Trends in the development of artificial intelligence technologies in Russia. TAdviser 2025 Review

Over the past few years, the AI technology market has grown largely due to hype, but interest in AI has become more and more meaningful. The quality of content generated by AI and the amount of information processing is constantly growing. If a year ago it was quite possible to simply surprise with the capabilities of AI products, now companies are methodically analyzing the quality, safety and relevance of AI-based solutions, as well as assessing their economic efficiency from the point of view of their business.

Traditional barriers affect the development of the market: an acute shortage of personnel, including experienced specialists in the field of AI and machine learning, a high level of costs in general for IT personnel, including software developers and machine learning specialists, availability and quality of data necessary for machine learning, lag of the regulatory sphere from the technical capabilities of AI, conservative and wary attitude towards AI products.

A new review of the AI market from TAdviser is devoted to current technological trends in Russia and the world. We discussed with experts estimates of the volume and dynamics of the market, talked about investments, the applicability of various AI technologies in the economy, the prospects for multimodal AI, NLP, data analysis segments, ML, Speech Recognition, GenAI. Learned how to escape deepfakes and other growing security threats with AI. They also touched upon the topic of confrontation between the largest AI powers and tried to find Russia's place on the global market.

According to the analysis of the Artificial Intelligence Competence Center of the Moscow Institute of Physics and Technology, the growth of the AI market in the Russian Federation in 2023 amounted to 37%, the volume - ₽900 billion. The volume of venture capital investments in AI for the second year in a row remained at the level of $10 million. State financing of AI for the year - ₽9,2 billion.^[1]

According to Statista, the volume of the AI market in the United States in the same year amounted to $37.2 billion, about the same amount was the market of all European countries, including Russia. According to the same data, Russia accounts for $5 billion, which, in my opinion, is somewhat closer to the truth, "said Nikolai Trzhaskal, Product Director of Preferentum, SL Soft.

According to the expert, the analysis of the Center overestimates the market volume by half due to the inclusion of revenue not directly related to the use of AI. The venture capital market in the United States amounted to $170 billion in 2023. Of these, AI accounted for 36% (that is, $61.2 billion). In comparison, the $10 million invested in AI by a Russian venture, even with almost $100 million added by the state, is a drop in the ocean, which in no way can allow Russian AI business to develop outside large corporations.

Maxim Ivanov, Director of Artificial Intelligence at Sber Business Software, confirmed that the growth rate of the AI market in 2023 was estimated at 30-40% both in Russia and in the world, and in 2024 some reports quite significantly increased the estimate, for example, in Bain Technology Report, the growth rate forecast until 2027 was already at the level of 40-55%.

Oleg Korolev, head of development at AI Lab Avito, according to open data, concludes that interest in AI continues to grow in the Russian market. According to the Venture Guide platform, in 2024, a significant increase in investment in AI startups to $33 million was visible - this is more than a threefold increase in just a year.^[2] Global investment in AI will increase by 60% in the next three years, according^[3] by consultancy BCG.^[4] Leading companies are already achieving 10-20% performance growth by introducing AI into everyday operations. At the same time, BCG analysts note that the success of AI initiatives is 70% dependent on people and processes, 20% on products, and only 10% on the algorithms themselves. The main trend of the future Oleg Korolev called all kinds of resource optimizations for training and the work of models.

Some of the experts surveyed by TAdviser agreed that the amount of venture financing in the Russian Federation in this period indicates the conservatism of private capital, and at the same time, in general, the achieved indicators should be recognized as good.

According to Oleg Rogov, candidate of physical and mathematical sciences, head of the Trusted and Secure Intellectual Systems group, AIRI estimates of growth rates of 37% and a total market volume of ₽900 billion indicate high dynamics, but a stable level of venture capital investment shows the caution of private capital and a possible lack of success in commercializing startups. Government funding (₽9,2 billion) is a positive factor. In 2025, further growth can be expected thanks to government orders and large AI implementations in the corporate sector, but the overall situation will strongly depend on macroeconomic stability.

Ekaterina Ionova, director of Lukomorye IT ecosystem projects (part of Rostelecom), also believes that analysts' estimates indicate high demand for technology, but the volume of venture capital investments indicates investor caution. Government funding seemed insufficient for large-scale projects, which shifts the focus to the commercial sector.

Stanislav Ziganshin, Marketing Director of the Video Analytics and Machine Vision Department of SATEL, believes that the achieved indicators for 2023 are a positive signal, especially given the difficult economic conditions. Venture has been at a low level for two years in a row, and in this regard, the expert reflects on the high dependence of the industry on state support. In 2024, the situation developed in a similar way: private investment had limited access to international venture funds and had high risks. Growth in 2024 may be 30-40%, but without active private capital it will be difficult to accelerate development.

Mikhail Telegin, Deputy General Director for Strategic Projects at OBIT, said that the dynamics of investment in the AI industry has significantly decreased in recent years. Since 2022, there has been a sharp decline in cash injections due to sanctions, blocking access to high-performance equipment and technologies. At the same time, the market itself is showing multiple growth, and this can be associated with the transition from "experimental" AI to its active commercial operation. These trends became especially noticeable in 2024 and will intensify in 2025.

The introduction and use of AI technologies involves significant investments. According to the HSE, large-scale projects are funded primarily at the federal level, while relatively small companies that are in the focus of attention of local authorities receive support from the regions.^[5]

Nikolai Trzhaskal believes that the policy of state financing of AI projects is not effective enough, since the state, by definition, cannot act as an effective venture capital investor, and this is exactly what is primarily required for the growth of the high-tech industry. The high discount rate makes it difficult for small companies to access government contracts and contracts with companies with state participation, since they often do not have the opportunity to implement long-term projects with 100% postpayment, and the accepted profit rates do not allow using ultra-expensive credit funds.

All this leads, as a result, to the consolidation of business, which negatively affects the spread of innovations beyond the cities with a population of over one million, the expert noted.

Although the federal focus on large projects creates a base for advanced development, obtaining the funding necessary for breakthrough ideas requires more active participation not only of the state, but also of enterprises of various sectors of the economy throughout the country. As Oleg Rogov said, regional subsidies cannot always cover the costs of training AI models in all important areas, given even special local support programs.

More active interaction of enterprises with university startups will help speed up the launch of breakthrough solutions within the country and thereby contribute to the dynamics of the development of the economy as a whole, the expert hopes.

Stanislav Ziganshin believes that state financing of large AI projects is logical, but limits the access of startups to resources. Regional support remains pinpoint. For the accelerated development of the market, it is important to simplify the access of medium and small companies to grants and subsidies.

Avenir Voronov, Innovation Director of the Analytical Solutions Department of KORUS Consulting Group of Companies, also noted that the current practice is quite logical, but assesses it more positively. Although at the level of federal projects, small companies are "screened out" already at the stage of filing documents and are not able to compete with large players even in formal terms, but at the regional level they may well compete with large players by understanding local realities and individual, point-by-point elaboration of solutions based on local tasks.

Dmitry Kondrashkin, AI Director of VK, spoke about the practice of co-financing AI projects by the state at the end of 2024, which implies investments in projects for testing and pilot implementation of AI technologies to solve applied problems. Often, such programs involve co-financing by an industry partner or company that will use AI-based systems in its practice. This approach allows us to focus research on the development of solutions that have the potential for application and create economic value in industries. At the same time, investments in the development of basic models are mainly carried out at the expense of large technology companies in cooperation with research centers.

It will be possible to assess the effect of investments in industrial solutions after the completion of financed projects and the start of the application of technologies in practice, the expert noted.

In medicine, the amount of state support in various formats can be 70-75% of the total investment in AI projects, Alexander Gusev,' director of development at Webiomed, shared his assessment . The first violin is played here by development institutions such as the Skolkovo Foundation, NTI. In addition, state financing of purchases of medical products with AI has become a serious catalyst for their practical application.

