AI in the codebase: how VTB and T-Bank rolled out coding assistants

At the Conversations conference, developers from VTB, T-Bank, Dodo Engineering and S7 TechLab talked about bringing AI into their development work through coding assistants, and about how team roles have changed over a year of use.

What has changed after a year of AI in the codebase?

Roughly a year ago, major Russian IT teams began rolling out AI coding assistants at scale. Developer sentiment went through familiar stages: enthusiasm at the start, disillusionment once they ran into hallucinations and poor-quality code, and finally a calm acceptance of the tool as part of the daily routine.

The key question today isn't "does AI work at all," but "has it changed anything measurable in the team's metrics." According to the panelists, the productivity gains are real, but they show up in a different place than expected. AI doesn't speed up building new features. Instead, it cuts the effort spent on routine work: tests, documentation and refactoring boilerplate code.

• Participating companies: VTB, T-Bank, Dodo Engineering, S7 TechLab
• Organizer of the Conversations discussion: Just AI
• Main effect of AI: savings on routine tasks, not faster development of new features
• The tools have become part of the daily workflow for most teams

Should developers be forced to use AI?

One of the most heated questions in the discussion was whether developers should be required to use AI assistants. The panelists' experience showed that a top-down mandate works worse than organic adoption. VTB and T-Bank relied on internal advocates: developers who showed their colleagues real cases where the tools saved time. When AI takes root through personal example, resistance drops noticeably.

S7 TechLab and Dodo Engineering noted that pushback most often arises when the assistant generates code that takes a long time to review. If the ratio of "code written by AI" to "code accepted without significant edits" is too low, trust in the tool erodes, and developers go back to their old methods.

"AI doesn't automatically speed up development — it redistributes

where the time goes."

Who writes the code now, and who only reviews it?

Here the panel identified an unexpected shift in roles. On teams that actively use an AI assistant, the developer's job is moving from writing code to reviewing and validating it. That changes what's expected of junior and mid-level engineers: the ability to quickly read, understand and evaluate someone else's code, including generated code, now matters more than the ability to write it quickly from scratch.

The participants identified two real risks. In the first, a developer accepts code "blindly" and piles up technical debt. In the second, they spend so much time understanding the generated code that the entire gain from AI disappears. Both scenarios have already occurred in production teams.

What this means

AI coding assistants have become part of the production cycle at major Russian IT teams. Along with productivity gains, though, they have raised new questions about role structure, code review quality and tech debt management. The main takeaway from the discussion: AI changes processes rather than simply speeding them up.

FAQ

Which companies took part in the discussion?

Representatives of VTB, T-Bank, Dodo Engineering and S7 TechLab took part in the Conversations technical discussion organized by Just AI. All four companies shared their experience of bringing AI assistants into production development.

Where do AI assistants actually boost productivity?

According to the participants, the gains show up mainly in routine tasks: writing tests, documentation and refactoring boilerplate code. They don't show up in how fast new features get built.

What stages did AI assistant adoption go through at

Russian companies? Based on the experience of VTB, T-Bank and other large companies, coding assistant adoption went through three stages: enthusiasm at the start, disillusionment once teams ran into hallucinations and poor-quality code, and then a calm integration of the tool into the workflow.