George Sivulka, founder of the Hebbia startup, noted that many organizations faced management challenges when implementing AI agents due to the lack of the necessary infrastructure to control their work and costs. According to him, companies mistakenly interpreted the growth in the consumption of computing tokens as an indicator of successful digitalization and increased overall productivity. In practice, this has led to the fact that artificial intelligence has not so much reduced labor costs as fundamentally changed the structure of operating costs, making software a more expensive resource compared to human capital. This practice, dubbed "tokenmaxing" in the industry, forces organizations to rethink approaches to evaluating the effectiveness of automated systems and stop encouraging meaningless activity.

The main reason for the inefficient operation of algorithms is not the technical limitations of the models themselves, but rather vague internal instructions and a lack of relevant context. The expert emphasizes that only a small proportion of employees have the skills to formulate requests that ensure high-quality fulfillment of assigned tasks. This leads to cyclical processes when the agent repeatedly accesses the model to adjust its own actions, spending additional computing resources without a commensurate commercial result. As a result, corporations are facing a sharp increase in bills from AI service providers, which forces management to introduce strict internal limits and warning systems for exceeding consumption thresholds.

As a solution to the problem, experts suggest not the abandonment of technology, but the formation of an integrated management infrastructure for agent systems. This includes a detailed description of business processes, the introduction of metrics for evaluating the quality of results, and the application of a model routing strategy in which simple operations are delegated to more cost-effective algorithms, and advanced models are used exclusively for key scenarios. In addition, successful integration requires overcoming internal barriers: employees may be reluctant to transfer critical work knowledge to AI systems for fear of losing their own professional significance. Therefore, building trusting relationships and establishing clear rules of interaction are becoming prerequisites for the sustainable development of corporate artificial intelligence.