning assistant built on the locally developed KazLLM large language model created by its local operating and software development units, Beeline Kazakhstan and Qazcode. AI Tutor is embedded into Beeline’s Janymda super-app and offers modules to support Kazakh language learning for children and adults, although its capabilities will be expanded in the future.
Tabidze noted that operators are playing a ‘different game’ from hyperscalers with their approach to AI, noting that operators offer the cheapest digital distribution channel, allowing them to push products at a mass level. This is essential for customer-facing AI, as it allows consumers to see the value immediately. Similarly, it allows easier access for businesses to use KazLLM ‘as a service’ and train it using their own data to tailor it for their specific industry.
This ‘LLM-as-a-service’ concept leans on Veon’s tech expertise, with proprietary AI built on top of foundational models, i.e. the more widely known generative AI models such as Meta’s Llama, Google’s Gemini, or DeepSeek, which are used as a baseline. The LLM can then be monetised on a subscription basis to educational institutions etc – Tabidze claims that the main component for creating value is “where AI touches the people”, i.e. augmented and agentic intelligence. By putting this value in people’s hands, they understand it and so are willing to pay for it. Veon intends for companies to use KazLLM to develop internal agents that will help improve the model’s efficiency and drive digital inclusion in this market, for example by helping small businesses upscale their operations. Tabidze notes that developing Kazakh-language educational services delivered a huge nationwide benefit, highlighting the importance of identifying the key focus areas for each market.
“On top [of foundational models] you create the layer where data is secured, created by local people, which is trusted by not only governments or the private sector, but also people like us. When…you trust the company, you start to use it more in your daily work and life. [Our] Kazakh language [LLM], now it’s very good – it’s 40% more precise than if you try to do the same thing through – for example – ChatGPT in the Kazakh language. We started small, and then we scaled – and on the way, we’re learning…it’s very important that several parties are involved. You cannot develop it yourself. You have to have good relationships with every academic institution.”
By forming both public and private partnerships, operators foster trust from all sides, allowing them to become service enablers. International partnerships provide access to databases across different languages, improving its offering and thereby underlining the value of the LLM to the user base. Telcos have spent years amassing data on customer behaviour in a single silo, which is a huge advantage as most siloed data is unstructured and in different formats. Tabidze stated that while the model is transferrable, the datasets are not; structuring the data to be fed into the model is challenging, as is data enrichment on locally developed LLMs where wider datasets are not available. “We’ll have [data] specifically tailored for farming… for healthcare. if some other companies, startups or regular companies, have different very good data sets, that’s where partnership with companies comes into [play]…so you cannot have the closed ecosystem.”
Veon aims to create versions of KazLLM for its other markets, using local insight to help train the model in linguistic nuances, on top of the foundational models. Tabidze named Bangladesh, Pakistan and Ukraine as potential target markets.
