Sometimes the best AI pilot outcome is not creating AI agents

Sometimes the best outcome of an AI pilot is deciding not to create AI agents.

Speaking at TNF, Anton Rubenchik, CIO of SIBUR, shared a curious observation: if you seriously and thoroughly describe the rules of a process, some of the planned AI agents will turn out to be unnecessary.

This sounds almost paradoxical, because in the industry, AI leadership is often judged by the number of launches. But SIBUR follows a different logic. The company actively experiments with agents, tests new scenarios, and moves towards greater autonomy. However, the result of a pilot is evaluated by what has changed in the work: whether there are fewer manual operations and repeated approvals.

The pilot itself also works as a diagnostic tool. While the team explains to the agent where to take data and what rules to follow for decision-making, gaps between systems, undocumented exceptions, and habits that held the entire process together come to light.

After such a breakdown, it sometimes becomes clear that the task is cheaper and more reliable to solve without an LLM — by fixing an integration, removing an extra approval, or recording a rule as a standard algorithm.

The agent didn't go into industrial production, but the process became shorter and clearer. A perfectly worthy result of the experiment.