AI-assisted development without outsourcing the vision
AI becomes a serious production multiplier only when the human owner controls the brief, repository, evidence and final quality bar.
Human ownership
Positioning, visual language, prioritisation and final acceptance stay with the creator. The system is judged on the shipped result, not the prompt transcript. The decision that matters most is also the one least visible in any output: what not to build. A model will happily produce eleven more pages; deciding the client needed one verifiable proposition instead is the part that changes the outcome.
Where it is the wrong tool
Anywhere the answer has to be checkably true. A generated paragraph about a FIDIC clause reads exactly as confident when it is wrong as when it is right, so on subject-matter work the model drafts and the domain knowledge decides. Being able to reject an answer is the part that took years to acquire, and it is the part that does not transfer to whoever holds the subscription.
Evidence over mythology
Build logs, public repositories, analytics context and live URLs make the workflow auditable. The engineering log goes further and records the changes that were measured and then thrown away, which is the harder half to show. Tooling does not produce that record; being accountable for the result does.
The review loop is the work
Generated code is reviewed in repository context, against a running site, on real devices — not read once and accepted. The findings that matter most tend to come from measurement rather than reading, because the expensive mistakes are invisible in a diff: they are about what the page does on a phone, not what the code says.
What a client is actually buying
Not production capacity — that is cheap now and getting cheaper. What a rate buys is judgement under uncertainty: which of two plausible answers survives contact with a real Search Console export, which requested feature to argue against, and what to do at the point where the generated version is ninety per cent right and the last ten per cent is the part that fails in production. It also buys someone reachable in six months, which no model currently is. A generated site is a snapshot; a maintained one is an operation, and the gap between them opens months after launch.