Nazariy Popov, Senior Director of AI Engineering and Implementation at EPAM Ukraine, claims that predictions about the rapid replacement of developers by artificial intelligence have not come true. According to him, AI transformation is happening as an evolution, not a one-time revolution, and the first wave of hype is giving way to a sober calculation of payback.
Nazariy Popov, Senior Director of AI Engineering and Implementation at EPAM Ukraine, claims that predictions about the rapid replacement of developers by artificial intelligence have not come true. According to him, AI transformation is happening as an evolution, not a one-time revolution, and the first wave of hype is giving way to a sober calculation of payback.
The main mistake of optimists, according to Popov, is in a simplified idea of the developer’s work. Directly writing code makes up only a small percentage of the budget and terms of software creation. AI works well in greenfield projects, that is, from scratch, when the requirements can be described briefly and clearly. However, according to his estimate, more than 90% of budgets in global IT fall on brownfield: existing systems with a 10–15-year history, where a new function must fit into integrations, meet cybersecurity and GDPR requirements, and take into account outdated code. AI is not yet capable of reading such a context without human help.
Another problem: customers at the start often do not have a clear vision of the product themselves. Popov suggests imagining an ideal AI model and the task of creating a system at the level of the Rozetka marketplace: describing all the functionality accurately is extremely difficult. Formulating the right task for the algorithm, according to him, is 80% of the project’s success, and this work falls on the person. Therefore, the Human-in-the-Loop principle is increasingly being talked about: AI takes on the writing of code and routine, and the person understands the business, sets tasks, is responsible for the architecture and control of the result.
The market, says Popov, is beginning to realize that fully autonomous AI (AI-Autonomous), when agents independently perform tasks through thousands of attempts and iterations, is, first of all, expensive. As an example, he gives the task of raising the history of correspondence and writing a summary letter. An autonomous agent can “burn” tokens worth tens of dollars for one letter, while a person with an AI-assistant (AI-Assisted) will do it in three minutes with a cost of a few cents. A semantic layer like Second Brain helps to make the agent’s work cheaper, but even for local tasks this is a non-trivial task.
He cites the EPAM project to automatically migrate a legacy system from Cobol to Java as an example. Agents were run in autonomous mode to iteratively rewrite, test, and fix the code. The project was successful, but the bill for tokens alone exceeded the cost of the engineers who supported the process. Despite this, the total cost was still significantly lower than for manual migration. Popov’s conclusion: autonomy can pay off, but tokens become a separate, large expense item that needs to be planned in advance.
According to his assessment, a significant decrease in the price of tokens in the near future is not worth expecting. This can be compensated for by the effective use of tokens and AI-oriented architectures and processes. A business that expects to completely replace people with algorithms risks encountering operational and financial barriers. Popov recalls Gartner’s forecast: by 2029, businesses will have to rehire almost a third of employees laid off due to AI, but on much more expensive terms.
Popov believes that AI does not replace the IT industry, but accelerates its evolution. When writing code becomes cheaper, development does not disappear, but scales: the cheaper "kilogram of code" makes custom development available to thousands of new companies, so the market grows, not falls. He connects this with the Jevons paradox.
