Something strange happened to software while everyone was arguing about whether AI could write it. The question got settled, and almost nobody adjusted to the answer.
AI can write it. Software production has become nearly free. Early enterprise deployments of AI-native, spec-driven delivery are reporting calendar compression of 97 to 98 percent, three years of scoped work landing in six weeks, six weeks of work landing in two days.
Arthur Miller believes those numbers, and that is precisely why he wrote a book about the danger inside them.
“A 97 percent faster path to the wrong system is a 97 percent faster mistake,” Miller writes in AI-SDLC and the New Enterprise Architecture Approach, his new framework for enterprise software in the age of intelligent agents.
Miller has standing to worry about expensive mistakes. He is the enterprise architect behind Zelle and Real-Time Payments infrastructure, and a fintech chief technology officer who has spent his career building regulated financial systems, the kind where an error doesn’t crash an app; it moves money to the wrong place.
His argument starts with a reversal most organizations haven’t absorbed. For the entire history of software, writing code was the expensive part, so everything else, the requirements, the architecture documents, the verification, organized itself around protecting that scarce effort. Now generation is cheap, and the bottleneck has moved. What’s scarce is knowing what to build, confirming it was built right, and deciding who answers when it wasn’t.
The industry’s current answer to cheap generation has a name Miller treats as a warning label: vibe coding. Describe what you want, accept what the machine produces, ship it. It feels like the future. Miller’s case is that it quietly reintroduces every risk disciplined teams spent decades engineering out, just faster and at greater volume.
His alternative doesn’t ask organizations to slow down. It asks them to redirect what they already have. Companies are full of requirements documents, project plans, and release notes that exist for humans to skim and forget. Miller’s framework turns those artifacts into living, AI-legible specifications, climbing what he calls a maturity path from Guided to Contracted to Generated, so that speed flows from intent that machines can actually execute against, rather than from improvisation.
The deeper argument is about people. When agents do the producing, humans do the judging, and Miller’s book is blunt that judgment, not code, is now the whole of the job. Someone still has to own what gets built. Someone still has to answer for it.
Notably, the book holds no vendor’s brief. Named products are quarantined to a single dated appendix, on the logic that the tools will change before the reader finishes the chapter. The framework is built to outlast them.
Software’s history has a pattern: every time building gets cheaper, the companies that win are not the ones that build the most. They are the ones that still know what’s worth building.
The machines settled the first question. Miller’s book is about the one that’s left.