AI-SARA: AI-Powered Intelligent IT Support System for Employees

The project aims to transform IT support into an intelligent self-service model available to employees 24/7. AI-Sara provides fast and consistent answers to common IT-related questions in Russian and Kazakh using verified corporate knowledge and historical support data. The key objectives are to reduce repetitive manual consultations, shorten response time, improve user experience, increase reuse of corporate knowledge and scale IT support without proportional staff growth.

Today, AI-Sara already covers around 28% of IT consultation requests previously handled by the support team and serves as a foundation for further development of AI-driven IT Service Management.

•                      On-Premise LLM: Qwen3.5-397B-A17B (FP8)

•                      Corporate GPU infrastructure

•                      Corporate knowledge base with 2,267 verified Q&A items

•                      Historical dataset of 189,338 closed IT requests

•                      Retrieval-Augmented Generation (RAG) and knowledge retrieval mechanisms

•                      AD / Kerberos authentication

•                      K8s containerized services

The main challenge was to create a production-ready AI service suitable for a large financial organization.

The project required preparation and verification of corporate knowledge, use of a large historical dataset, response-quality control, Russian and Kazakh language support, and integration with corporate authentication.

Information security was a key requirement. The AI model had to operate fully within the Bank’s infrastructure, so it was deployed On-Premise on corporate GPU resources without external internet requests.

The team also had to address growing user demand, performance, monitoring, feedback collection, quality control and future scalability.

AI-Sara was created to reduce the volume of repetitive IT consultations and make support faster and more convenient for employees.

The solution uses verified corporate knowledge and historical IT support experience to answer user questions in natural language. Its knowledge base includes 2,267 verified Q&A items and data from 189,338 closed IT requests.

Since 25 May 2026, AI-Sara has been operating in production across the Bank’s corporate network.

During the first six months of operation:

·       8,206 employees registered in the service, 60% of the Bank’s employee;

·       average MAU reached 2,051 users;

·       20,431 dialogues were conducted;

·       31,105 messages were processed;

·       3,317 user ratings were collected;

·       2508 ratings were positive;

·       user satisfaction reached 81%;

·       average response time was less than 5 seconds.

According to project calculations, AI-Sara saved around 4,137 hours of support-team time and approximately 6,299 hours of employee time. The time required to receive an answer was reduced by around 90 times compared with a traditional support consultation.

IT consultation requests currently account for approximately 28% of the total support volume. Since its launch in May 2026, AI-Sara has independently resolved 21% of these consultation requests.

AI-Sara operates entirely within the Bank’s infrastructure. The LLM runs on corporate GPUs, data is not transferred outside the Bank, and no external internet requests are used.

The solution also includes analytics and monitoring tools that track usage, user feedback, response quality and activity patterns.

The architecture was designed for further scaling, including new knowledge areas, additional corporate data sources and new automation scenarios.

Republic of Kazakhstan. AI-Sara is available to employees of the Head Office, branches and regional units across the Bank’s corporate network throughout Kazakhstan.