You Just Hired a Million Bad Employees

There's a new essay out of a16z making the rounds, and its central claim is almost funny: for the first time in history, a human employee is cheaper than an AI one. Not because AI got worse, but because companies have been throwing tokens at problems the way bad managers throw headcount at them — infinitely, and without direction. The piece leans hard on a 19th-century analogy. American rail mileage grew 120-fold in a decade after the 1830s boom, and it took a fatal 1841 collision in Massachusetts, caused by nothing more than a coordination failure between two trains, to force the invention of modern management: written roles, reporting lines, regional managers. Rail eventually became roughly 60% of the stock market. The argument is that AI agents are having their own pre-collision moment right now.

The mechanism the essay zeroes in on is what it calls "looping." Most people, it argues, maybe 1 in 100, actually know how to give an AI clean context. Everyone else hands a vague task to an agent harness, watches it fail, and lets it call itself over and over trying to self-correct. That's not intelligence at work, it's a workaround for a human who never explained the job properly. And it's expensive. Elon Musk cut 80% of X's staff and the company got better, not worse — the essay's evidence that a similar 80% of corporate token spend is doing essentially nothing, just looping to justify its own existence.

Where this gets more interesting is the political angle. Employees, understandably, don't want to hand their hard-won tribal knowledge over to a system built to replace them. The essay points to Meta, where staff who own stock and are financially rewarded for AI succeeding are still furious that their own work is being used to train it. If that tension exists inside a company whose workers are literally paid to want AI to win, imagine every other industry. Medieval guilds guarded their secrets for centuries; AI is the first technology asking for all of it at once, immediately.

The fix the essay proposes isn't more prompting workshops. It's evals — the same reason coding became the one AI use case that actually scaled without political blowback, since code either runs or it doesn't. No fuzzy judgment calls, no hidden agendas. Every other white-collar function is waiting on someone to build the equivalent yardstick, and whichever firm builds the best, most specific eval suite for its own weird internal processes will have a real moat, not a generic chatbot wrapper. The essay's closing bet is on "AI transformation companies" over the trendy new AI-native startups trying to poach services revenue from incumbents — because the actual leverage still sits inside existing companies, in the processes that already work.