Raiffeisenbank deployed AI agents to 500 engineers, but speed didn't increase. Why

Raiffeisenbank concluded the second part of an experiment on implementing AI in an engineering organization of 500 specialists: local acceleration of individual roles through coding agents did not change key metrics such as Lead Time, Throughput, and Defect Rate.

Why copilots did not accelerate the team

Distributing coding agents to engineers alone does not change the speed of the system — this is the conclusion of the first part of the experiment that Raiffeisenbank shared in a blog post on Habr. According to the authors, you can increase MAU LLM (monthly active users of the language model), then increase MAU API (number of API calls), distribute coding agents to people — and still not see changes in three engineering metrics: Lead Time (time from idea to result), Throughput (throughput of the development process), and Defect Rate (share of defects in released code). These three metrics are a standard indicator of the health of an engineering process, and their unchanged state after mass AI implementation is what the blog authors call the main conclusion of the first part of the experiment.

• The experiment covers an engineering organization of 500 Raiffeisenbank specialists
• Metrics that did not change after AI implementation: Lead Time, Throughput, Defect Rate
• Local acceleration runs up against requirements, review, testing, approvals, neighboring teams, and business experts
• The company's next step is the transition to Agentic Engineering

Where speed is actually lost

Accelerating one person or one role does not accelerate the entire system — the task still goes through all the remaining stages of the process, explain the authors of the Raiffeisenbank blog. One employee may start doing their part faster, but the result still waits in queue for review, testing, approval, or a neighboring team's decision. The authors describe this as a problem of local optimization: accelerating one link in the conveyor does not change overall throughput if the bottleneck is located elsewhere in the process — with reviewers, testers, or approving departments.

"One person or one role can start doing their part faster, but the task still waits for requirements, review, testing, approvals, neighboring teams, or a business expert," — states the

Raiffeisenbank blog on Habr.

What's next: Agentic Engineering

Raiffeisenbank did not stop at distributing tools to individual engineers and decided to go further — in the direction that the company calls Agentic Engineering. In the industry, this term typically describes the transition from point AI assistants for one developer to agents embedded in the process at the level of teams and stages, not just at individual workplaces. What exactly Raiffeisenbank puts into this concept and how it will solve the problem of approvals and queues between roles, the company promises to reveal in the continuation of the series of articles on Habr — the material is directly called the second part of the story about implementing AI on 500 engineers.

What this means

The Raiffeisenbank case shows: distributing AI tools to engineers one by one does not accelerate development if bottlenecks remain at the interfaces between roles and teams — real effect appears only when acceleration affects the process as a whole, not the work of one person.