𝓲

Фото: «Freepik»

According to the index of readiness of priority sectors of the Russian economy for the introduction of AI, healthcare was added to the leading financial sector and ICT in 2023. AI becomes possible to use in education, science, social sphere and many other spheres. At the same time, the influence of negative factors noted in previous years has decreased significantly.^[6]

{{quote | The degree of penetration of AI into health care was pleasantly surprising. In our experience, the trade and construction sector is also actively mastering AI. This is primarily due to the process digital maturity of companies. Their digitalization strategies can include dozens of projects in the process, and the same amount will lie in the backlog, "said Ilya Novoseltsev, director of new developments at Rubius.}}

According to Roman Styatyugin, director of the VK Predict Analytical Products Center, the financial sector and ICT are indeed the most predisposed to using AI and obtaining business effects, which is typical for any industry with a high degree of digitalization and the availability of a ready-made infrastructure for data processing and analysis. Such companies have the expertise, expertise and teams needed to apply AI. And here are important not only IT competencies that allow you to develop a solution, quickly implement it and ensure operation, but also competencies for interacting with business in formalizing tasks, assessing economic effects, and having methods that allow you to link mathematical metrics of models to business metrics. The expert noted that healthcare is indeed one of the industries in which the potential for using AI is very high.

The fact that healthcare has been added to the leading industries indicates the expansion of the use of AI in the real sector, Stanislav Ziganshin believes.

State digitalization programs have had an effect: the barriers to implementation have become lower, he said. - Accelerated development of AI in education, industry and transport can be expected in the future.

The potential of AI technologies in medicine, genetics and healthcare is obvious, Avenir Voronov believes, because the cumulative knowledge base with which AI is "trained" is constantly growing, respectively, the accuracy of diagnoses and forecasting is also growing.

Thanks to the digitization of public health (including with the help of AI tools), scientists will be able to respond to public requests more relevant, the expert noted.

" Alexander Gusev' stated that AI in healthcare has been actively growing over the past two years and expressed confidence in further multiple growth in terms of market metrics, the level of use and maturity of the solutions present on the market. According to the expert, the segment is one of the leaders in the general digital health market, which includes basic informatization, telemedicine, management systems and various services for citizens.

At the end of 2024, the total volume of the AI market in healthcare, according to consolidated estimates of various experts, amounted to about ₽15 billion. This amount includes state purchases of medical devices with AI technologies, Moscow's investments in a city experiment in computer vision and the launch of the MosMedII project, internal investments of a number of major players in the commercial and industrial medicine market, as well as research and development in the research field, including for farm. industries.

Dmitry Domarev, CEO of SberMedII, emphasized the fact that in 2023 36.4% of medical organizations used AI, while in 2021 - only 2.5%. The most popular technologies were decision support systems (71%) and computer vision (69%). 84% of medical organizations using AI also prefer intelligent decision support systems. 26.5% of medical organizations using AI services assess the economic effect of them as significant or multiple.

As part of our activities and practice of introducing AI solutions into regional healthcare, we also see the success of such cases, the expert said. - It is possible to reduce the burden on medical workers, increase the quality and speed of diagnostics, increase the availability of medical services in remote areas.

𝓲

Фото: «Freepik»

According to the Institute for Statistical Research and Knowledge Economics of the Higher School of Economics, the possibility of intellectualization of industrial robots, primarily due to the integration of AI solutions, is one of the key areas of development of industrial robotics. It is expected that the global AI market in robotics, which amounts to $17 billion, will show stable growth (about 25% annually).^[7]

As Nikolai Trzhaskal said, robotization in Russia, as in the entire space of Eastern Europe and post-Soviet Central Asia, has always lagged behind Western countries and, in the last decade, China. This was due, among other things, to the fact that the cost of labor is still lower than in the West or in Japan. In recent years, sanctions pressure has also contributed to the slowdown in robotization, which makes it difficult for the advanced robotics solutions to enter the Russian market.

The relatively low cost of labor is a key deterrent to the automation of Russian plants using AI, also believes Anton Chikin, head of the intelligence department at Jet Infosystems. According to the expert, because of this, the economic benefit of complete automation is not always obvious to enterprises. In addition, there are situations when the task of optimizing processes is easier and cheaper to solve with other tools.

Dmitry Demidov, head of the NORBIT Innovation Laboratory (part of the LANIT group), agrees with this opinion:

Even if we take the largest international companies, we regularly hear stories about huge factories where tens of thousands of people who assemble equipment work. So far, it's more cost-effective than completely transferring work to robots.

 
The problem of outdated production infrastructure also negatively affects the development of industrial robotics. Many plants use equipment that is difficult to integrate with new technologies. It copes with its tasks, but how to physically install cameras on it to implement a computer vision service, for example?

We take into operation such complex projects and create a comprehensive solution that allows us to equip such units with cameras. Usually it includes an engineering solution, hardware and software, - said Anton Chikin.

If we take the average condition of all enterprises in the country, then a very large percentage of equipment works here, which is either impossible at all, or it is extremely difficult to automate at least somehow. So here you no longer have to talk about AI, you first need to go through a serious modernization process.

 
The introduction of AI services occurs only after confirming their value to business. At the same time, projects with AI are difficult to assess and calculate the return on investment.

It is necessary to have a confirmed practice of applying a solution at least several enterprises from one area. It is necessary to confirm in practice that the solution is safe and cost-effective.

Our task, in addition to directly creating services for our customers, is to determine those fragments of production where the introduction of AI will definitely bring profit, and not "apply" it at random to any point of production, - said Oleg Rogov.

Today, there are not many companies in the Russian industry who are ready to experiment with such technologies at their production sites, there is a certain conservatism of enterprises.

Entrepreneurs are often unprepared for risks, such as managing complex algorithms, and often underestimate the real impact of implementation. At the same time, the acceptable return on investment was reduced to 1-1.5 years, said Ilya Novoseltsev.

Although in Russia there are already successful examples. The most recent - at Avtotor in Kaliningrad, a fully robotic line for assembling automobile bodies was launched. In most cases, this is still not a deserted plant, rather, a deserted conveyor, but the growth rate in the use of industrial robots is decent, the expert said.

 
Full automation with AI is an expensive and complex process that can stretch for years. This is a matter of long-term investment and systematic modernization. Such projects need capital temporary and financial resources, which is due to the low level of automation of Russian plants using AI, experts say.

The transition of industry to AI-rails requires investment, primarily in the infrastructure base. It is necessary to plan the implementation, work with employees at the plant to avoid sabotage, purchase equipment and, possibly, stop production lines for a while.

{{quote | Russian factories are a prime example of a request for multimodality. The key problem here is the lack of end-to-end integration. In real life, it is impossible to automate one part of production and say: "That's it, I'm great!." You need to automate the entire process - from software development to controlling a specific device, machine in real time. So far, this is hindered by the high cost of implementation, - said Nikita Nazarov,' technical director of the IT company HFLabs.}}

 
Many experts mention the shortage of qualified specialists in enterprises as one of the slowing factors. Training qualified specialists in system integration and big data analysis can help accelerate the spread of industrial AI solutions in Russia.

{{quote | We must first overcome the boundary of basic automation and robotization, and only then reflect on the use of AI as an advanced tool for optimizing and controlling the operation of robotic systems. For 2023, there were about 15% of industries that are ready for this. We should also not forget about profitability. AI and robotization are needed where the current maturity of the production process does not cover consumer demand, "said Nikita Kupriyanov, Technology Director of Zyfra Group.}}

According to the expert, basic automation grows annually, which means the opportunity opens up to use AI as an automation tool for more and more production sites, which should be used, since the potential is great here.

𝓲

Фото: «Pexels.com»

Interviewed by TAdviser, the experts are unanimous: DeepSeek's Chinese neural network R1, which made a lot of noise, is a promising development, but so far it cannot be called the "killer" of ChatGPT, Silicon Valley and its startups, although such concerns have been expressed. Marketing in the AI technology market plays no less role than, for example, in the automotive industry or telecom, so competitors of well-known large brands can appear every season. And the release of the next LLM provokes an inevitable comparison.

DeepSeek is largely based on the neat application of recent advances in large language models. This is a free and almost unlimited neural network that can solve more complex problems than traditional LLM models (such as models of the Llama-3 family), which require many steps to get a solution.

In addition to simply generating text, it can solve more complex logical problems through Chain of Thought Reasoning: the model "reflexes" and forms a plan before answering. Therefore, DeepSeek-R1 is called the "reflecting" LLM based on DeepSeek V3. This is a smart agent who does not respond immediately, but at first thinks almost like a person.

