Hebbia CEO says AI tokenmaxxing costs expose weak agent management
George Sivulka says companies let AI agents spend too freely without the management systems needed to make them useful.
By Sofia Marchetti · World Affairs Correspondent
3 min read
AI tokenmaxxing costs are now a management problem for companies that rushed to deploy agents, according to George Sivulka, the 27-year-old Stanford graduate who founded Hebbia. In an essay published by Andreessen Horowitz’s a16z newsletter, Sivulka argues that businesses added AI labor before creating the controls needed to direct it.
Fortune reported that Hebbia works with clients including BlackRock, KKR and the U.S. Air Force, giving Sivulka exposure to large-company AI rollouts. His warning to executives was blunt: in his view, firms effectively “hired a million bad employees” by giving agents broad access and too little oversight.
What is tokenmaxxing?
In Sivulka’s framing, tokenmaxxing was the brief rush to spend heavily on AI agents in the hope that more automated work would mean lower labor costs. He says that bet backfired because poorly directed agents can burn through usage without producing useful results.
Sivulka compares the moment to the railroad boom of the 1830s and 1840s. He wrote that U.S. track mileage rose about 120-fold in a decade before an 1841 fatal crash in Massachusetts forced railroads to build modern management systems, including clearer roles and reporting lines.
He argues AI agents have created a similar need inside companies. In his essay, Sivulka says agents can scale bad instructions instantly, and that only about “1 in 100 employees” knows how to give AI enough context to complete work well.
Why are AI agent costs rising?
Sivulka says the problem is less about the number of tokens used than whether workers know how to use them well. He describes “looping,” where agents repeatedly call themselves to fix vague or flawed directions, as a costly pattern of spending tokens to spend more tokens.
UBS Global Research reported similar concerns after hosting AI-native companies at its 5th Annual UBS Private AI and Software event in Menlo Park. One company executive told UBS that internal engineers had not been trained to think in terms of token budgets, while customers were asking about the issue in nearly every conversation.
UBS also cited one unnamed AI firm whose Anthropic spending rose from $20,000 in December to an expected $1 million in July. The firm told UBS it was adding monthly alerts and internal governors, including limits on general and administrative staff using frontier models.
Other companies have also moved to control spending. UBS estimated that token-cost concern had become a real issue for about 60% of organizations, while OpenAI has said AI costs became a major issue after not coming up at the start of the year. The Washington Times reported that Uber installed spending guardrails for internal AI coding tools.
Palantir CEO Alex Karp has made a related complaint. On CNBC’s Squawk Box, Karp said AI labs had oversold their models and criticized enterprises for wasting time and money on tokens without enough return-on-investment discipline.
What fix does Sivulka propose?
Sivulka says companies need better management of AI work, including what he calls context engineering. His phrase for the next advantage is the “100x token,” meaning well-directed AI usage that can do far more valuable work at scale.
UBS said some AI firms are already moving in that direction through model routing, or assigning different parts of a task to different models. One executive told UBS that firms can cut costs by knowing which models are best suited to specific work and by using cheaper or faster models for routine tasks.
Sivulka also warned that companies may face resistance from employees who do not want to hand over institutional knowledge to AI systems. He pointed to Meta as an example, saying some equity-holding employees have pushed back on the use of their work context as training data.
This story draws on original reporting from Fortune.