AI in development: vibe coding instead of developers

Bitrix24 Product Lead Katya Obraztsova published a detailed case study on Habr AI in July 2026: a team of five people without dedicated developers built a multi-user service from scratch in three months of evenings and weekends for an internal corporate marathon — using AI assistants as the primary development engine.

What exactly they built

The task looked straightforward: an internal tool was needed for Bitrix24's annual sports marathon — employees log physical activity, see colleagues' results, and compete in teams. No ready-made solutions for this specific format existed, so the team decided to build a product from scratch and chose vibecoding as the primary development tool.

• Team — 5 people, without a dedicated backend developer
• Timeline — approximately 3 months, work only during non-work time
• Approach — vibecoding: AI assistants as the primary code-writing tool
• Result — a working multi-user service for internal users

Obraztsova immediately warns: no theory about microservices, no pretty architectural diagrams — only an honest breakdown of the pitfalls that the team encounters when product managers tackle a genuinely technical project.

Why quick wins end

The first weeks deliver the classic vibecoding effect: code appears quickly, features come together one after another. The AI assistant writes components, handles errors, turns a task into working code in seconds. Product managers feel like developers — and it's a genuine 'wow moment' at the start.

But over time, the pace slows down. AI tools maintain context well for individual components but start to struggle when the system grows to dozens of interconnected modules. Inconsistent dependencies, logic duplication, hard-to-diagnose errors at layer boundaries — all of this requires something AI doesn't do for you: architectural thinking.

'There won't be advice about microservices architecture and pretty diagrams here.

There will be an honest story about what happens when product managers take on a complex technical project, what pitfalls they encounter, and what ultimately results,' writes Obraztsova.

How vibecoding becomes an AI-driven team

The next stage is conscious reorganization. The team moves from chaotic prompting to structured work: each participant specializes in a specific part of the product and knows how to properly formulate tasks for AI specifically for their domain.

This shift changes the required competencies. A product lead in an AI-driven team spends more time decomposing tasks and setting proper context for AI than writing code manually. Prompt design and architectural planning become equally important skills — and it's at their intersection that true team productivity emerges.

In the end, the five arrived at a model where AI handles routine code writing, while people focus on what AI does poorly: making architectural decisions, managing technical debt, and synchronizing between modules.

What this means

The Bitrix24 case documents a maturing pattern among AI-oriented teams: vibecoding genuinely lowers the entry barrier for non-technical specialists and provides a fast start — but doesn't eliminate the need for systems thinking. Beyond the prototype stage, a product doesn't advance without an architectural approach. Companies that realize this earlier avoid costly refactoring. AI accelerates execution, but doesn't replace design.