In the Russian market, the experts examined 22 publicly documented cases. Four of them corresponded to the L4 level, and all were developed internally by large companies: Afanasy Ivanov at T-Bank, Markus at Sber, Alfa-Bank’s virtual developers, and Ruban at Severstal. This market structure does not indicate that Russia is lagging behind. Russian customers are more likely to require on-premises deployment, the use of confidential data, integration with legacy systems, and regulatory compliance. As a result, the platform-based approach is developing faster than off-the-shelf SaaS products.“The share of mature systems is even lower in real-world corporate implementations. Of the 43 international cases, only 10, or 23%, reached the L4–L5 level, meaning they can execute an end-to-end process with minimal human involvement. Most solutions remain at L3: they can perform a sequence of operations and have certain autonomous capabilities, but they still require substantial employee oversight.”
Maxim Bolotskikh, Partner and Head of the AI and Advanced Technology Practice
Agent or assistant?
The report proposes six operational criteria for agentic capabilities: an agentic loop; planning; tool use; self-correction; memory; autonomous task completion. A solution is considered a full-fledged agent if it meets at least five of the six criteria. The authors also use a five-level maturity scale, from L1 to L5, ranging from a large language model and an AI assistant to an autonomous agent and a multi-agent system.Claude Code, Devin, OpenAI Codex, OpenAI Operator, Cursor, Harvey, and Manus met the threshold for a full-fledged agent. Three more products were classified as agents with limitations, while five were categorized as AI assistants. Two solutions were agentic frameworks, and one was a workflow automation tool. According to the authors, the number of products marketed as agents does not, in itself, demonstrate a company’s maturity. What matters more is the degree of autonomy within a specific process and the measurable results achieved.“Today, the word ‘agent’ often describes a product’s marketing positioning rather than its architecture. Executives need to determine whether a system can independently break a goal down into steps, operate within external systems, assess results, correct errors, and complete a task from start to finish. Without these capabilities, a company is buying an assistant while expecting end-to-end automation, leading it to overestimate both the required budget and the potential impact.”
Maxim Bolotskikh, Partner and Head of the AI and Advanced Technology Practice
Widespread adoption — limited impact
The widespread adoption of generative AI has not yet translated into equally widespread financial impact. More than 70% of Russian companies use generative AI in at least one business function, while all 150 of the largest organizations in the sample have conducted at least one pilot involving key AI technologies. However, only 9% of companies reported an impact exceeding 5% of EBITDA. There has nevertheless been positive progress: over the past two years, the share of organizations without a confirmed impact has fallen from 32% to 22%. The experts explain this gap through an “impact attenuation cascade.” A technology may significantly accelerate an individual task, but when it is introduced into an actual project and scaled across a company, the local improvement encounters approvals, testing, integrations, legacy systems, and unchanged performance metrics.“The problem is usually not the model itself, but everything around it. If one stage is accelerated while the rest of the process remains unchanged, the bottleneck simply moves elsewhere. Implementation should therefore begin not with the purchase of a license, but with the selection of one repeatable use case, the appointment of a business owner, the establishment of a baseline metric, and the adjustment of employee KPIs alongside the process.”
Marina Dorokhova, Director in the AI and Advanced Technology Practice
The Russian Market: Platforms Instead of Ready-Made Products
The four most mature publicly available Russian solutions—Yandex AI Studio, Sber’s GigaChat Enterprise, Just AI’s Agent Platform, and MTS’s MWS AI Agents Platform—are platforms for building agents. They meet most of the architectural criteria, but customers must use them to create solutions tailored to their own processes. The remaining products in the Russian sample fall into the categories of assistants, legal support tools, and conversational systems, according to the report.“In Western markets, an AI agent can often be purchased as a ready-made product. In Russia, it generally has to be assembled and integrated into the corporate environment. This is a rational response to requirements related to data, security, and integrations. However, this approach requires an architectural solution based on on-premises deployment, modularity, and the ability to replace the model or vendor without rebuilding the entire system.”
Sergey Kobelev, AI Adviser at Yakov & Partners
