Sber, Yandex and red_mad_robot show how AI is changing the developer's role

AI is taking over more and more of the routine part of software development, and the engineer's role is shifting from writing code by hand to defining tasks and controlling the results. The examples of Sber, Yandex, T-Technologies and red_mad_robot show that this shift already affects not only release speed but also the requirements placed on developers themselves.

How development is changing

Sber describes its move toward AI PDLC through a maturity scale from level zero to level five. The company is currently at the supervised automation level: AI is already built into most development stages, but the final decision remains with a human. According to the company, around 14 thousand developers use GigaCode, and almost 80% of them use the tool daily. The share of AI-generated code accepted by teams grew from 45% at the start of 2025 to 69% by the end of the year. This is no longer a point assistant but a new working pattern.

The next step is rebuilding the environment itself. In November 2025, Sber moved its developers from JetBrains IDE to its own GigaIDE PRO, which runs seven AI agents: from documentation and logging to automated tests and analytics. More than half of their suggestions are accepted. The economics of hiring are changing too: in 2024 a new engineer needed 71 days to reach full productivity, now it takes 36 days. In this logic, people spend less time on mechanics and more on steering the process.

"AI does 90% of the implementation, while 90% of managing the intent

stays with people."

What happens to the people

The technological gain does not cancel out the human costs. Sber's researchers note that juniors increasingly find themselves in an odd role: rather than learning to write code, they need to evaluate output produced by a model, even though they still lack a foundation of their own. More experienced engineers face a different problem: part of the satisfaction of the job used to come from solving hard problems, and with agents comes a feeling that you are no longer the author but a conveyor-belt inspector. Hence the anxiety about skills, status and one's own value on the market.

At the production level this is even more noticeable. In a large corporation you cannot simply switch on a fashionable agent mode and wait for a miracle: the system must understand internal APIs, security policies and regulatory constraints. That is why "vibe coding" hardly works in its pure form for large companies. The GigaCode case is telling: in December it completed a full release of a tool for uploading and checking libraries without a single manually written line of code. The release included five major tasks and several fixes, and the team accepted the result without serious comments. This changes the engineer's specialization: less manual assembly, more task definition, review and architectural responsibility.

How the impact is measured

Yandex and T-Technologies show that measuring impact by lines of code alone is already meaningless. At Yandex, about 70% of developers regularly use AI assistants when writing code and make 10–20% more commits on average. But code generation is only part of the job: a lot of time goes to searching for information, design, debugging and review. The internal AI Chat noticeably reduced trips to the wiki, and DeepAgent, by the company's estimate, delivers a tenfold speedup on complex codebase research tasks.

T-Technologies, for its part, looks not at the volume of text but at Developer Experience.

• time from first commit to deployment
• share of focus time and number of context switches
• speed of the first review
• share of flaky tests
• time it takes a new engineer to get up to speed

According to T-Technologies' internal telemetry, the share of regular AI users grew from 17% to 85% in ten months, and four weeks after first trying the assistant, 80% keep using it in the IDE and 75% in the web interface.

red_mad_robot went even further: the company built prototypes for web, iOS and Android in 48 hours, produced about 80 thousand lines of code and 208 commits, and the team's role effectively came down to a single AI engineer who orchestrated agents. At a model cost of about 27 dollars per prototype, this is not just faster delivery but a new product economy.

What this means

The developer is not disappearing, but the work is quickly moving up a level: from writing code to formulating intent, verifying results and managing agents. For companies, the main risk is no longer that AI writes too much, but that the business manages to speed up releases without managing to rebuild training, metrics and engineering culture for the new reality.