 
One of the main values, the main breakthrough, in addition to a given high level of quality, is the speed, cost of creating, training and functioning the model.

Its developers have achieved a significant reduction in costs, using limited resources through optimization of mathematics, the effective operation of data centers, their own software and adapted learning processes, said Oleg Korolev.

R1 copes with tasks at the level of OpenAI's o1 model, available to paid users. The uniqueness and reason for the general attention is that R1 turned out to be the first in its class available openly and free of charge, not for researchers and business, but for the general public, and besides, in integration with Internet search.

The model is available for testing absolutely free, and not for $20 per month, as in the case of ChatGPT. At the same time, testing the model as a whole shows similar quality results to ChatGPT. In addition, DeepSeek can be used through the API (as well as the OpenAI product), but the price of the R1 token is 30 times cheaper than that of o1, - said Anton Chikin.

DeepSeek requires less computing power compared to large external solutions, and, importantly, there is no need to rebuild software for it, and there is also the possibility of direct access to servers. In addition, the neural network has great potential at the architecture level.

So, with a volume of parameters that is ten times less than that of competitors, the quality of the generated image is no worse. That is, the network can still be trained and increase the number of parameters, - said Avenir Voronov.

{{quote | The new neuron also demonstrated the possibility of effective training of models on less powerful and cheaper chips, - said Anna Malysheva, head of marketing at Zyfra Group.}}

Open code allows developers to adapt algorithms to their tasks - from automating document flow to creating chat bots, which accelerates the introduction of AI into small businesses and startups, - said Ekaterina Ionova.

 
Despite the fact that the resonant release of DeepSeek was not a premiere, and back in the spring of 2024, the younger product model had already caused a violent reaction in the Chinese market, R1 could still be called a rather raw product that could not cope with the load of many users that fell on it.

{{quote | There was no revolution, and the attention that was attracted to the release of the new DeepSeek model was rather a viral effect due to the fall in Nvidia shares, which is associated with the release of the new model (although it is not a fact that this is the only reason for the fall in stock prices), - commented Alexander Krushinsky, Director of the Voice Digital Technologies Department of the company BSS.}}

According to Alexander Krushinsky,' DeepSeek models are at the level of leaders of the ChatGPT-4o type, but there is no high-quality separation. According to the Chatbot Arena rating, DeepSeek was in 4th place. The architecture of the model and the approaches to its training use the already known approaches of "reinforcement learning," when the responses of the models are evaluated by people and the model is adjusted taking into account this assessment. It is argued that training the model cost significantly less than teaching ChatGPT, but this is difficult to verify: there may be many nuances in the assessment method, or simply hiding the real capacities that DeepSeek had.

{{quote | The manufacturability of DeepSeek, both V3 and R1 ‒ is about the same as GPT-4o and GPT-o1. With V3 and GPT-4o, everything seems to be clear, with this earlier we have all played enough. In terms of technology, nothing breakthrough happened, - said Sergey Litvinov, head of the center of competence of big data and artificial intelligence of the LANIT group of companies .}}

As for security, the expert noted, then all AI models are about the same. They remain "black boxes," and there are a large number of scientific discussions on this topic, about the results of which we will still hear.

Alexey Chistyakov, technical leader of Bercut, believes that DeepSeek is not an Open Source product despite being positioned as an "open" alternative to Western models, since it does not provide full access to its code and does not allow free change or distribution of the model.

In fact, this is another closed model, more optimized, while free, but not giving users fundamental advantages over, for example, OpenAI, the expert noted.

{{quote | It is important to look at changes in distance: the same OpenAI has multiplied performance and capabilities over two years. DeepSeek has also gone a solid way, but whether it will show the same growth rate in the horizon of the year is still unknown. I would like to see more than one model from High-Flyer in opensorse, - shared Nikita Nazarov.}}

As for the quality of DeepSeek's work in Russian and its safety, in these respects it is far from ideal, since it studied mainly on Chinese and English data, Dmitry Kondrashkin said. According to the VK expert, English words and even Chinese characters are often found in DeepSeek's answers in Russian. Nevertheless, further training of DeepSeek for the Russian language is quite possible.

Nikolai Trzhaskal began testing DeepSeek even before it became mainstream:

I can say that he is far from his local brother Qwen, and even further to advanced OpenAI models. I personally found him very useful in implementing a multi-agent environment, where he acts as a critic of artifacts generated by GPT-4o. Here is such a wonderful synergy of the two worlds and will lead humanity to the next round of development.

 
DeepSeek challenges traditional monetization models like subscriptions, jeopardizes the monopoly of Western AI giants, primarily thanks to free access.

Despite the fact that the model itself does not demonstrate revolutionary results, the fact that the Chinese company was able to create a product close to the leaders suggests that at the current stage, large American corporations have failed to maintain a monopoly on AI. Modern generative models have great potential, and the monopoly ownership of these technologies by several corporations within the framework of one country would pose a threat and limit competition, and therefore further technological development. This opinion was shared by Alexander Krushinsky.

{{quote | This story, in my opinion, launched a paradigm shift from the race for computing power towards optimizing learning and working with artificial intelligence, "says Mikhail Telegin.}}'

Before the advent of the Chinese neural network, it was believed that a huge amount of computer hardware was needed to reinforce language models, and Nvidia was supposed to be the main beneficiary of this huge development. DeepSeek demonstrated to the market that it is not necessary to inject a similar amount of resources to train AI models with characteristics similar to California developments.

{{quote | DeepSeek certainly had a very strong impact on Nvidia because they came up with and implemented several innovations that allowed a radical way, about 20 times, to reduce the cost of training and the use of models, commented' Maxim Ivanov.}}

{{quote | The industry is strengthened by the belief that without huge money and expensive video cards, you can create equally high-quality models. It gave the market outside of America a positive. DeepSeek is definitely a mouse that can scare an elephant, Nikita Nazarov commented .}}

DeepSeek demonstrates the key principle - progress in the field of AI is not only in increasing hardware capacity, but also in optimizing training algorithms and infrastructure. The open publication DeepSeek of their research works and technical documentation contributes to the development of the entire industry.

{{quote | This will allow companies to learn this experience and try to implement such optimizations in their models. Such transparency, while not revealing all the details, will lead to valuable research into DeepSeek's teaching methods. Potentially, this case could accelerate the democratization of the development of advanced AI, - said Oleg Korolyov.}}

 
Companies need time to compare R1 with other models on their application tasks. Whether interest in Chinese development continues after May 2025 depends on how OpenAI responds and how DeepSeek develops. Most likely he will take part of the niche. However, someone else can demonstrate such inexpensive training of neural networks. A scenario is discussed in which the market is radically restructured, and in addition to large players there will be a huge number of startups that will develop something competitive with small resources. Avenir Voronov believes that the R1 case is inspiring, and new Google and Microsoft will be born in the coming years.

Losing in this race is almost impossible, everyone can find their niche. And as the case with DeepSeek shows, large players should not relax: today's developments will not cost anything tomorrow if they are not constantly developed, the expert warned.

"Classic Disraptor" called DeepSeekNikolai Trzhaskal. The expert predicts that many engineers in the West and East will now look for additional opportunities to solve complex problems using fewer computing resources.

DeepSeek managed to keep the market off balance, but new ideas, developments are spreading quickly, and the same approaches will begin to use grandees, which means that again everything will ultimately be reduced to resources, computer hardware and more. The difference from the previous situation will be that potentially one grandee will become more on the market. That is, the influence will not be as radical as in some scenarios, but the competition will definitely intensify. This opinion was expressed by Maxim Ivanov.

Does not predict the mass transition of users to DeepSeek, but Anna Malysheva sees it as one of the full-fledged instruments in the orchestra of AI assistants. According to her, she R1 is very likely to be the loudest technology market event of 2025 and the threat of a temporary outflow of users from OpenAI and, Google as well as a decrease in demand for high-performance GPUs from such large companies as. Nvidia

 
What objective lesson from the DeepSeek case should be learned quickly?

First, the world saw that China has AI ambitions, China is actively developing its own LLMs and seeks to reduce dependence on Western technology. It is also critical for Russia to develop its AI models in order to avoid dependence on foreign suppliers.

