TBC Uzbekistan Cuts Cost per Interaction From $0.35 to $0.03 With In-House AI 

Building artificial intelligence capacity inside a bank is easier to announce than to complete, and the record of TBC Uzbekistan offers a useful case study in what completion looks like. Between January and November 2025 the bank ran a dedicated transformation programme that absorbed 45,000 man-hours, produced two AI products now working in production, and cut the cost of each customer interaction from $0.35 to $0.03, a reduction of more than 90 percent. 

The programme was led by Andrey Malygin, Head of AI Platform, and its stated aim was to create an internal AI foundation that does not depend on outside vendors. That ambition shaped nearly every technical decision that followed. 

From a pilot idea to a vendor-independent platform 

Many financial institutions buy AI services as finished boxes and plug them into existing systems. The approach taken in Tashkent was different. The team built its own platform, supported by a technology stack that includes PyTorch and TensorFlow for machine learning, LangChain for orchestration, and Kubernetes and Airflow for deployment and scheduling. Monitoring runs through Prometheus and Grafana, and data governance rules define who may use which data and how models are checked after launch. 

Scale makes this independence more valuable each year. As the number of conversations grows into the millions, even a small saving per interaction compounds, and a platform that the bank controls can be tuned for that volume without renegotiating a contract. 

Independence from vendors has three advantages. It reduces long-term licensing costs, it keeps sensitive customer information inside the country, and it lets the bank change models quickly when behaviour or regulation shifts. All development and infrastructure remained in Uzbekistan, with the team based in Tashkent. 

Two assistants working in production 

The practical output of the programme is a pair of assistants. The Sales Assistant supports conversations with prospective customers, while the Collections Assistant handles reminders and repayment dialogue. Both operate at a volume that human teams could not match economically: the platform processed more than 1.6 million calls and 690,000 chats. 

Confirmed savings reached $2.3 million, and the bank projected that the figure would rise to $3.7 million by the end of the year. Those numbers are significant because they rest on a measurable baseline, the earlier cost of $0.35 per interaction, rather than on a general claim of efficiency. 

A language problem that had to be solved first 

One of the more distinctive elements was the work on language. The team developed what it describes as the first Uzbek-language large language model, with speech recognition and speech synthesis. General-purpose models are strong in widely used languages but perform less reliably in Uzbek, particularly in telephone conversations with regional accents, mixed Uzbek and Russian speech and informal vocabulary. A model that understands how customers really speak is a precondition for useful voice assistants. 

The same capability improves quality for text chats. When an assistant recognises spelling variants and colloquial phrases, fewer conversations are escalated to human agents, and fewer customers abandon a request out of frustration. 

What customers actually ask 

Automated assistants spend much of their time on repetitive, factual questions, and money-related queries are among the most common. Customers who deal with several currencies want quick answers about conversion, and the wording of their questions changes with the news cycle. The steady flow of searches such as “funt kursi” illustrates how widely people consult currency information and how closely they follow it in everyday decisions. For TBC Bank Uzbekistan, a well-trained assistant that can respond to such requests at any hour, in the customer’s own language, turns a recurring burden on the contact centre into a smooth part of the service. 

Speed is a second benefit. A customer who writes at midnight receives an answer within seconds rather than the next morning, and a repayment reminder can reach a borrower at the moment that is most convenient for them. Both effects improve satisfaction and, in the case of collections, can reduce the number of accounts that fall into longer arrears. 

The effect on cost follows directly. If a question that once required an agent can be handled automatically, the unit cost falls to cents, and the staff time released can be moved to work that machines do poorly, such as resolving disputes or advising people in difficult financial situations. 

People, governance and the culture behind the numbers 

Technology alone did not deliver the result. The bank created an in-house AI team and a Machine Learning Competence Center, and it introduced an AI-ization programme designed to turn employees into active users of the tools. The aim was cultural as well as technical: an organisation in which colleagues in sales, risk and operations expect to use models in their daily work, rather than leave them to a specialist department. 

The training side deserves attention. Employees who understand what a model can and cannot do ask better questions, spot errors earlier and propose new uses. A bank that treats AI literacy as a core skill, much like spreadsheet literacy a generation ago, tends to find savings in places that a central team would never think to look. 

Governance matters just as much. Monitoring pipelines track model performance, and defined processes decide when a model should be retrained or withdrawn. In regulated industries, a system that cannot be audited is a liability, so building controls into the platform from the start protects both customers and the institution. 

Lessons for banks in emerging markets 

The case carries several lessons. First, measurable targets focus the effort: a drop from $0.35 to $0.03 is easy to verify. Second, local language support is not a detail, since it determines whether automation works at all. Third, building internal skills pays back over time, because each new use case becomes cheaper than the last. 

Sequencing was also sensible. By launching two focused assistants rather than a general tool for every task, the team could measure results precisely, correct errors quickly and build confidence among colleagues before widening the scope. Each completed use case supplied data, code and experience for the next. 

There are limits as well. Models need constant monitoring, customer trust has to be earned through accurate answers, and sensitive conversations still call for people. Yet the 2025 results show that an emerging-market bank can build a serious AI foundation at home, and that doing so can change both its cost structure and the experience of its customers.