The main architectural trap is to “sew” AI directly into existing systems / Economic news of Krasnoyarsk and the Krasnoyarsk Territory / Newslab.Ru

The risks of introducing AI into production IT systems lie not in the technologies themselves, but in the architecture of the solutions. Improper integration of tools into an existing enterprise framework can lead to vendor lock-in, loss of accumulated data, and inefficient use of expensive computing resources. This was stated by Sergey Golitsyn, head of T1 AI (IT holding T1) at CIPR-26.

Speaking at the session “Made in Russia. Smart ERP, MES, CRM for a new stage of digitalization of production,” Sergey Golitsyn emphasized that the main architectural trap is to “sew” AI directly into existing ERP, MES or CRM systems. In this case, the logic of the models and prompts become strictly tied to a specific product, and a potential change of IT supplier turns into an expensive and risky project.

The second difficulty is the inability to reuse AI solutions in different circuits. The same model may be needed by production, logistics, and service, but with an integrated approach it works within only one vertical. The business actually loses its accumulated intellectual capital, since any major restructuring of the IT landscape results in the loss of trained models and scenarios that have been created over the years.

The third risk relates to the management of computing resources. Models for predictive analytics, computer vision, and text processing require expensive GPU clusters. If each system – ERP, MES, CRM – tries to manage them in its own way, which provokes an increase in costs.

“If you try to manage GPU clusters from each individual system, the result will be extremely expensive and inefficient. It would be more correct to bring this to the level of a single AI platform, through which all business applications access models and resources according to the same rules,” notes Golitsyn.

The expert emphasized that Russian industry is at a point where a showcase of cases has already been formed. There are already quite a lot of solutions in which neural network models work inside ERP and MES. Among the most common scenarios, Sergei Golitsyn identified three: flaw detection, industrial safety monitoring and predictive analytics, which allows one to predict, among other things, the timing of equipment failure.

“We see a growing demand for such tools, which is explained by the rapid financial effect of their implementation,” noted Sergei Golitsyn.

The increase in the number of such implementations takes industrial automation in Russia to a new level. Over time, as Sergei Golitsyn notes, the system will be able not only to carry out a given scenario, but also, within certain limits, to choose it itself, but this is not the prospect of several years, but a much longer story.