Roman Smirnov noted that a number of solutions are built on open-source models,

[tanjimajuha20]

Maria Bar-Biryukova names three risk factors at once: "Firstly, not all companies can ensure data security when implementing AI technologies. Big data in e-commerce is aggregated from a large number of different systems and on different platforms - this can complicate their collection, storage, processing, use and protection. Secondly, companies, especially small ones, may not have enough resources to invest in such technologies. So far, with the exception of the simplest descriptive generative tools, this is still an "expensive toy". Finally, the mass implementation of AI in the future may entail changes in pakistan whatsapp number database teams: the simplest positions, such as data entry operators, may simply disappear. The e-commerce sector is one of the most actively developing industries, including from the point of view of the labor market. Therefore, managers need to be prepared to combine AI technologies with the work of individual employees."
and businesses should understand that everything that goes into the neural network trains it, and as a result, the risk of leakage exists and will be a weak point of the technology for some time. But this, according to the expert, is not the only possible problem: "Many people underestimate the issue of regulating interaction with AI. Along with the capabilities of language models, the number of threats and methods of application is growing. Today, various regulatory legal acts and regulators are only being limited and described, which imply liability for illegal actions using AI technologies. Another factor is the difficulty in understanding the result. Some developers say that if neural networks are trained by people, then these neural networks will somehow rely on human experience and interpretation of concepts. It is important to understand that absolutely all data requires additional processing and verification. Many forget that AI is a tool that employees need to be trained to use. Inform about risks and opportunities, show effective mechanics and methods of application, test the result and conduct fact-checking of information and data from LLM. This is a large, complex job."

Andrey Rybintsev reminds that the use of neural networks requires significant computing resources, which can be obtained by purchasing equipment or renting from cloud services: "For machine learning models to work effectively, it is necessary to have high-quality and extensive data sets that will be used to train algorithms. This can be compared to building a strong house: without a high-quality foundation in the form of correctly selected and structured data, any innovative artificial intelligence system will be unstable. It is also important to take into account the specifics of the industry and the needs of users when choosing and setting up machine learning algorithms. This is similar to choosing tools for a workshop: each must be precisely tailored to the task so that the final product is not only effective, but also in demand among the target audience. An acute problem is the shortage of qualified specialists in the field of artificial intelligence, capable of not only developing and setting up algorithms, but also applying them in the context of a specific business. For example, at Avito, as the scale of the company and the number of employees grew, we faced the dilemma that we simply could not find a sufficient number of employees with the qualifications we needed on the market. Therefore, we decided to start systematically sharing the accumulated applied knowledge in the field of machine learning through the education of young specialists and internship programs - we launched our own master's program at MIPT and plan to begin cooperation with 11 more key specialized universities by 2028."

Maxim Melsitov sees user data leakage as the main risk: "The main risk is the potential leakage of user and transaction data, as well as the complexity and high investment in training commercially available models on our own data. We see a way out of this situation in the use of the RAG (Retrieval-Augmented Generation) architectural model - an alternative approach to data processing in LLM. Companies also often face a shortage of highly qualified personnel, a lack of resources for upgrading information systems and infrastructure, as well as experience in implementing AI and, as a result, a lack of understanding of all the possibilities of its application. In such cases, they turn to large integrators who have expertise in this area and the necessary specialists to solve their problems."