The company is implementing a 90-day AI-native development pilot. The process is divided into three stages: the first 30 days focus on measuring and limiting tasks, the next 30 days involve building a corporate engineering harness, and the final 30 days test the agentic cycle in a sandbox environment.
The pilot is conducted with a single experimental team selected based on high qualification, task complexity, and motivation. An external AI tracker acts as an independent mentor to monitor progress and apply global AI-native development practices without being tied to any specific tool provider.
To ensure security and compliance, the system uses a corporate gateway for model access, which masks sensitive data and allows model switching without rebuilding the process. The architecture includes a separate data layer and guardrails for classification, masking, and blocking prohibited transfers, specifically considering Russian law 152-FZ for personal data localization.
The approach emphasizes that simply providing AI licenses is insufficient. The company refers to METR research from 2025 and 2026, where experienced developers initially worked 19% slower with AI tools, despite believing they were 20% faster. The goal is to move beyond 'expensive autocomplete' to a system where requirements are machine-readable and changes are automatically verified.
