AI projects often stumble not because the models themselves are weak, but because companies and their employees have different understandings of the same processes. Alexey Mezentsev, head of OTP Bank’s AI Lab, shared this insight at the Youth Day of the ”Game Theory and Management” conference. His blunt takeaway: training an AI system on incomplete or outdated data leads it to confidently make wrong decisions.
Mezentsev explained that automation goes off track when businesses digitize only official rules and protocols. On paper, procedures look neat, but in reality, employees frequently add their own steps-skipping redundant parts, manually double-checking, or handling tasks differently than the documented instructions. If AI overlooks this informal layer, it sees a distorted picture and ends up accelerating the wrong workflows instead of scaling the intended ones.
According to OTP Bank’s top AI executive, successful AI implementation must start not with model selection but with aligning on the actual ”reality”: how processes truly operate, where exceptions occur, and what employees consider a normal outcome versus a failure. When staff input is integrated this way, they become collaborators rather than obstacles to automation, providing crucial context. Without this alignment, companies risk investing in costly AI systems that confidently get things wrong.
Mezentsev also challenged the typical strategic debates around AI project priorities. Instead, he advocates for a down-to-earth approach: test hypotheses with small pilots and immediately measure impact in financial terms. His logic is straightforward-if a hypothesis can’t prove its value quickly in a limited experiment, it’s too early to scale it into a major AI initiative with dedicated teams and big budgets.
This mindset matches current banking industry trends. McKinsey estimates that banking is among the sectors where generative AI could deliver substantial economic impact-provided AI efforts are woven into real business processes rather than isolated pilot projects. Gartner has repeatedly pointed out that many AI initiatives stall between pilot phases and full deployment due to challenges around data quality, governance, and aligning with business goals. In Russia, banks like Sber, Tinkoff Bank, and VTB are already leveraging AI for customer support, credit scoring, fraud detection, and internal operations.
At the GTM 2026 conference, where mathematician Alexey Savvateev and other industry leaders also spoke, Mezentsev added one more key factor for successful AI launch: the tangible benefit of any new system must be perceptible not only to IT departments or managers but to everyone involved-from frontline employees to customers. This is especially critical in banking, where complex processes and strict regulations make errors in the ”reality map” used to train AI all the more costly.
In the coming years, banks that can bridge the gap between data-driven AI models and the messy, human realities of day-to-day workflows will have a clear advantage. The real test is no longer just building smart algorithms but embedding them into organizational contexts that reflect how business truly operates.