Yandex at AI Dev Day Showed How AI Is Already Changing Development at Avito, Ozon, and T-Bank

Yandex held an AI Dev Day and essentially showed that AI in development has already moved beyond demo mode. Companies like Avito, Ozon, T-Bank, Sber, and Yandex Go discussed not hype, but metrics, limitations, and places where agents are already saving hours.

From Assistants to Agents

The main shift heard in almost every presentation is that the market is moving away from simple IDE suggestions to agents that can change code themselves, write tests, search for internal information, do reviews, and even handle incidents. Avito specifically emphasized that fine-tuning open-source models for the company doesn't always give better results: often what matters more is giving a strong model good context, access to documentation, and tools via MCP. Other teams are making a similar bet: not on a "magic model," but on a working combination of agent + context + infrastructure.

At Yandex, this already looks like a system platform. Around 57% of engineers use the internal assistant, and in backend and frontend the share reaches 60–75%. More than 90% of internal infrastructure is already covered by MCP servers, and among active users, 23% of code is generated automatically in agent mode.

Meanwhile, the focus is shifting from simple adoption metrics to a tougher question: where exactly can an agent work almost autonomously, with humans intervening only in complex cases.

Where It's Already Working

The most useful part of AI Dev Day wasn't talk about the future, but numbers from production. Companies now measure not just assistant user count, but time to release, review speed, test quality, and incident costs. Essentially, AI started being evaluated as a regular engineering tool: if it doesn't move business metrics, a wow factor alone isn't enough.

• At Yandex, developers using AI commit about 10% more, and in Go, Python, and JavaScript up to 20%; individual technical tasks for agents dropped from 20 to 2 minutes.
• At T-Bank, median merge time fell by 12%, and for ambassador teams lead time dropped 30% over a year; unit test generation increased four times.
• At Ozon, around 1100 people use the agent assistant daily, about a quarter or third of all development, and auto-review in GitLab turned out to be so demanded that the rollout had to be limited.
• At Sber, the AI system for designers cut design review from an hour to two minutes, and creation of a new screen from 16 hours to five minutes.
• At Yandex Go, AI already helps process about 400 incidents per day and saves about 30 minutes on postmortem for each case.

It's also important that the spectrum of tasks is expanding rapidly. It's not just about code generation anymore, but also searching internal knowledge, automatic checklists, supporting analysts, design review, and SRE scenarios. If a year ago AI often looked like smart autocomplete, now they're trying to embed it across the entire SDLC—from task formulation to analyzing consequences in production.

Why Euphoria Is Limited

Despite strong cases, speakers were quite sober about limitations. Avito directly said that they haven't yet seen noticeable acceleration of the entire development cycle: at most 4–5% in individual teams. The reason is simple—code writing is only part of work time, and bottlenecks lie in review, coordination, testing, and delivery of changes.

So bulk code generation by itself doesn't solve the problem: it may just move the bottleneck further along the process. Another pain point is quality and control. Models can "cut corners" with tests, fake tool calls, produce bland template design, or make mistakes in internal domain languages.

At Yandex Go, for example, for SRE scenarios they eventually switched to English prompts because results in Russian were worse. Plus, truly useful agents need an expensive foundation: observability, service dependency graph, change audit, service catalogs, and proper scenario metrics, not just nice demos.

"There won't be an end of the world.

There will be harsh, but fascinating evolution."

That's exactly why almost all presentations boiled down to one idea: human-in-the-loop hasn't gone anywhere. Ozon is cautious about external models due to code leakage risks. Sber had to separately battle conservative design and hallucinations. Yandex measures not just benefits but side effects like supporting generated code. AI here is no longer a toy, but not an autonomous employee either—more of an accelerator that requires good rules, tools, and constant validation.

What This Means

After AI Dev Day, it's hard to say that corporate AI for development is just an experiment. The experiment is over: now there's competition over who has better context, metrics, processes, and infrastructure. AI doesn't cancel out developers, but noticeably changes the price of routine, the role of review, and demands on those who can manage agents, not just write code by hand.