FEATURE
Inside Veon’s Augmented Intelligence vision
In a technology landscape increasingly dominated by hyperscalers and global AI platforms, operator Veon is carving out a distinctly different path rooted in local languages and sovereign infrastructure.
Speaking to Mobile World Live, Lasha Tabidze, Veon’s chief digital services officer explained the company’s AI approach is anchored in a people-first philosophy of Augmented Intelligence, focused on practical deployments across three verticals: consumer, enterprise, and society.
“The first and the most important one is when it touches people,” said Tabidze. “We create agents who specifically operate in the verticals people need them. It’s education, it’s healthcare, it’s financial services.”
While global AI giants like OpenAI, Google and Meta Platforms race ahead with massive general-purpose AI models, Veon is betting on localised, language-specific large language models (LLMs) trained on regional data, built for underserved markets like Kazakhstan, Uzbekistan, Pakistan and most recently Ukraine.
AI only speaks English
“There is still a bias in the world that AI speaks English and only English,” Tabidze (pictured) explained. “For example, lots of people even today, despite not being native English speakers, like me, prompt in English because there is no normal solution in my Georgian language.”
This is the gap Veon aims to bridge. The group’s AI agents are built on local LLMs and trained domestically with native datasets, incorporating regional contexts and linguistic nuances.
Just last week, Kyivstar signed a deal with Ukraine’s Ministry of Digital Transformation to develop a national LLM trained exclusively on local data sources to incorporate the nuance of dialects, terminology and cultural context. The AI model, set to release in December, will be utilised in sectors such as in education, finance, health, legal and more. “Ukrainian LLM will empower users to access augmented intelligence tools with the full cultural context and depth of their native language and national resources,” Veon Group CEO Kaan Terzioglu explained in the announcement.
Meanwhile in Kazakhstan, the company launched KazLLM, trained on Kazakh-language content and supported by local academic and tech institutions. Veon claims that results from the model are already outpacing expectations; Tabidze touted that KazLLM performs 40 per cent better than ChatGPT when responding in Kazakh because it was trained on unique, often offline datasets sourced from universities and research centres.
There is still a bias in the world that AI speaks English and only English.
LASHA TABIDZE
VEON
However, Tabidze flagged a persistent challenge in developing local LLMs: the scarcity of local language training data, much of which requires digitisation, including books, documents and offline archives. “Enrichment of data is absolutely different for the locally developed LLM rather than for the ones that just use all available data online,” he explained. “Our university and research centre… have books and resources which are simply not available online.”
A different race
On the enterprise side, from agentic HR assistants to AI-powered legal advisors, the company is developing tools trained on local datasets and tailored to sector-specific use cases. “We are not trying to create an almighty Kazakh AI agent,” said Tabidze. “We will have specifically tailored models for education… with different datasets, for farming… for healthcare.”
Indeed, in Kazakhstan, Veon deployed an AI-powered local language tutor in a commercial pilot, with plans in motion for similar local LLM initiatives in Bengali, Urdu, Uzbek, and Ukrainian.
To this end, the Dubai-based company is deliberately avoiding competing with the likes of hyperscalers such as OpenAI and Google. “This is a different race,” Tabidze commented. “We cannot make a big bet that any of them (hyperscalers) will come to Uzbekistan… and start to create an Uzbek LLM.”
Instead, Tabidze called for telecoms companies to use their unique position to build local AI with two major advantages: invaluable behavioural data from years of customer interactions and mass-scale digital distribution channels. Indeed, with over 125 million monthly active users across its apps as of the end of Q1 2025, Veon controls a direct pipeline to consumers and enterprises especially in regions underserved by global AI players.
In addition, he urged operators to rely on open-source architectures such as Llama or Gemini and layer proprietary local training data into domestic infrastructure. “You cannot do inference having the infrastructure somewhere far, far away,” Tabidze noted, emphasising digital sovereignty and local inference. “It’s important that interaction is going fast, data exchange is going fast, and agent is coming back very fast.”
Ultimately, Veon’s AI ambitions are a means to achieving real-world digital inclusion. “It’s not about creating digital products,” said Tabidze. “It’s about delivering digital products to people despite the language they speak… Then you really don’t leave anybody behind.”