AI Adoption Didn't Speed Up Development: Raiffeisenbank's Experience

The AI adoption paradox: at Raiffeisenbank, 500 engineers actively use AI tools, both chats and coding agents, yet development speed has not changed.

The problem dashboards don't show

AI adoption in a large company usually looks encouraging: active users grow, engineers learn new tools, and management receives positive reports. That is often enough to call the rollout a success.

The problem surfaces later, when people start looking at business metrics. Kinyabulatov's team compared figures before and after adoption: task cycle time, feature delivery speed, and the number of iterations before production. The 500 engineers had learned the tools, but team productivity stayed at its previous level.

According to Kinyabulatov, this pattern is not unique to Raiffeisenbank. Most companies rolling out AI in development at scale run into the same thing: tool usage metrics rise before anything changes in real workflows.

Why usage ≠ speedup

The team tested several hypotheses about how AI affects development speed. The initial logic seemed reasonable:

• AI chats cut the time spent searching for documentation and explanations
• Coding agents take over writing boilerplate and repetitive code
• Less routine work gives engineers time for complex tasks

In practice, each of these links turned out to be nonlinear. An engineer getting fast answers from a chat could spend the saved time on extra clarifications and rephrasing rather than moving the task forward. A coding agent generated code that had to be checked and reworked, which sometimes took longer than writing it by hand.

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

the tools and getting used to the new reality, but when you check the metrics, it very often turns out that nothing has actually become faster," Kinyabulatov writes.

Which metrics separated habit from results

The central challenge was building a measurement system that shows not the fact that tools are used, but their effect on outcomes. The team looked for correlations between activity in AI tools and real performance indicators: cycle time, throughput, and the share of tasks that went through production without being sent back for rework.

In parallel, they studied behavioral patterns: what exactly was different in the work of engineers whose productivity did grow. The key difference turned out to be not how often they used the tools, but how they integrated them into their workflow.

This analysis led to the concept of Agentic Engineering, an approach in which the engineer builds chains of AI agents that take over entire stages of work, rather than using AI as an interactive reference for individual questions.

What this means

The Raiffeisenbank case describes a systemic problem with AI transformations: visible tool usage metrics create an illusion of progress while real productivity stays the same. The real effect comes not from adopting the technology but from rebuilding workflows around it. Kinyabulatov promises a detailed account of how the team built Agentic Engineering in the second part.

Why didn't AI adoption speed up development?

The number of users and the frequency of requests are growing, but real metrics show that development speed has not changed.

Why doesn't AI adoption deliver faster development?

According to Raiffeisenbank's experience, a growing number of AI tool users and more frequent requests do not guarantee faster development. Marat Kinyabulatov, who is responsible for the effectiveness of the bank's engineering teams, examined why AI activity does not convert into faster real work.