AI has quietly changed the economics of making things. Writing a policy, a test plan, or a chunk of code used to take real time, which meant someone usually asked whether it was worth doing at all. AI removed that question, and it will keep producing more of anything if you let it.
A software engineer named Steve Yegge recently gave a clear real-world example of where that leads. He runs a fleet of roughly 50 to 60 AI agents to help build Wyvern, a video game he has worked on for thirty years. The system managing those agents has grown to around 600,000 lines of code.
The game itself, the actual product being sold, is only about twice that size. The agents have also written 450 internal rules and enforcement mechanisms to govern their own behavior. The rulebook grew so large that Yegge had to create a role just to prune outdated rules.
None of this is fake work. Players reportedly asked him to slow down new features because the pace was too fast, which is a strange complaint to have and a sign the system genuinely works. But it also means a factory built to produce a game has partly turned into a factory that mostly maintains and governs itself.
That pattern shows up in harder numbers too. An analysis of more than 10,000 developers found that teams using AI heavily completed noticeably more tasks and merged far more code changes than teams that did not. But the same teams saw review time climb sharply, code changes grow much bigger, and bugs per person increase as well.
Across whole companies, the study found no clear link between heavier AI use and better business results. Output went up. Value did not follow at the same rate.
Google's own research on software teams found something similar: as AI use rose, release stability got worse. Separate research tracking hundreds of millions of lines of code found that code rewritten again within two weeks of being committed has roughly doubled since AI coding tools became common.
This is not only a software problem. A survey of over a thousand office workers found that 4 in 10 had recently received AI-made work that looked polished but fell apart under real use, forcing them to redo it themselves. Researchers estimated this costs a company with 10,000 employees around 9 million dollars a year in wasted time.
The pattern across all of this is the same. AI has made the cost of creating something close to zero, but it has not made the cost of understanding, checking, and maintaining that thing any cheaper. Every policy, test, agent, or report someone creates is now a small ongoing cost to the business, whether or not it ever gets used.
For any business bringing AI into daily work, the useful habit is not asking what more the AI could produce. It is asking, before anything gets kept, whether it would actually be missed if it did not exist. That question is now the expensive part of the job, and AI cannot do it for you.