AI is genuinely useful in website production, and it is useful in a narrower band than the marketing suggests. This is where it sits in my process.

Problem

Producing a website involves six phases with very different failure modes. Applying AI uniformly across all of them produces a site that demos well and fails in production.

My System

PhaseAI involvementWhy
ResearchHighSummarising competitors and extracting patterns is fast and cheap
DirectionNoneThis is the client’s positioning, not a generated artefact
PrototypeMediumLayout options are useful; final composition is not
BuildMediumComponent scaffolding yes; architecture no
TestingLowIt misses real-device and accessibility problems
DeploymentNoneDeterministic steps should stay deterministic
Where value moved
The bottleneck moved from writing the code to reviewing the code. Budget time for review or the speed gain is fictional.

Workflow

  1. Research pass. Competitor teardown, content inventory and the constraint list, drafted with AI and then corrected by hand.
  2. Direction document. Written before any design tool is opened. One page: audience, promise, content structure, constraints.
  3. Prototype in Figma. AI-generated layout options are used as raw material, never as the deliverable.
  4. Build in a static generator. Components are hand-reviewed. Anything touching the content model is written by me.
  5. Test on real devices. Two phones, one tablet, keyboard-only navigation, and a Lighthouse run.
  6. Deploy. Nothing AI-assisted happens here.

Why It Works

Because the phases where judgement matters most are the phases where I use it least. Research is cheap to redo. Direction is not.

Limitations

AI-generated front-end code has three recurring faults: it uses a framework when none is needed, it produces inaccessible interactive components, and it ignores the content model entirely. All three are invisible in a screenshot and obvious in month three.

Improvements

I am testing a review checklist generated per project from the direction document, so that review is systematic rather than remembered. Early results are promising and not yet trustworthy.

Related reading