Expert Interview | Jimsher Chelidze: "The State Can Remove Barriers. The Journey Must Be Undertaken by the Companies Themselves"

We continue to publish an extensive cycle of interviews with Jimsher Chelidze, an expert in industrial artificial intelligence and a member of the Industrial Committee of the Financial and Business Association of Euro-Asian Cooperation. In the first part of the conversation, he explained why the race for frontier models is not our path, and that the competitive advantage lies on the shop floor: in data, engineering, and access to real-world assets. In the second part, Chelidze discussed how the role of the engineer is changing in the era of AI assistants, and why companies risk losing a generation of expertise by winning three years of productivity. In the third part, the expert dissected the main barriers preventing industrial enterprises from moving from isolated pilot projects to the systematic implementation of AI—ranging from the lack of a strategy to personnel resistance and security policies.

In the fourth and final part, the focus is on the role of the state: what has already been done, where support mechanisms are failing, and what must change in regulation, infrastructure, and financing approaches.

The answers are published unabridged.

Jimsher Chelidze is the General Director of Chelidze and Partners LLC, Business Partner for Digital Development at Horizontal Drilling Center LLC, and a member of the Industrial Committee of the Financial and Business Association of Euro-Asian Cooperation. He is the author of articles on the philosophy of technology and five books on digital transformation, as well as the creator of two AI products. He has practical experience working with Gazprom Neft, LUKOIL, the Ministry of Energy of Russia, Gazprom Burenie, and other industrial companies in Russia, Kazakhstan, and China.

In May 2026, at CIPR, the Prime Minister personally reviewed digital solutions in industry. How do you assess the role of the state, and what should change in regulation and support mechanisms?

Let me start with what has actually been done—because it is unfair to criticize without acknowledging the achievements.

First, the state has created demand. Industry would not have generated this demand on its own. Through the industrial competence centers, 175 projects were launched by May 2026, 120 of which have been completed; out of 139 particularly significant projects, 105 have been implemented—that is 76%; solutions have been deployed at more than 930 enterprises. Without the state framework, this volume of work would not have materialized.

Second, the legitimization of the topic. The presence of the head of government at an industry conference is, surprisingly, a practical tool for any head of digital transformation. It signals to the board of directors that the topic is not optional. It eliminates half the debate of "why do we even need this."

Third, the framework. The Law "On Support for the Development of Artificial Intelligence Technologies," adopted by the State Duma on July 8, enters into force on September 1, 2026; key provisions take effect on March 1, 2027. It is a framework, risk-oriented law and—crucially—regulates only large foundational models, from one billion parameters upwards. Standard corporate AI: scoring, computer vision on the production line, predictive maintenance—does not fall under its scope.

My assessment: the framework nature here is an advantage, not a flaw. The worst thing that could have been done is to regulate the industry in detail before practices emerge. We would have ended up with a European scenario: detailed requirements and no products to apply them to.

And two things that industry reviews missed. First: the law directly legalized the path through open weights—a model assembled on open licenses, fine-tuned by a Russian legal entity, and deployed in a Russian data center receives national status. This is sensible and represents an opportunity for industry teams. Second, the banking sector is the only industry explicitly named in the text of the law. Cases where only domestic models are permissible, and the requirements for the risks of their application, are established by the Government in coordination with the Bank of Russia. For the financial market, this is the main provision of the law—not the labeling that everyone is writing about.

Now, where the mechanisms are not working. Four systemic deficits

That very figure: about 187 billion rubles in costs for 2022–2025 versus roughly 1.6 billion rubles in revenue for developers from the implementation and scaling of the created products. We have built solutions—but we have not built products and product companies.

What specifically needs to change: shift part of the support from the "create" stage to the "sell to the tenth client" stage. Subsidize not development, but implementation for the second and subsequent customers. Conditionally: pay for the fifth, tenth, and twentieth implementation, not the first. This is one line in the support rules—and it instantly changes the developer's behavior: they start investing in documentation, implementation methodology, and a partner network. That is exactly what makes a product export-ready.

Industry estimates under discussion: the gap with the US in available compute is hundreds of times over, around 10,000 accelerators across all Russian AI data centers, parallel import supplies have dropped from thousands of servers to dozens. At the same time, the limiting factor has become connected electrical capacity.

For industrial AI, frontier models, I repeat, are not critical. But fine-tuning industry models, running models in the loop, digital twins—these require compute. And this is precisely the kind of infrastructure the market will not build on its own: it is capital-intensive, has a long payback period, and is by nature a public good. Here, the role of the state is primary—just as it was during electrification.

