Companies have rushed to deploy AI agents without the necessary infrastructure to monitor their performance and costs. This conclusion comes from George Sivulka, founder of the startup Hebbia, in an essay for the a16z newsletter, as reported by Fortune.
He linked the issue to "tokenmaxxing"—a practice where organizations tracked and encouraged increased token consumption when working with AI. According to him, companies mistakenly interpreted rising token usage as a sign of AI implementation and productivity gains.
“You just hired a million bad employees,” Sivulka wrote.
In his view, AI agents have not so much reduced labor costs as altered the structure of expenses. He stated in his essay that “for the first time in history, people have become cheaper than software.”
Hebbia develops corporate AI tools. According to Fortune, among its clients are BlackRock, KKR, and the U.S. Air Force.
The Problem Is Not with the Models
Sivulka estimates that AI agents often fail not due to weak models, but because of unclear instructions within companies. He claims that “about one in 100 employees” knows how to provide AI with enough context for quality work.
Poor instructions lead to “loops”: the agent tries to correct its own actions, repeatedly consults the model, and consumes additional tokens without proportional results. Sivulka described this as “token consumption for the sake of token consumption.” Ultimately, he concluded that companies are facing a management crisis rather than a technological one.
Fortune connected this thesis to a broader discussion about AI spending. According to the publication, at the beginning of 2026, several companies tracked and encouraged token consumption as a measure of employee activity.
Later, some of these initiatives were rolled back. Previously, the publication reported that Meta and Amazon used internal ratings of token consumption but abandoned them after employees began using AI for metrics rather than results.
Costs Have Become a Separate Risk
Fortune cites data from UBS Global Research, indicating that nearly all executives from AI companies at a private bank event discussed the issue of token expenses in the corporate environment.
One unnamed firm reported to UBS that its expenses on Anthropic rose from $20,000 in December to nearly $1 million in July. Management did not completely restrict usage but began implementing internal limits and warnings about exceeding thresholds.
According to journalists, UBS previously assessed concerns about token expenses as a significant issue for about 60% of organizations. The publication also mentioned an unnamed company that, according to Axios, received a bill of approximately $500 million in one month for using Claude after launching AI tools without limits.
Companies Are Shifting to Model Management
Sivulka believes the solution lies not in abandoning AI but in establishing a management infrastructure for agent systems. This includes clearly defining processes, evaluating quality outcomes, controlling expenses, and distributing tasks among different models.
Fortune reports that some companies are already transitioning to model routing. In this approach, simple operations are assigned to cheaper or faster models, while advanced, costly models are reserved for key scenarios.
Sivulka also warned of a new internal issue related to employees retaining essential work context. He noted that individuals may be reluctant to share knowledge and processes with AI systems that could render them irreplaceable. The expert emphasized that without trust and clear rules, workers will be unwilling to assist systems that could change their roles.
In July, Financial Times columnist Sarah O’Connor referred to employee knowledge as a resource for corporate AI.