The speed claims around AI-assisted coding are, for once, mostly true. A working first version of something that used to take two weeks can now take two days. Boilerplate, CRUD scaffolding, first-draft UI, even reasonably complex first-pass business logic — AI coding agents handle a meaningful chunk of this well, and pretending otherwise at this point is just outdated.
What hasn't changed is what happens after that first version exists. Speed of generation was never the part of software development that made it expensive — the expensive part was always making software that keeps working as real users, real data and real edge cases hit it over time. That part didn't get automated, and treating AI-generated code as production-ready because it ran once is how "vibe coded" became a slightly derogatory term.
Concretely, here's what we've found still needs a human, every time: architecture decisions that are expensive to reverse — data models, tenancy design, how services talk to each other — because an AI agent optimizing for "make this feature work" has no visibility into decisions made three files away last week. Security review, because an agent will write code that satisfies the prompt without necessarily satisfying "this input could be malicious." Edge cases nobody thought to describe, because AI is very good at solving the problem you stated and not the one you forgot to mention. And the actual decision to ship — someone has to be accountable for that, and "the AI wrote it" has never been a satisfying answer when something breaks.
None of this is an argument against AI-assisted development — we use it constantly, and it's a genuine and significant speed advantage when applied well. It's an argument against skipping the review, testing and hardening step because the first version already worked. A prototype that runs is not the same claim as software that's safe to put in front of paying customers, and the gap between those two things is exactly where engineering judgement still lives.
The practical version of this we apply on every project: AI accelerates discovery, scaffolding, implementation and test-writing. A human reviews every non-trivial change, owns every architecture decision, and signs off before anything reaches production. Fast where AI earns it, careful where it still matters — and those are not in tension with each other.