People and leaks: why domestic business refuses to implement AI
The introduction of artificial intelligence capabilities in Russian companies has faced a paradox. On the one hand, 59.6% of organizations are already actively using neural networks. On the other hand, 42.5% are afraid to scale these technologies because of the risks of confidential data leakage. The market is actually divided into two camps: those who build closed IT circuits, and those who "feed" their data to public clouds. Why business is so careful and how it tries to protect its secrets is in the Izvestia article.
Public neural networks and the shadow use of AI
Deploying local language models and secure infrastructure contours requires significant resources. Companies that cannot afford local hardware and software systems are forced to use public services, where data can be spent on upgrading third-party algorithms by default.
High risks in the field of data security and confidentiality are the main reason for businesses to abandon the introduction of artificial intelligence. At the same time, about 40% of IT specialists reduce risk using a hybrid model, when AI only performs the role of an assistant, leaving control, correction and final approval of decisions to a person, according to a joint study by the UCSB system integrator and the Solar group of companies.
According to the authors of the study, which involved 102 companies, among which there are those who do not yet use AI, the main reason for the failure is high risks to data security and confidentiality, including leaks of source code and commercial information (42.5%). In second place is the lack of competencies within the company — 35%. Another 32.5% of respondents indicated the lack of clear and economically sound scenarios for the use of AI in business. The high cost of implementation (20%) and regulatory uncertainty (25%) play a less significant role.
At the same time, according to the survey, about 59% of companies already use various neural network models. One third uses large language models, and another third does not use such tools yet, but plans to implement them in the near future. And only 3% of organizations are ready to entrust artificial intelligence with independent decision-making. The most common model remains the AI assistant format: in 40% of companies, neural networks are used to analyze information and prepare recommendations, but the final decision is still made by a human.
The choice between public and private AI systems directly depends on the industry. The "dangerous" sectors that require local models include banks, the public sector, energy, industry, medicine and telecom, while for "less sensitive" areas such as trade, media, marketing, logistics, education and HR, it is acceptable to use public services provided that confidential information does not get into them..
Leaks and the human factor
The main threat is the uncontrolled transfer of data to neural networks by employees. According to the study, the volume of confidential information going into public neural networks increased 30-fold in 2025. At the same time, 60% of companies do not have a clear policy of using AI: commercial offers, contracts, financial reports and personal data of clients regularly get into prompta.
An additional threat is posed by "external" agent-based AI systems — voice and text assistants for customer support, automation of B2B sales and lead management, IT support and infrastructure management, supply chain and logistics management, financial and insurance agents, multi-agent corporate systems and external platforms for automating business processes. According to the research, 62% of companies are testing such systems today. And in case of compromise, attackers can gain access to the entire IT infrastructure of the organization. But a complete ban on neural networks in business does not work.
— It is important for businesses to use local solutions that operate in a closed loop on their own equipment, without Internet access. This ensures confidentiality and compliance with regulatory requirements," says Sergey Golitsyn, Head of T1 AI.
For this purpose, local models are used — AI systems deployed on the company's own servers for full control over data and security. RAG (Retrieval-Augmented Generation) technologies are also used, which allow AI to search for and use up-to-date information from external databases, which makes the answers accurate and up-to-date. Compact AI is also used — small neural networks that require few resources.
Why classical protection systems do not save
Data leak prevention and access control systems reduce risks, but do not eliminate them completely: it is impossible to revoke data that has already been used to train a public model. In the context of working with artificial intelligence, these systems serve as multi-level protection, ensuring that sensitive information does not leave the protected perimeter and is accessible only to authorized AI components. However, because of this, the level of developers' trust in cloud-based AI tools has fallen from 40% in 2024 to 29% in 2025, according to T1 data.
To protect corporate models, experts recommend isolating neural networks at the network level and implementing special filters — software mechanisms that analyze and control data flows at the input and output of the model. These filters serve as a barrier that prevents the model from processing, storing, or transmitting unwanted or confidential information.
— Special filters protect against hidden malicious instructions that intruders "hide" in prompta. This is a basic part of AI cybersecurity," explains Onlanta product manager Dmitry Vasiliev.
The main problem lies in the complexity of total control over the actions of personnel and the need for strict separation of scenarios for the use of neural networks. This means that for each specific business task or for each department of the company, there should be an isolated and strictly limited use case for artificial intelligence in order to minimize risks and prevent mixing of data and authority.
In addition to total control, other approaches are used to protect corporate neural networks. The key is the "architectural" separation, in which isolated AI instances are created for different tasks and departments that do not have access to other people's data. In addition, mathematical "noise" (the so-called differential privacy) is added to the training data so that the model cannot remember and give out specific confidential information. Strict separation of access rights for users and logging of all their actions is also applied, which allows incidents to be investigated and risks associated with the human factor to be minimized.
Переведено сервисом «Яндекс Переводчик»
