Foreign media reports that after companies began to aggressively adopt AI agents, the market is shifting from "using models more" to "how to manage models effectively." Fortune, citing an article by George Sivulka, founder of AI company Hebbia, states that the previous wave of enterprise-level AI agent expansion was too rapid, but most companies did not establish corresponding management systems, resulting in increased costs, repetitive processes, and in some scenarios, even less stable than human collaboration.

Companies expand first, then improve management.

Sivulka calls this brief surge "tokenmaxxing." His core argument is that companies haven't bought more efficient digital employees, but rather introduced a large number of agents with inconsistent execution quality all at once. The problem isn't just with model capabilities, but more importantly, that few people within the company can clearly explain, break down, and provide sufficient context for tasks.

The article mentions that Hebbia's clients include BlackRock, KKR, and the U.S. Air Force. Sivulka argues that many companies, without budget constraints, process delineation, and reporting mechanisms, directly integrate AI agents into their workflows, amplifying existing organizational inefficiencies. Once the agent receives ambiguous instructions, it repeatedly calls upon the model for self-correction, resulting in high-frequency resource consumption.

Cost pressures are the first to surface.

Fortune cited a survey by UBS Global Research, stating that many AI-native companies privately acknowledge that one of the most pressing concerns for enterprise clients is token costs. Some companies previously lacked a clear concept of token budgets, but clients have begun to persistently inquire about spending control.

UBS, citing an unnamed AI company, reported that its spending on Anthropic has surged from $20,000 in December to nearly $1 million in July, a roughly 50-fold increase in seven months. Despite this, the company has not asked employees to stop using the model; instead, it has begun setting monthly threshold reminders and restricting the use of cutting-edge models for certain administrative positions.

Public statements are also increasing. The article mentions that OpenAI has stated that AI costs have become a prominent issue this year, and Uber has begun setting up spending barriers. UBS estimates that approximately 60% of organizations now consider token costs a real concern, and this percentage may be rising.

Enterprises are starting to switch to model routing

Sivulka believes that what's truly lacking right now isn't more models, but more mature management methods. He emphasizes "context engineering," which enables employees to more accurately define tasks, add context, and assign different subtasks to more appropriate models.

UBS research shows that some companies have begun to adopt a "model routing" approach to project management. Instead of assigning a single task entirely to the same model, they allocate different stages to different models to balance effectiveness, speed, and cost. For scenarios requiring human-like interaction or key business results, companies retain cutting-edge models; for standardized processes, they tend to use cheaper or faster models.

Following this line of thought, the focus of competition among companies may no longer be solely on the amount of computing power they purchase or the number of models they integrate, but rather on who can establish the capabilities for budgeting, division of labor, and task decomposition first. The article also mentions that as companies hope to entrust more internal knowledge to AI systems, employees' concerns about knowledge retention and job replacement may further increase.