The corporate world has encountered a paradox: the widespread adoption of AI agents has not led to the expected productivity growth, but has instead spawned a new class of managerial and financial problems. Hebbia startup founder George Sivulka, in his essay for the a16z newsletter, diagnoses a systemic error — companies rushed to deploy agents without adequate control infrastructure, and are now reaping the consequences of this haste.

The key problem, in his view, lies in the practice of "tokenmaxxing" — the pursuit of the number of tokens consumed as the main indicator of activity. Organizations mistakenly took the explosive growth in AI spending as a sign of effective implementation. Sivulka puts it bluntly: "You just hired a million bad employees." And this is not a metaphor. Agents have not so much reduced labor costs as changed their structure, and for the worse. For the first time in history, as the expert notes, people have become cheaper than software.

A managerial crisis, not a technological one

Sivulka's analysis shows that failures in AI agent operations are often linked not to model imperfections, but to companies' inability to formulate clear tasks. Only "roughly one in 100 employees" can provide an agent with enough context for quality work. The result is "loops": agents endlessly adjust their actions, overspending tokens without proportionate results. This is, in essence, "spending tokens on spending tokens." We are witnessing not a technological, but a deep managerial crisis.

The financial consequences have been shocking. According to UBS Global Research, virtually all top managers of AI companies at closed events acknowledge the problem of cost control. One example — expenses on Anthropic at an unnamed firm grew from $20,000 in December to nearly $1 million in July. Another case, mentioned in industry sources, is a company that launched AI tools without limits and received a $500 million bill from Claude in a single month. UBS estimates that about 60% of organizations are already seriously concerned about uncontrolled token spending.

New managerial infrastructure as a necessity

The solution, according to Sivulka, lies not in abandoning AI, but in creating a managerial infrastructure: clear process descriptions, quality assessment of results, cost control, and, crucially, task distribution among models of varying costs. This refers to model routing — where simple operations are handed off to cheap models, while expensive advanced ones are used only for critically important scenarios.

However, there is also a hidden threat. Employees, realizing that their knowledge is a key resource for training AI, may begin to sabotage the process. People will not want to hand over to systems what makes them irreplaceable. Without trust and transparent rules, workers will not help those who could replace them. This turns AI adoption from a technological task into a highly complex social and managerial problem that no one has yet solved.