The Bank of Russia’s proposals for the safe use of artificial intelligence (AI) in the financial sector will lead to increased costs for banks, experts say. In addition, implementation of the regulator’s recommendations will require significant improvements to the already created infrastructure. The biggest problem may be the requirement for bank employees to confirm AI decisions in payment transactions every time, which will make the use of this technology pointless. Only in Sberbank, according to its own data, the decision to issue loans to individuals using AI is already made in 100% of cases.
The Bank of Russia has published guidelines for the safe construction and use of artificial intelligence in the financial sector. In particular, banks will need to monitor the reliability and relevance of the input data when training AI, the integrity of the model itself, the transparency and predictability of its operation, and the reliability of the output data. In total, the regulator proposed more than 20 measures aimed at neutralizing current threats to the security of AI technologies.
According to the Central Bank, currently 69% of Russian banks use or plan to use AI. It is most often used in risk management, customer services, analytics and forecasting, anti-fraud, document automation and claims handling. In particular, 11 out of 12 credit reporting institutions use traditional AI in credit scoring and risk management, 9 – for profiling clients, personalizing offers of banking products and services, and in other functions.
According to Sberbank, in 100% of cases the decision to issue loans to individuals is already made using AI. In the first quarter of 2026, 65% of requests from state bank clients in voice and text channels were also processed using AI, every second loan transaction with medium and large businesses takes place without human intervention. According to the head of Sberbank, German Gref, in 2026 alone, costs for technology development will increase to 350 billion rubles, and in total, in 2024–2026, the bank plans to invest 600 billion rubles in AI. Since 2025, MTS Bank has been implementing a program that involves the introduction of AI tools into all key processes.
At the same time, banks are very loyal to the recommendations of the Central Bank. Sberbank said that it had built an AI cybersecurity management system in advance, within which it developed and published its own threat model. The State Bank notes that the Central Bank’s recommendations will not create new obligations, since they are already working “according to the principles that are now being formalized for the entire industry.” Denis Popov, managing expert of the PSB Analytics and Expertise Center, believes that the use of AI “requires the formation of a system of rules aimed at increasing digital security and hygiene and meeting all requirements for the protection of bank secrecy.”
Experts are more conservative. According to Artem Brudanin, head of cybersecurity at RTM Group, “the recommendations cover literally everything: from “poisoning” training samples to clearing incoming requests from injections and protecting against popular attacks.”
In his opinion, if a bank cannot guarantee the explainability of its model’s decisions or protect it from threats, “it will now simply not be allowed into the critical infrastructure circuit.”
Experts believe that the most difficult thing for banks will be to implement the Central Bank’s recommendation to confirm the AI decision by a bank employee “in critical processes with high information security risks, in particular in payment transactions.” “This is a direct cooling of hotheads who dreamed of completely replacing employees with algorithms for the sake of quick savings,” states Artem Brudanin. MBA professor of business practice in digital finance at RANEPA Alexey Voylukov explains that if the entire flow of payment orders through the bank is confirmed by an employee, then “this will make the use of AI pointless.”
As a result, implementation of the regulator’s recommendations will lead to a significant increase in bank costs. Timur Aitov, Chairman of the Commission on Financial Security of the Council of the Russian Chamber of Commerce and Industry, explains that the main expenses are “refining data control systems, strengthening monitoring of models and introducing human participation in critical operations.” According to Artem Brudanin, budgets will have to be set aside not just for the purchase of servers or licenses, but for the construction of a full-fledged MLSecOps (Machine Learning Security Operations – a set of measures aimed at ensuring the security of AI).
According to the expert, for large players this means a complete restructuring of development pipelines, the purchase of FSTEC-certified security equipment for supply chains, and the hiring of scarce specialists at the intersection of Data Science and information security. “In practice, creating systems for monitoring data drift and conducting regular checks of the Red Team format for models will result in millions of expenses,” Mr. Brudanin is sure. Moreover, as Alexey Voylukov points out, “major players have already created their own threat models and are taking measures to stop them,” but taking into account the fact that everyone does this in their own way, “it cannot be ruled out that they will have to rebuild their systems.” Bankers also expect an increase in costs, however, according to Denis Popov, they “should not be perceived by the banking business as an additional burden.”