Specifically, I would do three things: priority grid connection for AI data centers; locating them in nodes with surplus generation—we have regions with a power surplus, and this is a rare natural advantage; and—fundamentally—guaranteed access to compute for medium-sized businesses, universities, and industry consortia, not just for three or four ecosystems. Otherwise, we will get an oligopoly, and all industry solutions will be born only where there is an in-house data center. For industry, this is a bad scenario: a metallurgist has the data but no compute; an ecosystem has the compute but no data.


This is the issue that genuinely blocks AI in the loop—both in industry and in banking. Who is responsible when the system makes a mistake? The adopted law did not resolve this issue: the relevant article simply refers to general legislation. And as long as the default answer is "the person who signed," no reasonable executive will hand over decision-making to the system: they keep all the risk to themselves and give all the savings to the company.

What is needed from regulation is not to "permit AI." It is to define the distribution of liability between the model developer, the integrator, and the operator; to set requirements for the traceability and explainability of the decision differentially, by risk class; and, critically, to legalize a space for error in low-risk classes.

Without the latter, the risk-oriented approach will remain a declaration. Because in practice, any corporate lawyer will interpret any application as high-risk—you don't get punished for over-insuring, but you do get punished for an incident. As long as the asymmetry is like this, risk orientation exists only on paper.

Industrial AI hits a wall with industry-specific data, which no single entity possesses alone. For example, I evaluated an AI project for predictive analytics on equipment failures. The results were excellent. One exception—it lacked data. Meanwhile, another company had the exact same turbines, with a history of daily operations and failures over 10 years.

Datasets on equipment failures, defects, and incidents cannot be accumulated within the boundaries of a single company, and the effect of combining them is non-linear. This is a classic case where the state or an industry association should not just provide money, but remove the legal and antitrust barriers to the exchange of anonymized data and establish a standard for such exchange.

And note: this is exactly the same problem I discussed in relation to a single company—just one level up. Within an enterprise, a unified data layer is not built because it has no customer. Within an industry, industry datasets do not appear for exactly the same reason. The same disease at two different scales.

For the EAEU, this is particularly interesting. The market of each individual country is small; together, it is already significant. The joint statement by the heads of state on the responsible development of AI and elevating the topic to the priorities of the chairmanship is the right first step. But the declaration will become a tool only when it is followed by two things: mutual recognition of requirements for AI systems—so that a solution certified in one Union country does not have to undergo the same procedure five times—and a regime for the cross-border exchange of anonymized data. Otherwise, each country will build its own regulatory fence, and instead of one medium-sized market, we will get five small ones. And five small markets do not create export products—they create five local contractors.

What I would change—based on the same three pillars

At the state level, the exact same framework applies as within a company.

Management.

● Change support metrics: not "how much has been substituted," but "revenue from scaling and exports per ruble of state support." We measure what we want to get.

● Subsidize implementation, not development—starting from the second customer.

● Legal certainty regarding liability and traceability, differentiated by risk class; expansion of experimental legal regimes to in-the-loop industrial AI—today they are mostly about transport, telemedicine, and unmanned vehicles.

Technology.

● Infrastructure: priority grid connection, placement near surplus generation, guaranteed compute quota for medium-sized businesses, universities, and industry consortia—not just for three ecosystems.

● Industry-wide anonymized data pools, and within the EAEU framework—mutual recognition of requirements and a cross-border regime for anonymized data.

People.

● Personnel for verification, not just development. We are training data specialists—and almost no one is training those who know how to verify AI in the subject matter: engineers who understand the data, model risk managers, "translators" between the shop floor and the model. This deficit cannot be closed quickly with money. Separately, university access to compute: without it, we will be training specialists using slides.

And one thing the state cannot do

No subsidy and no regulation can replace the maturity of the company itself. Most of the failures I have seen had nothing to do with either money or laws: there was no process owner, no baseline, no measured effect. Waiting for the state to create conditions—and then we will finally implement AI—is the same type of thinking as waiting for a supplier to bring a magic box.

The state can remove barriers. The journey must be undertaken by the companies themselves.

And a projection onto the financial sector. Financiers have a head start here: the Bank of Russia has long accustomed the market to a risk-oriented approach and model validation—this is essentially a ready-made framework for managing model risk, which industry desperately lacks. The exchange is useful in both directions: industry should adopt the banks' discipline in model validation, and banks should adopt from industry the culture of unacceptable events and mandatory failure scenarios. In industry, we are used to designing based on the assumption that everything will fail. In AI, this is exactly the habit that is needed.