George Sivulka, in his recent article “You Just Hired a Million Bad Employees,” makes an argument that stopped me mid-scroll: for the first time in history, humans are cheaper than software. Not because AI failed. Because almost nobody has learned to manage it yet.
The article backs this up with two charts worth sitting with. The first tracks annualized AI spend per employee at top firms, climbing from around $25,000 in early 2025 toward $200,000 by the end of 2026, closing in on what a fully loaded software engineer costs. The second tracks headcount at companies that have adopted AI, and it does not show layoffs. It shows growth, about ten percent more headcount over two years at the heaviest AI spenders, while low-spend firms stay flat.
Both numbers say the same thing from different angles. Spending on AI and getting value from AI are not the same activity, and most companies are doing a great deal of the first without much of the second. That gap is the entire premise of my book. Somewhere between 88 and 95 percent of companies are running AI today. Only about one in twenty captures real value from it at scale. Having AI and making money from AI turn out to be almost entirely different things, and Sivulka’s spend chart is what that difference looks like once it reaches the income statement.
The headcount chart points somewhere I did not expect an a16z newsletter to go: management. Sivulka tells the story of the American railroad in the 1830s, track mileage multiplying more than a hundredfold in a decade, until two trains collided in Massachusetts because nobody had built the coordination to prevent it. The fix was not a better train. It was hiring managers, writing down roles, and building a hierarchy that could hold the complexity the technology had created. Modern management was invented to keep the railroad from killing people.
The first chapter of my book opens with a similar idea, borrowed from an older source. Adam Smith’s pin factory took a single job and broke it into narrow steps, and output multiplied because each person could specialize. AI is doing the same thing to cognitive work that Smith’s factory did to physical work, and the productivity is showing up in the headcount numbers Sivulka charted. But the early layoffs are not landing evenly. They are landing on middle management, the coordinating layer between the front line and the executives, the same rung the railroads had to invent from scratch two centuries ago. Cloudflare’s chief executive said their cuts targeted “measurers.” Google trimmed more than a third of the managers running its smaller teams while cloud revenue grew 63 percent. AI is not removing the need for coordination. It is removing the old coordinating structure and leaving companies to build a new one, on a faster clock than the railroads ever had.
Sivulka’s list of seven parallels lands on a related point under a name I like: evals are the new OKRs. My book’s builder chapter says the same thing in blunter terms. A tool you cannot test is not a tool you are maintaining. It is a tool you are hoping about. Every company that gets serious about AI eventually hits the same wall, that you cannot manage what you have never defined as good, and that definition is not a technical detail. It is the actual work.
None of this argues against AI. It argues for treating it like the workforce it has quietly become, one that needs the same things every workforce has always needed: a clear definition of success, a real management layer, and someone willing to do the unglamorous work of building both.
Read the original piece: You Just Hired a Million Bad Employees, George Sivulka, a16z.
More on the ideas above: AI’s Missing Playbook: Aligning AI With the Goals That Move Your Business
It’s time to build.