Secondly, experts note, it seems that today in the global market there is a time for innovators and challengers to challenge the status quo. It's time to invest in your own laboratories and developments, experiment and not be afraid of competition with big players. In other words, today you do not need to be a mega-corporation to make a revolution in AI. Unexpected twists are possible even in markets where everyone is used to "stability," and the global AI market may await many more discoveries.

Thirdly, in the market, due to the "DeepSeek dysrapt," values ​ ​ are being updated. On the one hand, the focus is shifting from technological superiority to service, ecosystem. When basic tools become free, the key advantage becomes not the model itself, but platforms that teach it to solve narrow problems, integrate with other systems or provide security. This forces even large companies to move from selling smart hardware to creating unique add-ons, where value is formed through personalization and user trust. On the other hand, DeepSeek demonstrated profit from international cooperation based on open solutions. Everyone saw that there was a huge demand for them.

𝓲

Фото: «Freepik»

According to the Artificial Intelligence Competence Center of the Moscow Institute of Physics and Technology, the importance of two segments of the AI market is overwhelmingly great: in 2023, the NLP segment (Natural language processing) amounted to 61.3% of the entire AI market of the Russian Federation (in 2021 - 32.8%). The second most important segment is the data analysis segment (33.6%).

According to these surveys, the segments of computer vision and recommendation systems in Russia have significantly decreased or even found themselves on the verge of stagnation. However, this is unlikely, as these technologies continue to be actively applied in various industries. This opinion was expressed by Anton Chikin.

As Maxim Ivanov' said , the imbalance towards NLP in the Center's data is primarily caused by the growth in revenue of Yandex, Avito and VK, which, according to the researchers' methodology, belong to the NLP segment, although these companies are engaged in a much wider range of technologies.

As Nikolai Trzhaskal said, segmentation is adopted in world practice: AI robotics, autonomous machine and sensor technologies, computer vision, ML (recommendation systems, data analytics and predictive analytics), NLP. The proportions between all of these components are approximately the same. ML accounts for 32-37%, and NLP - about 18-25%. If we take the data of the Center as a basis, then robotics, autonomous machines and sensors, computer vision remains only 5.9%.

In my opinion, such figures do not correspond to reality, given the increased growth of the military-industrial complex industry, where all three of these segments are key drivers for ensuring defense capability. Also, taking into account the need for import substitution of analytics, the figure of 33.6% seems somewhat underestimated, and the volume of NLP is at least 2-2.5 times overestimated, the expert noted.

For example, a slightly different "pie" of data has developed in healthcare. The computer vision sector traditionally leads here, since the use of AI for the analysis of medical images is one of the first and most developed areas. But then there is the analysis of electronic medical records, where NLP technologies are actively used.

As Alexander Gusev' said , up to 90% of records accumulated in electronic medical records are unstructured medical texts that cannot be correctly analyzed without technologies for working with a natural language. And the expert called data analysis the third segment in medicine: predictive analytics, RWD research, support for management decision-making.

Market segmentation is very often quite arbitrary. In practice, we have seen an increase in AI applications in the field of automation of business processes, which are differently attributed in a particular classification of market segments. The driving force behind this growth is the business's natural desire to reduce costs, - commented Dmitry Kondrashkin.

Many experts confirm this opinion when they say that technologies in the NLP and data analysis segments have reached a high level of maturity, directly affect the efficiency and competitiveness of companies in a digitalization environment, and bring "fast" profits.

For example, Ekaterina Ionova believes that the growth of NLP is associated with the demand for automation of communications: chat bots, voice assistants, text analysis and tonality systems, automatic translation help companies process customer requests faster, reducing costs.

Oleg Rogov noted that data analysis is a fundamental area that companies use to predict, optimize their business and make decisions based on large amounts of information.

{{quote | These two segments provide the greatest economic effect for large businesses. NLP is widely used to automate service in contact centers, which gives huge savings on PHY operators. Data analysis technologies are widely used to increase sales in advisory and analytical systems, - commented Alexander Krushinsky.}}

The main driver of NLP growth is large language models (LLM), stated Vladislav Tushkanov, head of the research and development group of machine learning technologies at Kaspersky Lab. According to the expert, the quality and ability of models to solve various problems is growing (for example, the potential length of documents that LLM can process has increased by several orders of magnitude), their reliability and speed are growing, and the cost is falling.

So, the GPT-4 cost $60 per million tokens, while a more powerful and modern GPT-4o costs already $10 per million tokens. Competing with it in some quality tasks, the Llama-3.1.405B is available from cloud providers for less than $5 - with the potential to deploy such a model in its own infrastructure for processing confidential data. All this allows you to use LLM in application tasks: from automation of technical support to applications in software development and cybersecurity.

An additional impetus to the development of this segment in Russia is given by the fact that closed and open models from leading foreign developers are working better with the Russian language, and in parallel with this, models from Russian vendors are being actively improved - both trained from scratch and adapted on the basis of open models.

In 2022-2023. the emergence and distribution of powerful language models, such as Claude, led to the active development of domestic analogues GigaChat, YandexGPT. The localization of technologies, import substitution played an important role in strengthening the positions of the NLP segment. In addition, along with the growth in the number of text data and digital communications (instant messengers, social networks, online appeals), the demand for intelligent text processing is also growing. The attractiveness of NLP technologies is increased by the availability of computing power and the development of transformer-based models (GPT, BERT).

Government services, including Public services, are introducing them to automate the processing of citizens' requests. Banks, retail and telecommunications companies are actively using them to improve the quality of customer service.

Dmitry Goltsov, Deputy General Director for Commercial Affairs of Megaputer, noted two popular areas of NLP application: customer feedback analysis and competitive intelligence.

In recent years, with the development of NLP technologies, issues of combating synthetic negative reviews have come on the agenda, as well as creating artificial feedback in their own interests. Competitive intelligence based on data from open sources - OSINT (Open Source INTelligence) - cannot be called a young direction. This is a systematic collection and analysis of information on advanced technological trends and key plans and actions of competitors.

The data analysis segment, in turn, is growing due to the fact that the business is looking for analytical tools - companies need demand forecasting, risk analysis, marketing automation and supply chain optimization. The market for IoT and industrial analytics is also growing, that is, huge amounts of data appear that need to be analyzed. In addition, enterprises use machine learning to predict equipment failures, prevent fraud, and improve operational efficiency. This was told by Alexey Chistyakov.

The overall level of readiness of companies to use analysis technologies is growing, as well as an understanding of the value of data for these tasks. The technical competence of companies is increasing, which allows more effective implementation of ML solutions, said Sergey Litvinov.

Huge amounts of information are critical for training AI models, but working with them requires a lot of high-level specialists and computing power. Cloud platforms are also growing from large Russian and global players - they provide infrastructure for cleaning, deduplication and preparing for analysis of large amounts of data. Not all companies can afford to maintain such an infrastructure on their own, said Oleg Korolev.

The Russian cloud services market is becoming a key platform for introducing advanced AI solutions. This is primarily due to the growing needs of the business in processing large amounts of data, training neural networks and automation of processes. This was announced by Alexander Obukhov, Product Director of RTK-DPC, owner of the Public Cloud product.

According to research, the MLaaS and GPUaaS market is growing at an average of 20-30% per year, and the demand for these services is only increasing. Such infrastructure solutions give companies powerful resources to work with AI. The use of GPUaaS is more profitable than the independent purchase and support of specialized equipment. For example, when developing answering bots, neural network training takes several months, and then only minimal adjustments are required. According to Alexander Obukhov, several years ago companies had to either buy expensive servers that were idle after training, or work with limited resources.

𝓲

Фото: «Freepik»

Against the background of NLP and analytics, the volumes of the Speech Recognition and GenAI (Generative AI) segments still look insignificant. FinTech's cybersecurity is growing noticeably relative to its own performance.

The last two segments are developing with AI successfully due to the circumstances that have developed some time ago: cybersecurity - against the background of growing threats, FinTech is traditionally a strong point of the Russian high-tech industry, which receives an additional impetus due to the exponential growth of digital financial services and electronic payments.

