T-Bank Researchers Develop Training Methods That Could Make Enterprise AI Cheaper
Researchers at T-Bank’s Artificial Intelligence Center have unveiled two new methods for training AI models. The methods are expected to help Russian businesses cut their spending on enterprise AI.
The new training methods are designed to let companies run neural networks locally, directly on their own servers. That matters because confidential information stays within the company instead of being sent to outside AI providers. This is particularly important for banks, insurance companies, healthcare providers and government agencies, where data protection requirements are especially strict. The researchers presented the methods at EMNLP 2026, an international conference on natural language processing and artificial intelligence, in Budapest.
200 Services in One Model
Today, some companies cannot send corporate data to external AI models because of stringent security requirements, so they run models on their own infrastructure. This can create problems: deploying new models without shutting down existing ones puts additional strain on computing resources that are already limited.
Researchers are looking for ways to address these challenges. One of the new training methods developed by T-Bank’s Artificial Intelligence Center is designed to lower the cost of processing requests to language models; the other aims to reduce the expense of extracting information from documents. A model built on the compact Qwen3-32B and trained using the first method brings together more than 200 of the bank’s internal services. It processes requests at 2.8 to 3.9 times lower cost than a larger model. In testing, the researchers also found that the new model delivered better answer quality than its larger counterpart.
A Prompt to the Rescue
The second method, designed for vision-language models, helps tackle the high cost of processing large volumes of documents and extracting the information they contain. It lets users retrieve information through prompts, using natural-language queries. The AI can readily locate details such as account information, dates and amounts. According to T-Bank’s research, a model built on Qwen3.5-35B-A3B-Base and trained with this method outperforms other available models at understanding both one-page and multipage documents and extracting data from them.
The test results show that using the model cuts document-processing costs by more than 80% compared with manual processing and by more than 50% compared with the best open model included in the tests. The system can then handle the extracted data automatically, send it to an employee for confirmation, or route it for manual processing, depending on the quality of the information it finds and extracts.
Both AI training methods are already in active use at T-Bank. They demonstrate that a compact language model can be fine-tuned for real-world enterprise tasks. Over time, it could replace more expensive solutions without sacrificing performance.
Taking Customer Service to the Next Level
Developing and deploying new methods for training AI models is now a priority for Russia’s IT industry. These methods can help companies cut costs without investing in their own AI infrastructure. For businesses, another important advantage is that the models can process information without sending it to external AI providers.
In the future, lower AI costs could accelerate the digital transformation of Russian businesses and spur the development of new software products. These technologies could also make customer service faster and more convenient at banks, insurance companies and government agencies.
T-Bank researchers’ technologies could also become sought-after exports. The high cost of enterprise AI is a challenge worldwide, and businesses everywhere are looking for ways to bring those costs down. The most promising markets for Russian developers include the Commonwealth of Independent States, the Middle East and Asia, provided the technologies are adapted to local languages and legal requirements.
