Russian businesses froze 90% of generative AI projects and failed to bring them to production

Large Russian companies have brought only a small portion of their generative AI pilots to full-scale implementation by March 2026. Most initiatives with LLMs, chatbots, and agents never left testing mode, were rebuilt for narrower tasks, or were shut down entirely.

Why pilots stalled

According to an assessment by the consulting company "Intelligent Analytics," only 7–10% of pilot projects that large businesses launched in 2025 reached industrial production. The sample included about 50 companies from IT, manufacturing, finance, public sector, transport, and logistics. The remaining initiatives by March 2026 were stuck in pilot phase, radically rebuilt, or halted.

This shows how difficult it is to translate beautiful demos with large language models, chatbots, and AI agents into a working corporate product. At the same time, market participants do not consider such conversion to be something sensational. Some projects are still in development and theoretically could reach production later.

But the gap between the number of launches and actual implementations is already noticeable: in 2025, companies often pursued not the most useful scenarios, but those that gave a quick wow-effect for presentations and PR. When the integration phase, ROI measurement, and accountability for results arrived, many initiatives began to stall.

Where the economics break

About 30–40% of pilots, according to survey participants, were closed because they did not deliver the expected financial effect. The main problem turned out to be not the generative AI idea itself, but weak integration with the company's everyday processes. In many cases, models were not embedded in CRM, ERP, document management, and other corporate systems, so they remained a separate showcase rather than a working tool. As soon as the business tried to calculate time savings, reduced employee workload, or impact on revenue, the beautiful concept quickly lost credibility.

• No deep integration with CRM, ERP, and internal systems
• Projects were launched for PR effect rather than for a specific business task
• There was insufficient quality and representative data for fine-tuning
• Models lacked multimodality, maturity, and security requirements

Technical failures are also evident. In one case, a company independently fine-tuned a Chinese model for legal department tasks, but collected too little data, and the assistant's accuracy did not exceed 30%, after which the project was shut down. In another case, the support service wanted to automate document and image processing, but as of August 2025, available models did not support the required multimodal scenario in full. In other words, the problem was often not AI hype, but unrealistic expectations for specific tools.

Timelines shift to the right

Half of the surveyed executives saw their transition timelines from pilot to industrial production shift from 2025 and early 2026 to the second half or end of 2026. The reasons are quite pragmatic: companies had to spend more time training employees, modernizing infrastructure, and resolving information security and data protection issues. Meanwhile, spending on experiments is already substantial.

In 2024, large and medium-sized Russian organizations spent 90.3 billion rubles on AI implementation and use, and a typical pilot budget in 2025, not including infrastructure, was estimated at 5–15 million rubles. An additional complication is that agent solutions quickly expose old problems in the processes themselves.

If intelligent search through a knowledge base can be implemented locally, a more complex AI agent almost immediately runs into unplanned integrations, informal rules, and manual workarounds that employees have long grown accustomed to. According to Cloud.ru's assessment, the culture of familiarity with AI tools in companies can already reach 80–90%, but actual integration into business processes remains at the 5–10% level.

The most growth potential is seen in legal services, consulting, manufacturing, medicine, public sector, and education.

What it means

The Russian market is not abandoning generative AI, but is moving out of the hype phase and transitioning into a phase of strict testing for usefulness and ROI. The next wave of implementations will likely focus not on loud showcases, but on narrow, well-measurable scenarios with clear integration, security, and a clear owner of results within the business.