George Sivulka highlights management challenges with AI agents, emphasizing the need for better infrastructure and processes to control costs and improve efficiency.
Businesses are facing significant hurdles in managing AI agents effectively, as they often lack the necessary infrastructure to monitor performance and costs, according to George Sivulka, founder of the AI startup Hebbia. In an essay for the a16z newsletter, Sivulka critiques the trend of "tokenmaxxing," where companies mistakenly view increased token usage as an indicator of successful AI implementation and productivity.
Sivulka warns, “You just hired a million bad employees,” emphasizing that AI agents have not necessarily cut labor costs but have instead changed the nature of company expenses, suggesting that people have become less expensive than software.
Hebbia, which develops AI tools for corporate clients like BlackRock and the U.S. Air Force, points out that AI failures are often rooted in poor internal directives rather than flawed models. Sivulka estimates that only about 1% of employees understand how to provide AI systems with the necessary context for effective performance, leading to inefficient "loops" where AI agents waste resources trying to rectify their own mistakes.
In a broader context, discussions at a recent UBS Global Research event revealed that nearly all executives from AI firms expressed concerns about rising token costs. One company reported expenses on AI services skyrocketing from $20,000 to nearly $1 million in just seven months, prompting management to impose internal limits on usage.
Furthermore, a separate report highlighted a company that received a staggering $500 million bill in one month for AI usage, underscoring the financial risks associated with unchecked AI deployment. Sivulka argues that the solution lies not in abandoning AI but in developing a robust management framework that includes clear processes, quality evaluations, and expense controls. Some companies are beginning to adopt model routing, assigning simpler tasks to cost-effective models while reserving advanced tasks for more sophisticated systems.
However, Sivulka cautions that employees may hesitate to share critical work knowledge with AI systems, fearing job displacement. He stresses the importance of building trust and establishing clear guidelines to encourage collaboration between workers and AI technologies.