It is worth honestly admitting that in general, the level of digitalization of the Russian banking sector and the quality of services for corporate and individuals are, if not the most, then definitely one of the best in the world. With the introduction of legislation on working with digital financial assets, we can expect further breakthroughs in this direction, - said Nikolai Trzhaskal.

The role of cybersecurity is growing around the world, and in Russia, additional growth drivers are the need for import substitution and the current geopolitical situation in which the country must defend its cyberspace with a vengeance. Russian companies are actively working in this area and have a recognized reputation not only at home, but also abroad.

Speech Recognition and GenAI still occupy a smaller share due to a relatively late start and higher barriers to the market (requirements for computing resources, specific application scenarios), but are growing relative to previous periods.

Both of these segments currently require access to the most advanced computing power. This access is seriously complicated by the sanctions policy against Russia. In addition, the main boom in the growth of generative AI began in 2024 and will continue for the next 5 years, increasing at least one and a half times annually. In the context of a shortage of hardware, it would be reasonable to assume that the efforts of Russian startups will be aimed at creating less "voracious" models that study on smaller amounts of data and are able to produce results while spending less computing resources. This is the way the developers of the DeepSeek model, which has become a global hit, went. However, without a radical change in the situation with venture capital investments, it is hardly possible to expect the same breakthroughs.

Шаблон:Quote 'author=Avenir Voronov, Innovation Director, Analytical Solutions Department, KORUS Consulting Group

It is also possible that the reason for the low volume of the GenAI segment is due to the fact that this is technology, while NLP or cybersecurity are applications, and in a business context it is the practical result that matters. It is worth noting that the areas of GenAI and NLP are closely related, so the growth of NLP is largely due to the active introduction of generative models (and both classes of technologies can be part of FinTech solutions).

The Speech Recognition and speech analytics segment also develops in conjunction with language models, which provides it with good prospects. These technologies are increasingly used in voice assistants, automated call centers, biometric identification and multimodal AI systems. To date, their maturity is assessed by experts as quite high. As models improve and computing power grows, their application will only expand.

Шаблон:Quote 'author=Alexey Lyubimov, CEO of 3iTech

𝓲

Фото: «iuriimotov / Freepik»

Multimodal artificial intelligence can interpret several types of data at once (text, images, audio, including voice, natural language, video, tables and graphs, equipment performance indicators, and much more). Such models can take into account not only individual data elements, but also their relationships, which leads to a better understanding of user intentions and query context.

OpenAI's CLIP can interpret the text descriptions of objects depicted in the picture and recognize their visual features. Gen2 from Runway AI generates video clips based on text and graphic prompts or changes the captured video. DeepMind's Flamingo takes video, images and text as input and generates text responses. Russia has its own developments, for example, LLM GigaChat from Sberbank is today supplemented by image recognition.

This direction began to actively develop relatively recently. Gartner predicts that by 2027, the 40% of GenAI solutions will apply multimodality (in 2023 it was 1%). The consulting company puts multimodal Genia on the same level with Open Source LLM in the potential to influence organizations over the next five years.^[8] Analysts at MarketsandMarkets estimated a CAGR of 35% in this AI class (2023-2028).^[9]

Шаблон:Quote 'author=Ekaterina Ionova, director of Lukomorye IT ecosystem projects (part of Rostelecom)

Шаблон:Quote 'author=Anton Chikin, Head of Data Mining at Jet Infosystems

Шаблон:Quote 'author=Stanislav Ziganshin, Marketing Director, Video Analytics and Machine Vision Department, SATEL

Шаблон:Quote 'author=Nikolay Trzhaskal, Chief Product Officer of Preferentum, SL Soft

Шаблон:Quote 'author=Oleg Korolev, Head of Development at AI Lab Avito

 
Business

In the corporate segment, where data is rarely limited to one format, multimodal AI is in demand for speeding up, automating the solution of various application problems: analyzing information of various modalities for semantic search in internal corporate knowledge bases, including graphs, images, presentations, drawings; intelligent BI systems - extracting insights from data and generating graphs and dashboards; sammarization of audio and video content of meetings and much more.

Marketing

GenAI solutions are already changing marketing and creative industries, helping to create content in different formats (text, images, audio). It is expected that multimodal AI should bring this process to the automatic generation of marketing materials with minimal human participation: preparation of text, images, video, content recommendations for various formats, etc.

Client Service

Multimodal AI improves the capabilities of chatbots and voice assistants: they analyze the tone of the voice, recognize emotions in the video and in the text of the request, thereby penetrating deeper into the context, the client's intentions, in order to more accurately respond to requests.

Retail

In specialised areas such as retail, multimodal AI can simultaneously analyse photos of retail shelves, customer behaviour data and customer feedback. Technology-based solutions can provide comprehensive conclusions about consumer preferences and assess the effectiveness of product delivery. And they will do it in minutes instead of the days that it would take for a person to draw up such a report.

Medicine

Today, most medical AI services are limited to specific tasks, using only one type of data, such as CT scans or patient complaints. This approach is different from how doctors work: for diagnosis and complex treatment, they use data from various sources.

Шаблон:Quote 'author=Alexander Gusev, Development Director of Webiomed

With multimodal AI, medicine should get tools to match and combine data from snapshots, tomograms, cardio- and encephalograms, assays, symptom descriptions, data in medical records (medical history), and build a more comprehensive health analysis.

Industry

Multimodal AI in industry are intelligent forecasting systems, automation of complex processes. Industrial design and generation of design documents, including 3D layouts. The option to work with drawings and their text descriptions at the same time to improve the accuracy of the design documentation analysis. In the future - full-fledged generative design.

Security area

In security, the combination of audio and video analytics will improve the accuracy of recognizing potentially dangerous situations.

Smart cities

In the future, such solutions will become an important part of smart ecosystems, for example, in smart cities. Multimodal AI will analyze the flows of people, weather, data from cameras, manage transport.

 
The practical (commercial) application of multimodality is still quite limited, and at its core multimodal models are a new generation of Genii, and so far it is more about starting to massively use ordinary generative models, said Maxim Ivanov.'

Alexander Krushinsky assesses the significance of the segment as temporarily low. According to him, in general today there are not many cases with an understandable business effect that can be implemented through multimodal AI, but cannot be implemented through a combination of several monomodal ones, although there are such cases. The expert gave an example with an intellectual reading of emotions based on the analysis of both facial expressions, voice timbre and the content of remarks.

Building serious solutions based on multimodal AI requires very serious initial investments, Dmitry Goltsov believes. Therefore, at the moment, the main interest in building solutions based on multimodal AI is observed from large business players who have the necessary investment and personnel reserve to experiment in such a young technological field with not fully proven economic effects.

The growth of the direction depends on solving problems: training models on heterogeneous data, protecting privacy and reducing computing costs. Those who invest in the integration of multimodal systems will be able to offer unique products, but mass implementation will take time and collaboration between business, developers and regulators. This opinion was expressed by Ekaterina Ionova.

 
Multimodal AI will demonstrate accelerated development and can become the core for a huge number of new applications and services, and may even be used everywhere.

Avenir Voronov suggests that after the stage of optimization and segmentation, it may turn out that highly specialized, low-budget and resource-intensive MM models will be no less significant for the market than huge multimodal corporate or state systems.

Шаблон:Quote 'author=Nikita Nazarov, CTO of IT company HFLabs

𝓲

Фото: «Freepik»

21% of business representatives from 39 Russian companies surveyed by MTS AI and B1 Group admitted that they suffered from fraud using new technologies. The study assessed the threats from spoofing and deepfakes and the readiness of businesses to resist them. Messengers were named the most vulnerable communication channel.^[10] Experts interviewed by TAdviser largely share the concerns of survey respondents. However, there are also criticisms.

Mikhail Telegin noted that in Russia the use of AI by cybercriminals is not so widespread, in comparison with the global trend. According to Anton Chikin, the threat of AI fraud is real, but it is still difficult to assess the contribution of this class of technology to the overall level of cyber threats compared to traditional social engineering. Nikolai Trzhaskal believes that the sample of 39 companies is not representative, so you should not take the 21% figure seriously.

In this whole story, I would not talk about the danger of deepfakes, but about the danger of illiteracy in general and digital illiteracy - in particular, the expert noted.

