The IT holding T1 told how the use of AI changes the work of programmers
Understanding the code created by neural networks became the main difficulty when working with such tools: this problem was noted by 42% of respondents to the study of IT holding T1 and Tproger

Understanding the code created by neural networks became the main difficulty when working with such tools: this problem was noted by 42% of respondents to the study of IT holding T1 and Tproger. This problem is more common than model errors.
The results were presented on November 19 at the Impulse T1 conference in Moscow.
In addition, 39% of participants said that it takes longer to review and review AI-generated code. 36% said that the models do not take into account the existing code base and internal rules of the team, so developers have to adapt the result to the project.
The most acute problem is manifested when working with AI agents. Among those who pass the task to the agent as a whole, 51% of respondents reported difficulties with parsing written code; among those who use neural networks mainly for obtaining advice, this is 31%. At the same time, 59% of those who request individual fragments from AI face errors and “hallucinations”, while among users of full agents the share of such complaints is significantly lower – 17%.
A separate complexity is the organization of processes. Many companies do not yet have a systematic approach to what tasks can be assigned to neural networks, how to check the generated code and who is responsible for it. 43% of participants said that there were no clear rules, and only 10% reported mandatory verification of the results.
Sergey Polinenko, product leader of the Sphere platform of IT holding T1, explained: according to the study, more than half of respondents expect AI to take over the bulk of the work, and the role of developers will be reduced to setting tasks and quality control. According to him, it is important for businesses to bring products to the market faster without losing quality. To do this, you need to move from one-time solutions to end-to-end application of AI at all stages of development, taking into account security requirements - this will make vibcoding truly reliable.
The study also shows high engagement: 73% of developers turn to such tools almost every day, and 78% said they coped with the latter task faster than with a similar one without a neural network. At the same time, only 8% have undergone corporate training – most often, specialists master tools independently, within their own projects.
The survey was conducted by IT holding T1 and the publication Tproger. It was attended by 225 people (detailed questionnaire) and more than 11,000 (voting on social networks). Among the respondents are developers, team and project managers, specialists in DevOps, infrastructure and architecture, testers, analysts and aspiring developers. Data collection took place in October 2026.

