PSB Showed How It Implements AI in Banking: Chatbots, RAG and Business Services

PSB has revealed not just a set of AI experiments, but a fairly rigid and pragmatic scheme for implementing AI within the banking ecosystem. The bank emphasizes that it did not aim to be among the first to test every LLM tool, but instead consistently built solutions where business impact could be seen quickly: in serving small and medium-sized businesses, automating consultations, and reducing employee workload. As a result, PSB already operates a combination of scenario-based bots, a generative assistant, RAG-based search across its knowledge base, and pilots for internal teams.

In external services, the main focus is on entrepreneurs. PSB uses recommendation models in its CRM to analyze transactions and suggest more suitable tariffs, products, and reasons to contact customers. The logic is simple: AI should not entertain, but save money and time.

For example, the system can notice that a customer regularly overpays for operations beyond their limit and would benefit from switching to a different tariff. Or suggest the opportunity to transfer a counterparty to PSB if it reduces payment costs for both parties. According to the bank, approximately 98% of services for entrepreneurs are now available remotely, and 74% undergo end-to-end automation without human intervention at intermediate stages.

The key product here is the AI assistant "Katyusha," available in the internet banking app, mobile app, website, and VK. It is used not only as a chatbot for answering questions, but also as an applied tool for routine business tasks: preparing sales texts, creating marketplace product descriptions, responding to reviews, writing social media posts, and even formulating interview questions for candidates. The bank reports approximately four thousand requests per quarter and emphasizes that the service remains free for SMB segment clients.

Under the hood, a Russian LLM operates, which PSB began testing in an early pilot. The idea here is to gather the most frequent scenarios in one interface rather than sending clients to various external services. The next step is implementing RAG into Katyusha's consultation workflow.

According to PSB, this layer delivered the most noticeable quality jump: in the pilot, automation of requests increased by 7%, average issue resolution time was reduced eight-fold, and customer satisfaction in the chatbot increased by 21%. At the same time, the share of incorrect answers was reduced from 3% to 1%. This is an important result for the bank, because at high levels of automation, each additional percentage point becomes much harder to achieve than at the start.

A separate problem turned out to be the ethical filter of the LLM: in tests, it blocked almost 30% of requests due to phrasing in the knowledge base. After rewriting texts, replacing sensitive words, and adding post-processing, the share of such blocks was reduced to 1%, which made the pilot viable. A separate direction is the bank in messengers and internal AI for employees.

Since 2023, Katyusha has operated in VK not only as a consultant but also as an interface for basic RCC operations: issuing invoices, making payments, viewing statements and balances. In 2025, the number of active users of this channel grew by 35%, and over 60 thousand payments passed through it. Following this, PSB launched payment services in MAX as well.

Inside the bank, the pilot for employees started in October 2025: initially with 25 people who need to quickly respond to inquiries from sales offices and field managers. The goal was to free up about 10% of work time, but during the pilot it became clear that AI provides the most value not on typical questions, but in narrower scenarios without personal data, where accurate consultations and product guidance are important. From all of this emerges a rather clear strategy.

PSB is not trying to build a universal AI bank in one step, but distributes tasks by risk and benefit levels: cloud models for content and non-personalized scenarios, scenario-based bots for quick simple answers, RAG for scalable consultations, and LLM within the workflow for sensitive data and internal services. In 2026, the bank plans to deploy LLM within its perimeter and increase the share of RAG-based answers in the overall consultation flow from 4% to 16%. For the market, this is an exemplary case of how a major bank implements generative AI not for show, but as a working tool with clear metrics.