For seventy years, the software development lifecycle was organized around one economic fact: writing code was expensive. Every ceremony of enterprise delivery, the requirements phase, the design review, the release gate, existed to protect that costly effort from waste.
That fact just died. AI-native delivery is reporting 97 to 98 percent calendar compression in early enterprise deployments, three years of scoped work landing in six weeks, six weeks landing in two days. Generation is now nearly free.
Arthur Miller’s new book, AI-SDLC and the New Enterprise Architecture Approach, starts from the question most organizations are skipping: if code is cheap, what’s expensive now?
His answer is intent and verification. Knowing precisely what to build, proving it was built right, and establishing who answers when it wasn’t. Miller, the enterprise architect behind Zelle and Real-Time Payments infrastructure, argues these were always the real work. Cheap generation didn’t remove the lifecycle. It removed the excuse for treating the lifecycle as paperwork.
The framework’s mechanics start with a reframe of documents most companies already have. Requirements, architecture documents, project plans, release notes: today these are artifacts written for humans to skim. Miller’s approach converts them into living, AI-legible specifications, machine-readable statements of intent that agents can execute against and be verified against.
Those specifications climb a three-level maturity path. At Guided, specs steer human-AI collaboration, with people directing and machines accelerating. At Contracted, specs become binding interfaces, precise enough that outputs can be automatically checked against them. At Generated, the specification itself becomes the source of truth from which working systems are produced, with humans governing the spec rather than the code.
Around that path, the book rebuilds the full lifecycle in eight stages: Discovery, Design, Planning, Implementation, Verification, Release, Monitoring and Analysis. Each chapter walks a worked example, defines AI’s role at that stage, assigns tollgate ownership, and catalogs the common failure modes. The tollgate detail matters more than it sounds. In Miller’s model, agents produce and humans judge, which means every stage needs a named human who owns the gate, because accountability is the one artifact AI cannot generate.
The framework’s villain is what the industry has cheerfully named vibe coding: describe what you want, accept what the machine emits, ship it. Miller’s case is that this quietly reintroduces every risk disciplined teams spent decades engineering out, undocumented behavior, unverifiable correctness, unowned failure, at higher speed and volume than ever. As he writes, “a 97% faster path to the wrong system is a 97% faster mistake.”
For engineering leaders, the book’s most practical section is its migration playbook, which moves an organization from document-driven to spec-driven one artifact at a time rather than through a big-bang transformation. Reference appendices cover enterprise integration patterns, agentic AI and MCP orchestration, a metrics starter kit, and a symptom-to-diagnosis troubleshooting guide.
One structural choice signals how the book thinks. It holds no vendor’s brief, and named products are quarantined to a single dated appendix, on the logic that tools will churn before the reader finishes while the framework is designed to hold. Miller spent his career in regulated financial systems, where architectures outlive the products that implement them. The book is built the same way.
The industry has spent two years asking how fast AI can produce software. Miller’s framework is for the question after that one: how an enterprise stays in control of what it now produces at that speed.