In addition to deepfakes, the threats that cybercriminals can create with AI are diverse, and experts expect the trend to complicate attacks to continue.

Avenir Voronov said that deepfakes are one of the simplest technologies used by scammers, and full digital clones of individual personalities, or even entire companies, are more dangerous.

That is, you can work with a robot counterparty or a company of robots that behave absolutely like humans, but do not bear any responsibility for the services that they represent. The second example is when the employer is sure that the employee works in a remote format, and this employee does not actually exist, the expert shared.

LLM can be used to automate phishing and other attacks, said Vladislav Tushkanov.

Already now we see their use by cybercriminals for mass generation of phishing pages, there are many messages about the use of LLM for targeted phishing, the expert noted.

It is worth paying attention to the threat posed by the generation of fake documents, certificates and QR codes using AI. These fakes can be used in personalized attacks on specific individuals, which can lead to serious financial and reputational losses for companies.

Attackers will soon be able to launch large-scale and coordinated DDoS attacks with AI, which will make it difficult to protect and restore the normal functioning of systems. In addition, there is a possibility that in the future, using AI, it will be possible to quickly identify vulnerabilities in domains and systems. AI will find weaknesses faster than humans, which will significantly speed up the attack process.

Шаблон:Quote 'author=Vladislav Tushkanov, Head of Machine Learning Research and Development at Kaspersky Lab

Шаблон:Quote 'author=Dmitry Medvedev, Director of the Department of Applied Solutions "LANIT-TERKOM" (part of the LANIT group)

Шаблон:Quote 'author=Nikita Nazarov, CTO of IT company HFLabs

AI algorithms can generate personalized phishing emails, crack passwords. For example, neural networks analyze data from social networks in order to create convincing messages for specific recipients, flexibly build an attack strategy taking into account available information about a person, psychological techniques, human behavior in the process of cyber attack. The technology that allows deepfakes to be created is commoditized.

Шаблон:Quote 'author=Dmitry Medvedev, Director of the Department of Applied Solutions "LANIT-TERKOM" (part of the LANIT group)

Шаблон:Quote 'author=Emin Tagiyev, senior lecturer at the Department of Engineering Cybernetics, NUST MISIS

Шаблон:Quote 'author=Nikita Nazarov, CTO of IT company HFLabs

Nikita Nazarov recalled that until relatively recently there were many stories about cases when their managers allegedly wrote to company employees and asked to talk with law enforcement agencies. With AI, these scenarios can become much more complex and mislead even advanced users.

Шаблон:Quote 'author=Alexander Gusev, Development Director of Webiomed

Compliance with the minimum information security requirements can protect against many problems. Experts advise to conduct all working communication exclusively in work mail and internal corporate messengers, always configure two-factor authentication, and increase the information security literacy of employees. Implement and constantly update more advanced monitoring systems, invest in authentication systems, anti-fraud analytics, cybersecurity, and machine learning solutions. For example, there are systems that detect anomalies in the voice or video.

In countering cyber fraud, active cooperation between developers of AI-based solutions and players in the information security market can help.

𝓲

Фото: «Freepik»

The United States and China are key global AI centers in many market metrics: the number of patents, the availability of infrastructure and personnel, the availability of data for machine learning, etc. Both countries are working quite systematically and purposefully to strengthen their positions in the field of AI, realizing that this is a key factor for future economic growth, national security and global influence. They will continue to compete hard, which in general can have a very positive effect on the global AI market.

The United States is a world leader in AI, developing powerful models, integrating them into global products. The US invests huge amounts of money in AI, has a powerful research base and a strong startup ecosystem. Companies like OpenAI, Google DeepMind, Anthropic and Microsoft are dictating trends, and venture financing and access to advanced computing power are further dominating.

Шаблон:Quote 'author=Alexander Gusev, Development Director of Webiomed

China is betting on industrial application of AI, automation of cities and independence from Western technology, making it a key U.S. competitor. A striking manifestation of China's technological capabilities was the release of the new DeepSeek model, which was able to impose the fight on the American ChatGPT in the digital assistant market. In the future, China is also expected to respond to the United States not "head-on" and in a mirror, but traditionally more cunning and subtle behavior.

China's strength lies in state support, major tech giants, rapid infrastructure development for AI, mass implementation, data control. Thanks to scale and regulation, China can move beyond competitors in applied facial recognition solutions, smart cities.

The amount of resources of the United States and China is such that it will be extremely difficult to "catch up and overtake" them in developing fundamental tasks in the field of AI. Experts note that Russia is unlikely to be able to occupy its niche in the global AI market (unless we are talking about applied solutions), but you should take the best of foreign developments and use it in your interests, grow your hardware and software base, carry out technological modernization of your economy, using AI wherever it gives economic or social effect.

In one of the scenarios, the Russian Federation may go into the creation of a "sovereign AI" trained on Russian data: the state and corporations will allocate serious investments for the development of this area. In connection with this trend, Russian AI has prospects, but it will gain its client base not through technological capacities, but through geopolitical factors: due to localization and with a comparable capacity of Russian developments, "our AI" will spread in those regions where Russian influence is strong or, for example, there will be restrictions on the import of technologies from leading countries.

Russia can occupy a niche of specialized solutions in industry, cybersecurity, public administration, and the development of AI in the Russian Federation can be built around internal needs and priority industries.

Шаблон:Quote 'author=Mikhail Telegin, Deputy General Director for Strategic Projects "OBIT"

According to Ekaterina Ionova, Russia is able to develop technologies in areas where local features or security are important. For example, create AI for government agencies, defense tasks or replacing foreign solutions, as well as adapt algorithms to difficult conditions - from Arctic latitudes to underground mines. According to the expert, success depends on the balance between state support and market demands. National projects like Digital Economy and benefits for startups can become the basis, especially if they are supplemented by partnerships with BRICS countries to exchange experience. The key factor is the focus on real sectors: oil and gas, logistics, metallurgy, where business is already ready to introduce AI to optimize costs. This will help to gain a foothold in narrow but promising directions, turning restrictions into growth points.

Шаблон:Quote 'author=Nikita Kupriyanov, Technology Director of Zyfra Group

Шаблон:Quote 'author=Dmitry Demidov, head of the NORBIT Innovation Laboratory (part of the LANIT group)

Anton Chikin believes that Russia can occupy its niche in areas where localized language models are required (NLP for the Russian language, import substitution of Western solutions) and specialized AI solutions for defense and security.

As Dmitry Kondrashkin said, Russia has almost a full range of its own developments in the field of AI - from large language models to machine vision for industrial applications - second only to the United States and China and significantly surpassing the EU. According to the expert, the development of domestic AI is mainly made for solutions in the field of automation of business processes, security, as well as B2C products and services for the domestic market. A promising external market is the countries of the "global South," which do not want to fall into technological dependence on the United States and China.

Шаблон:Quote 'author=Dmitry Domarev, CEO of SberMedII

Alexey Chistyakov believes that Russia is quite capable of creating an international scientific center of expertise in the field of AI. The unification of scientists and developers from all over the world will allow the implementation of world-class Open Source projects. Russia can become a coordinator of such activities. According to the expert, in order to achieve real technological sovereignty, and even advantages, a radical new "business model" is needed, since it is obviously impossible to compete with the budgets of the United States and China.

Шаблон:Quote 'author=Nikolay Trzhaskal, Chief Product Officer of Preferentum, SL Soft

𝓲

Фото: «Freepik»

Vendor companies that took part in this TAdviser review shared information about their own latest developments and projects with AI, talked about what products are most proud of, as well as about the prospects for their further development, functionality and benefits for the target audience.

SL Soft

SL Soft aims to make intelligent document processing best practices available to a wide audience by gradually lowering the entry threshold, both financial and technological, for new customers. The company also conducts development with a high share of research components. Here work is carried out mainly with large corporations. For example, in 2024, together with TsNIImash JSC (the head institute of Roscosmos Corporation), the company launched a prototype system for analyzing technical standards for compliance and contradictions (SATSPIN). This system is designed to support the work of technical specialists (engineers, designers, designers) in solving search and analytical problems that arise when working with design and technical documentation. It solves the problems of mutual consistency of requirements and their compliance with new documents, based on the entire corpus (thousands of documents since the 1970s) of state and industry standards, directly or indirectly related to rocket and space technologies.

