A catalog by the ANO Digital Platforms organization now tracks 1,352 AI implementation projects across Russian companies and government agencies. Analysis shows that project success in production depends more on the surrounding infrastructure—such as semantic layers and deterministic code—than on the AI model itself.
Companies like First Cargo Company and Yandex utilize semantic layers and context-gathering tools to improve accuracy. In the case of X5 Group, a CV-based moderation system for 10 million photos per month achieved a 92 percent accuracy rate by combining 26 models with human oversight at confidence thresholds.
Reliability studies indicate that AI agents struggle with long chains of tool calls, with performance typically degrading after six steps. Solutions like GigaChain and the Model Context Protocol (MCP) are being adopted to standardize tool integration, allowing agents to operate with verified contracts rather than raw system access.
Financial results vary by use case. Production-focused AI, such as the system at Nevinnomyssky Azot, generated 270 ₽ million in 2024, while anti-fraud systems at Sber and T-Bank prevented losses of 360 ₽ billion and 170 ₽ million respectively. Corporate LLM platforms currently show broad reach but less direct financial impact.
Engineering teams are focusing on efficiency through techniques like semantic caching, PagedAttention, and FP8 quantization. These methods help reduce token consumption and GPU costs, which are critical for companies managing large-scale AI deployments.
