Search Without Internet Access
Nornickel has built a hybrid architecture with Yandex Cloud to run generative AI securely within its corporate network, with AI agents supporting more than 30 production and corporate processes.
The system connects to cloud-based large language models through a dedicated physical connection. Access is provided through this channel, and data is not stored by the cloud provider. This allows the company to run models within its corporate network without connecting to the public internet.
At present, this is one of the most advanced industrial generative AI deployments in Russia. A large mining and metals company is embedding agent-based systems and large language models into business processes while keeping control of data and security. According to Nornickel, the system has reduced task times for some processes from 30 days to just a few hours.
Cloud-Based, Yet Controlled
Nornickel and Yandex professionals developed the architecture. The system meets corporate security requirements and improves management processes across both production sites and office operations. Cloud technologies form the basis of the security model. The system connects to Yandex Cloud resources via a dedicated physical connection without storing data with the provider. As a result, AI agents operate within internal corporate networks without internet access. This is the first implementation of this type in Russia under real industrial conditions.
Experts say that real-world deployment conditions and strict requirements for security, scalability, and rollout speed are shaping a standard for using AI agents in industrial environments. Nornickel and Yandex are developing a cross-industry approach to using large language models in the cloud. Major banks already use similar architectures in Russia.
From Platform to Autonomous Development
The company develops advanced agent-based systems based on its domain-specific model MetalGPT. These systems are reshaping design processes, capital project management, inventory management, and financial control. The company integrates cloud-based large language models into existing workflows as digital assistants, primarily in the form of chatbots. Dozens of advanced agent systems are already in testing, along with many simpler personal assistants for employees.
Partners are planning the next stage, expanding access to AI tools for office staff and developing modules that allow users without technical expertise to build simple agents themselves. This marks a shift from AI tools used by a limited group of developers to widespread enterprise adoption.
From Pilot to Industry Standard
The current deployment builds on earlier initiatives. In 2024, Nornickel presented key IT projects at the “Digital Industry of Industrial Russia” conference, showing that this project is part of a broader digital transformation strategy. In December 2025, the company introduced MetalGPT-1, its own language model tailored for mining and metallurgy. The company is moving from in-house domain models toward secure use of external cloud-based LLMs within a protected environment.
Other major industrial players are moving in the same direction. In 2025, RUSAL reported deploying AI in its global aluminum production monitoring system ELTM, where neural networks identify process deviations and suggest root causes. This reflects a similar approach, embedding AI into core production management rather than limiting it to marketing or administrative workflows.
In September 2025, Severstal also launched its own generative AI platform, enabling employees to build AI assistants within a secure corporate environment.
Overall, Russian heavy industry is moving toward a model with protected environments and applied AI assistants for employees. A practical framework for secure use is emerging, tailored to companies with strict data requirements. This marks a shift from pilot projects to a corporate standard for large-scale industrial adoption.
