
AI Will Save Money
Replace unproductive humans *
(* That didn’t need to reread the manual daily)
They replaced the engineers to save money. The agents worked 24/7, never complained, never asked for equity, and had no interest in health insurance. It looked clean on the deck. Then the invoice showed up and it turned out the agents charge by the word. Every word going in. Every word coming out. Both directions.
For two years the pitch was elegant: humans are expensive, AI scales. What the deck left out was the amnesia. The IQ is there. The reasoning is sharp. The memory is gone by the next chat. Functional Alzheimer’s, billed hourly.
A human engineer walks in carrying twenty years of context. The 2019 database decision. The client who rewrote the requirements twice. Why that module was built that way. What failed in 2022 and why nobody touched it again. That memory came with the salary. Accessing it cost nothing extra. They just knew.
An agent knows nothing. Every task starts from zero. You provide the codebase, the instructions, the prior conversation, the relevant docs, and enough context for it to pretend it was here yesterday. You pay for every word going in. It answers. You pay for every word coming out. Then the next task arrives and you do it again, because the agent retained none of it. Same briefing. Same invoice. Every time.
As for why the AI rambles — it’s probably unrelated to output tokens costing more than input ones.
Jason Calacanis mentioned hitting $300 per day per agent using Claude’s API at partial utilization. That’s roughly $100,000 per year. Per agent. Chamath Palihapitiya reportedly started asking what the token budget was for their best developers. These are people who read P&Ls for breakfast. They forgot the AI forgets — until the bill made them remember.
There is also the part the pitch deck skipped entirely. The companies that let engineers go to save money now need engineers to manage the agents. Someone has to decide what to tell it this time, what to leave out, and where to set the limits before the meter runs. The agents don’t make those calls. They were hired to read.
The pitch was replacement. The invoice is a subscription to amnesia, billed by the syllable. Before letting the engineer go, price out the daily briefing — every task, both directions — and put a number on twenty years of context that never once appeared on an invoice. The agents work 24/7. So does the meter. The engineer went home at five and remembered everything in the morning. That used to be called overhead.
Editor’s Note: Think about the heroin-like pricing model — to get great output you have to pay for vast and growing input. And hope the AI understands the first time. Or you pay again. ROFL


Documenting AI absurdity isn’t just about reading articles—it’s about commiserating, laughing, and eye-rolling together. Connect with us and fellow logic-free observers to share your own AI mishaps and help build the definitive record of human-AI comedy.
Thanks for being part of the fun. Sharing helps keep the laughs coming!