Raiffeisenbank: 500 engineers adopted AI, but development speed didn't increase

Raiffeisen Bank's engineers actively use AI tools — chat interfaces and coding agents. On the charts, the activity looks convincing: the number of users is growing, the frequency of requests is increasing. But when the team checked the actual metrics, it turned out: the development speed hadn't changed. Marat Kinyabulatov, an expert in Agile practices and responsible for the efficiency of the bank's engineering teams, analyzed why the growth in usage doesn't translate into acceleration.

The Problem That Isn't Visible on Dashboards

The picture of AI implementation in a large company typically looks promising: active users are growing, engineers are mastering new tools, management receives positive reports. This turns out to be enough to consider the implementation successful.

The problem emerges later — when you start looking at business metrics. Kinyabulatov's team compared indicators before and after implementation: task cycle time, feature release speed, number of iterations before production. 500 engineers mastered the tools, but team productivity remained at the previous level.

According to Kinyabulatov, this picture is typical not only for Raiffeisen Bank. Most companies that are massively implementing AI in development face the same thing: tool usage metrics grow before anything changes in actual work processes.

Why Usage ≠ Acceleration

The team tested several hypotheses about the impact of AI on development speed. The initial logic looked reasonable:

• AI chats reduce time spent searching for documentation and explanations
• Coding agents handle writing template and repetitive code
• Reduced routine gives engineers time for complex tasks

In practice, each of these connections proved to be non-linear. An engineer getting quick answers from a chat could spend the saved time on additional clarifications and reformulations — rather than moving the task forward. A coding agent generated code that needed to be reviewed and refined: sometimes this took more time than writing it independently.

"On the charts, users and activity are growing, engineers are trying

out tools and getting used to the new reality, but when checking metrics, it very often turns out that nothing works faster," writes Kinyabulatov.

Which Metrics Helped Separate Habit from Result

The central challenge was creating a measurement system that shows not just the fact of tool usage, but their impact on results. The team searched for correlations between activity in AI tools and real performance indicators: cycle time, throughput, share of tasks that passed through production without requiring rework.

In parallel, they studied behavioral patterns — what exactly changed in the work of engineers whose productivity did increase. The key difference turned out not to be the frequency of tool usage, but the way they were integrated into the work process.

This analysis led to the concept of Agentic Engineering — an approach where an engineer builds chains of AI agents that handle entire work stages, rather than applying AI as an interactive reference for individual questions.

What This Means

The Raiffeisen Bank case describes a systemic problem of AI transformations: visible metrics of tool usage create an illusion of progress while actual productivity remains unchanged. The real effect comes not from implementing the technology, but from restructuring work processes around it. Kinyabulatov promises a detailed analysis of exactly how the team built Agentic Engineering in the second part.