This is the first step towards full-featured generative design, when AI will become not just an assistant, but a full-fledged partner in the design team, capable of solving many turnkey problems, "Nikolai Trzhaskal shared.

Rubius

Rubius develops an online master's degree in computer vision (together with Skillfactory), has its own video analytics product, a technical guild for ML, more than three dozen implemented R&D and AI projects.

{{quote | Most of the current AI projects fall under a non-disclosure agreement, so we can only talk about them in general. Of the latter: machine vision for quality control in production, equipment failure forecast, BI for supply chain with predictive analysis modules for demand and deviations. These solutions are among the most promising today because they provide measurable results for the business. Each of the projects in our portfolio went through the PoC, MVP stages and proved cost-effectiveness. This is our rule: do not rush into the pool with your head, but conduct high-quality analytics, test hypotheses, discuss results and only then decide to continue the project, "said Ilya Orlov, Rubius development manager.}}

3iTech

3iTech specializes in speech analytics using artificial intelligence and machine learning (AI, ML, NLU), including its own large language model 3i LLM, has been working with neural networks for more than 10 years. Creates voice and text bots, speech recognition and synthesis services, intelligent processing solutions for unstructured data and omnichannel analytics. Briefly, the mission of the company Alexey Lyubimov, CEO of 3iTech, formulated as help in automating work and improving communication between business and customers using AI.

Solutions are 3iTech in demand in companies where client communication plays a key role and innovation is vital in terms of competition. Customers use the company's solutions to analyze service quality, monitor script and regulatory compliance, automate call assessment, and identify insights that help improve business processes.

In 2024, 3iTech became the AI partner of Silver Mercury, the largest advertising and marketing communications festival. The company's AI development was applied as a member of the jury, and in parallel, projects were evaluated by over a thousand industry experts. AI was trained on the datacet of the festival for 24 previous years, analyzed over 1,600 creative projects of the contestants and selected the winners in 133 nominations, in clear accordance with the assessment criteria set by the organizer of the festival. In this case, AI from 3iTech coped with the task 7,000 times faster than a person.

In the Rostelecom contact center, AI solutions 3iTech helped to increase the labor productivity of operators by 21%, while the speed of receiving answers by contact center clients increased sevenfold. And in early 2025, the company announced its own bot factory - 3i GPT-A Factory.

3iTech is implementing a number of pilot projects in the field of monitoring negotiations in critical areas, as well as emergency alerts. As for the latter, for example, AI can instantly inform people about natural disasters, fires or other emergencies, up to and including adjusting evacuation routes in real time.

The company supported the project of teaching children to read in Russian according to a natural model (reading at a sound landmark): based on the development of Candidate of Sciences Alexei Kushnir, a service was created youcanread.ru.^[11] A in the company's plans -^[12] introduction of smart assistants for schoolchildren into the learning process.

VK

VK focuses on the application of AI, both for the main VK products and for the development of commercial B2B services based on them. The implementation of AI-based services for external customers is carried out by VK Predict. VK commercial services solve the following tasks of corporate customers: generation of marketing creatives, categorization of requests and preparation of response templates for client support services, semantic search by knowledge bases, search for insights in data and interpretation of graphs and dashboards, identification of trends.

{{quote | In our practice, we demonstrate the effectiveness of an approach with a combination of the use of gene and audience data - on their basis we develop AI services for marketing, client analytics and building personalized communications, relying on knowledge about the interests, social parameters and other impersonal characteristics of the target audience, - noted Roman Styatyugin.}}

Lukomorye (Rostelecom)

Lukomorye is an ecosystem of import-substituted products that provide a full cycle of management of the company's business processes from software development to the work of service divisions. Lukomorye has its own no-code Akola platform, which allows you to develop and implement corporate sites, web applications or your own ITSM systems, excluding the involvement of third-party specialists and using only the internal resources of companies.

There is also its own ITSM system - ESMP, which includes the ESMP Button module. It allows you to create chatbots with a low entry threshold thanks to low-code tools, process natural language and integrate with popular instant messengers. Linguo's ESMP module optimizes case processing by recognizing not only text, but also images, including screenshots. The company's own development - artificial intelligence "Sirin" is an intelligent center for project management: it automates the collection and update of data from different systems, forms a knowledge base, and also generates answers and texts based on information analysis.

Kaspersky Lab

At the end of 2024, Kaspersky Lab released new versions of products using LLM. The technology based on the internal LLM infrastructure of Kaspersky Lab is a function of obtaining a brief summary of threats (indicators of compromise) in the OSINT tab in the Kaspersky Threat Intelligence Portal. These are brief structured squeezes based on a potentially large amount of text data from various reports and research on cyber threats for quick decision-making about the future actions of TI analysts.

KIRA (Kaspersky Investigation and Response Assistant) is an LLM assistant designed to facilitate and speed up the work of security center (SOC) analysts when working in Kaspersky SIEM (KUMA) and XDR products. At the moment, the functionality is implemented here, which allows you to quickly study the command line (including obfuscated) from the final device, receiving brief explanations of actions and detailed analysis of components.

Jet Infosystems

Jet Infosystems is launching a specialized GPT laboratory that will become a technology expertise center and a key platform for developing and testing LLM-based solutions. The company already has more than 60 hypotheses for the applicability of LLM for large businesses. On their basis, the creation of MVP and prototypes began so that solutions could be tested in the cloud or on the basis of customer infrastructure to assess economic effects.

VizorLabs

As Stanislav Ziganshin' said , for VizorLabs 2025 will be the year of the mining industry. One of the company's flagship products is the conveyor belt control system - the first solution in Russia that allows you to detect damage, delamination and other defects in real-time mode, preventing accidents and downtime. The product has already proven its effectiveness at industrial facilities.

Another significant project is the monitoring of the flotation process using computer vision, which is in demand in the gold mining industry. The system analyzes the surface of the pulp, controlling the distribution of bubbles and adjusting the parameters of the process: gold extraction is significantly increased, reagent costs are reduced and stability is increased.

In addition, VizorLabs offers solutions for controlling the loading of mine dump trucks (ShAS), automated granulometry and monitoring of technological processes. The company integrates multimodal AI into its products and takes them to international markets.

GC "Zyfra"

One of the last significant projects with AI of the Figures Group of Companies is the creation of a digital twin of a gas condensate asset for Achim Development.

{{quote | This is a unique solution, the first in the world, which allows you to track the operation of a gas condensate asset at all stages in real time, "said Mikhail Aronson, General Director of Zyfra Group of Companies.}}

The innovation of the digital model consists in automatic update and adaptation to change. The high speed of data processing allows the twin to stably display the current state of the asset without the need for manual adjustments. When new information comes from production, the model instantly adjusts to this data. As a result, the accuracy of the operating modes of the equipment has increased at the facility, the mechanisms for planning production and using the necessary resources are being improved, and a higher level of safety has been achieved.

BSS

BSS is actively developing comprehensive speech solutions for contact center automation. Recent developments include a comprehensive CC automation solution consisting of intelligent voice and text robots, a voice analytics system, and a modern knowledge base.

Key areas of development using generative AI: introduction of geneAI into speech analytics for processing complex natural language criteria and automatic identification of significant trends in client requests; Integration of RAG into the bot creation platform AI assistant in the knowledge base.

Now we are actively experimenting with agent AI to radically simplify the bot training process, - commented Alexander Krushinsky.

Bercut

Starting in 2024, Bercut is actively investing in the development of its advanced product - the corporate ESB Bercut integration bus - and the expansion of AI/ML products in general.

ESB Bercut is an intelligent data bus based on the hybrid integration platform HIP Bercut, which in 2024 became one of the TOP-5 most popular Russian platforms for integrating data and applications according to TAdviser.^[13] The intelligent integration platform ESB Bercut is the basis for integration and self-creation of message processing flows using the AI/ML methodology. The solution provides a single access point for data exchange between different applications, and also allows you to control the routing and transformation of data. The built-in AI/ML assistant will speed up development, help in route building, testing, writing and checking code, documentation and monitoring.

The AI module in ESB Bercut is a separate market product that will be available in two formats: On-Premise and PaaS (Platform-as-a-Service).

