The useful part was speed
AI helped move the site from positioning, structure and copy into a working static build quickly. It was useful for generating page patterns, service-page structures, metadata, schema, markdown mirrors, supporting text files and deployment-ready packaging. That speed mattered because it created something concrete to inspect.
The hard part was not generation
The hard part was deciding what should survive. Several outputs looked plausible at first glance but were wrong in production terms: local preview links behaved differently from clean deployment URLs, the logo carried unreadable microcopy, footer navigation became cluttered, and a Notes card promised future content before any real note existed.
The failures were not dramatic. That was the point
None of those problems looked like a catastrophic engineering failure. They were smaller judgement failures: a stale header label on one page, a production build that was awkward to preview locally, a repeated UK-positioning phrase that made the site sound childish, and a layout correction that created a new visual misalignment. These are exactly the kinds of issues that make AI-assisted work feel finished before it is actually ready.
What had to happen
The useful pattern was simple: generate, inspect, challenge, harden and validate. Each fix had to be isolated, named, packaged and checked. Some changes were accepted. Some were reverted. The standard was not whether the code existed; the standard was whether the result could be owned.
What this means for AI-built products
A website is a small example, but the lesson scales. AI can produce a convincing first version of a product, workflow or internal tool. That does not answer the production questions: who owns it, how it fails, how it is tested, what assumptions are hidden, whether the language is credible, and whether the operating model can support it.
The practical lesson
Creation is cheaper now. Ownership is not. The work that matters after generation is still human: deciding what is true, what is useful, what is safe to publish, what should be removed, and what has been validated enough to stand behind.