This is a post recently shared on LinkedIn by George Sivulka, founder and CEO of Hebbia. For those leading AX transformation—whether fast‑follower CEOs or leaders driving AX change within their companies—the content is valuable, so I’ve summarized the key points for you.
The article is titled “Man vs. Token,” and it starts with a provocative line: “For the first time in history, humans have become cheaper than software.”
Many predicted that AI would replace human jobs, but the reality is the opposite. If the AI spend per employee at the top 1 % of companies continues to grow at the current rate (14.1 % month‑over‑month), it will soon exceed even the salary of a software engineer ($192 K) and the average wage ($98 K). Yet, companies that have aggressively adopted AI have actually increased hiring (+10.2 %). In other words, AI is not eliminating jobs; it’s creating more.
The author's way of unpacking this situation is intriguing. He draws an analogy to the railroads of the 1830s. When railroads were expanding rapidly, a major collision occurred because train operations were not properly coordinated, and in the aftermath, regional managers and a clear reporting structure emerged. The concept of 'management' as we know it today was born then. Whenever a new technology appears, we initially get excited about the technology itself, but ultimately it is systematic management that makes it run reliably. AI agents are now standing at the same crossroads.
His central point is this: AI agents fail in the same way people mess up work. Therefore, the answer already lies within the ways we have been managing organizations for a long time. He breaks down the parallels into seven categories.
1. Flooding with tokens = pushing forward by simply adding more people
The problem isn’t using many tokens. The issue is that only one in a hundred people knows how to provide proper context to AI. When the remaining 99 grab an agent, they end up generating endless repetitive loops.
2. Loop = meeting for the sake of meeting
In the end, a loop simply fills the gap left by people who can’t craft prompts properly by running endlessly until it works. Compared with a well‑defined task, a vaguely‑thrown‑together task can make the same work up to 100 times more expensive (what could be done for $4 ends up costing $310). Trying to save tokens ends up consuming even more tokens.
3. Wasted tokens = inflated headcount
Just as 80 % of employees make little tangible contribution, 80 % of today’s tokens do nothing. Even AI has inherited the same middle‑management hierarchy. In reality, only about 10 % of tokens are productive.
4. 100× token = 10× engineer
"Tokens are more accurate, faster, and never quit" is always qualified. They are accurate only when the prompt is accurate, and they can even be confidently wrong despite a perfect format. The real strength of AI is “scalability.” Mismanage it and it becomes more costly. The key is to identify the tokens that can amplify results a hundredfold and scale them.
5. Hiding know‑how = protecting a new revenue stream
This is the most subtle and political point. Employees are reluctant to hand over their know‑how to AI without hesitation. At Meta, even employees who own company stock were reportedly angry that the company was using their work context for AI training. Knowledge that only you possess has been a means of protecting one’s livelihood for centuries, dating back to medieval guilds, and AI is the first technology that demands you reveal it all at once. No one mentors their successors for free. Consequently, companies are structurally designed to push out their most critical technologies.
6. Evaluation Criteria (Evals) = OKR
The way to manage AI well is the same as the way to manage people well: first define 'what constitutes success'. Coding is the only field that has escaped political logic and exploded in growth because it embeds a clear evaluation criterion—"code either works or it doesn't"—which is why 99% of AI revenue comes from coding. Going forward, each company’s evaluation framework will become its most valuable asset and the core of its competitive advantage. Using the same evaluation criteria or agents that everyone else uses does not create differentiation.
7. The next 1 trillion-dollar opportunity = 'Transformation companies'
The author bluntly states. No one has yet been able to run AI reliably. Silicon Valley, convinced of this failure, is betting on 'AI‑native startups (neo‑firm)' instead of established companies, but the author thinks otherwise. He argues that companies that achieve AI transformation will grow ten times larger than any startup. Because each implementation spawns dozens of new use cases (Jevons paradox), transformation is not a one‑off project but an ongoing challenge. He cites Palantir as an example. By SaaS metrics, the company should have already collapsed, yet Palantir sold not software but the 'transformation' itself, and its stock rose 79% after the launch of ChatGPT (while SaaS stocks fell 2.9% over the same period). However, the meaning of transformation has also changed. It now goes beyond creating custom software; it involves setting evaluation criteria, conserving tokens, and deeply understanding the business enough to program it. Embedding each company's unique context into an agent will be the biggest economic challenge of the next decade, according to his conclusion.
And the closing message is striking. "Now is the time to manage."
The main point of this piece is that, just as with railroads, adopting AI agents isn’t merely about acquiring new tools—it’s about rebuilding systematic management. Just as important as the opportunities unlocked by new technology is the management capability to govern and forecast its use, which ultimately decides success or failure. Whether you’re looking at cost control or performance definition, it ultimately boils down to a management problem—something worth rethinking.