Sberbank Business Software

Maxim Ivanov highlighted four relevant AI cases of Sber Business Software: a software and hardware complex for monitoring the state of tomatoes in the greenhouse and reducing crop losses; forecasting passenger traffic and the number of stowaways for control; logistics of light petroleum products to reduce fuel transportation costs; AI assistant with Gigachat to handle customer calls.^[14][15]

NORBIT (LANIT)

Dmitry Demidov spoke about the development of a platform solution for word processing Norbit GPT, as well as an application solution AI Master Data for cleaning regulatory reference information based on Norbit GPT.

The target audience of AI Master Data is a large business that works with a large range of goods in procurement and sales. These can be industrial enterprises, retail companies, builders, miners and others. AI Master Data categorizes items, enriches reference records with attributes, eliminates duplicates and errors, and performs other data cleanup operations. The company begins to feel the business effect when it finds the necessary materials in its warehouse instead of re-ordering them due to errors in names, combines purchases into one large one and gets the best price, improves its forecasting functions.

HFLabs

HFLabs adapts existing neural networks and enhances their own solutions with their help, conducts experiments aimed at working with a gray zone in client data. For example, the AI solution of the Camouflage company depersonalizes personal data in documents for further safe use.

{{quote | As a rule, when working with large amounts of data, the participation of data stewards in complex and ambiguous cases is required - for example, to determine whether it is possible to combine two client records or whether the information in them, with all the similarity of the data, belongs to different people. In the future, AI could give clues when analyzing such situations. And although the capabilities of artificial intelligence are already enough to independently make decisions on such scenarios, so far the main blocking factor remains the question of who will be responsible in case of error, "said Nikita Nazarov.}}'

"Megaputer"

The latest developments of Megaputer include the creation of business solutions based on collaborative multi-agent AI. In this approach, different GenAI models work in conjunction, testing and refining each other's results, working sequentially, in parallel, or cyclically. Solutions use combinations of different AI models, including multimodal ones.

For example, visual language models (VLMs) are used to extract facts from large volumes of complex scientific and business documentation, which is a mixture of text, tables and graphs. Multimodal AI is complemented by the iterative application of OCR tools and classic NLP tools, which perform primary data processing, reduce hallucination levels and correct model performance, which allows you to increase the completeness and accuracy of the results obtained, even in tasks with a complex context, by at least 50% compared to the use of GenAI alone.

This technology allows you to solve complex business problems that cannot be solved with one LLM. The solution operates on the basis of the low-code analytical system PolyAnalyst, which supports work in a single graphical interface without programming. This allows analysts with low software development skills to clearly see the structure of the solution being created, quickly modify and supplement the system with new models. The platform provides a single point of access to various AI models: you can use both general-purpose AI models operating in the cloud, and additionally trained on customer data located in a closed loop.

Шаблон:Quote 'author=Dmitry Goltsov, Deputy General Director for Commercial Affairs of Megaputer

Webiomed

Webiomed is an AI-based system that helps identify health-threatening scenarios, assess the potential for developing complications of various chronic diseases. The platform's functionality includes the collection and processing of impersonal medical data, the extraction of features using AI technologies and their automatic verification and cleaning, the formation of a patient's digital profile, the creation of datasets based on profiles based on specified parameters, as well as functionality for the automatic interpretation and analysis of formed profiles, including using AI technologies and predictive analytics. In 2024, a management analytics module for industrial enterprises was added to Webiomed. He promptly informs managers about the health of employees: about groups and risk factors, provides information on the control of preventive measures, gives forecasts for health reasons.

SberMedII

The company's portfolio includes AI solutions for processing visual studies: "CT stroke," "CT chest," "Mammography," "X-ray of the knee joint." In the automatic mode, digital services highlight signs of pathologies with a color contour in the pictures, calculate the total volume of the lesion and make a preliminary description, determine the presence of signs of neoplasms in medical images.

Together with the Sberbank Artificial Intelligence Laboratory and the Moscow government, the company has developed services for processing medical text data. The system for supporting medical decision-making based on AI "TOP-3 diagnosis" helps in making preliminary diagnoses based on patient complaints and is used in medical organizations in 19 regions. The AIDA Digital Assistant (AI Diagnostic Assistant) provides a second independent opinion on the final diagnosis based on the patient's EMR data obtained over two years. The system is piloted in all adult polyclinics in Moscow and the Lipetsk region.

The next step may be the introduction of a digital medical assistant based on GigaChat into the work of medical organizations. This is one of the first generative models in the world to pass the exam in the areas of "Medical," "Cardiology," "Neurology" and "Pediatrics." An intelligent chatbot can take on the following tasks: collecting patient complaints (history) before admission with an assessment of the completeness of the data and the formulation of additional questions to reduce the burden on the doctor at the initial appointment and reduce the time to collect history; routing the patient on the recommendation of AI: the patient does not go to the therapist, but immediately gets to the right specialist; recommendation of assays and decoding of downloaded assays using AI.

The company plans to further develop and implement a new multimodal medical assistant, GigaDoc. The service was developed jointly by the Center for the Health Industry of Sberbank, scientists from the Sberbank Artificial Intelligence Laboratory and SberMedII. His work is based on technologies of remote photopletismography, computer vision and machine learning, which make it possible to assess some medical parameters (including body mass index, risk of type II diabetes, indicators of the state of the cardiovascular system, cholesterol level), accelerate primary diagnosis of health conditions and qualitative monitoring of chronic non-communicable diseases.

As Alexey Lyubimov,' CEO of 3iTech, said in an interview with TAdviser, the market is just being formed, and all players are looking for niches in which AI will give real savings or tangible additional profit. 2024 was the year of tests, and 2025-2026. will be the time of mass implementation.

If replacing people with AI gives savings - business will replace them. In Russia, this is especially important: the shortage of personnel even for simple operations pushes companies to automation, and in large organizations, saving even a share of a percent on each operation as a whole gives a huge economic effect.

According to the expert, it is necessary to take into account the humanitarian aspect of the widespread penetration of AI: if it takes on most of the tasks, the motivation for learning is reduced. So, for example, there is already a noticeable decrease in the importance of the role of a living teacher, analyst, psychologist, etc.

The TAdviser analytical center conducted a study of approaches to attracting and hiring talents in the field of AI and prepared a rating of key employer companies for AI in Russia.

According to TAdviser, in 2024, 90% of the top 100 largest companies in Russia from different sectors of the economy use machine learning (ML) and artificial intelligence technologies for internal business tasks, or for developing commercial products for the external market. The development of AI initiatives and the involvement of specialists in machine learning and AI technologies is becoming one of the key priorities in the framework of the national project "Data Economics" launched in 2025More...

Related Materials:

• Intelligent document processing. TAdviser Overview
  
• Artificial intelligence (AI, Artificial intelligence, AI)
  
• Artificial intelligence technologies. TAdviser Overview
  
• Artificial Intelligence (Russian market)
  
• Artificial Intelligence (China Market)
  
• Artificial intelligence in the United States
  
• Artificial intelligence in HR (recruitment, motivation and development of personnel)
  
• Artificial Intelligence in Manufacturing
  
• Artificial intelligence in industry
  
• Artificial intelligence in construction
  
• Artificial intelligence in agriculture (agro-industrial complex)
  
• Artificial intelligence in transport
  
• Artificial intelligence in logistics
  
• Artificial intelligence in retail
  
• Artificial Intelligence in Audit
  
• Artificial intelligence in banks
  
• Artificial intelligence in the courts
  
• Artificial Intelligence in Public Administration
  
• Artificial intelligence in military affairs
  
• Artificial Intelligence in Medicine
  
• Artificial intelligence in Russian medicine
  
• Artificial intelligence in radiology
  
• Artificial intelligence to solve demographic problems
  
• Artificial intelligence for writing texts in the media and literature
  
• Artificial intelligence in video
  
• Artificial intelligence in photography
  
• Artificial intelligence in painting
  
• Artificial Intelligence and Music Creation
  
• Artificial intelligence in the fashion industry
  
• Artificial intelligence in telecom
  
• Artificial Intelligence in Education
  
• Artificial intelligence in science
  
• Artificial intelligence in sports
  
• Artificial intelligence in crime