Where AI actually helps in web design (and where it still doesn't)

We have spent two years integrating AI into our build pipeline. We have used it on every project we have shipped since the start of 2024, in roles that range from 'co-pilot for a senior engineer' to 'first draft for a junior writer.' Here is the honest report card on where it actually helps, and where it still does not.

Where AI clearly helps: first drafts. Whether it is product copy, blog outlines, alt text, meta descriptions, or boilerplate code, AI gets the rough shape on the page in seconds. A senior writer or engineer can edit a rough first draft into a great final draft in a fraction of the time it would have taken to start from a blank page. This alone has cut our copy-to-launch timeline by roughly 30%.

Code scaffolding is a near-pure win. Routing files, form components with validation, API client wrappers, type definitions, test scaffolds — AI generates these in seconds and the failure mode (when it gets it wrong) is obvious to a senior engineer reviewing the diff. We treat AI like a fast junior dev: useful when the task is well-scoped, dangerous when the task requires architectural judgment.

Image work has been transformed. Background removal, upscaling, style transfer for mood-board work, generating placeholder imagery for early-stage prototypes — what used to take a designer half a day takes minutes. We still hand-art-direct hero imagery, but the hundreds of supporting images that fill out a marketing site are largely AI-assisted now.

Where AI still struggles: brand voice. Out of the box, AI writes in a flat, slightly corporate voice that screams 'AI wrote this.' We get useful results only when we feed it strong brand voice guidelines and edit aggressively after. The brands that publish raw AI output stand out for the wrong reasons — readers can tell.

Design judgment is still the human's job. AI can generate variations endlessly, but picking the variation that solves the brief — that is taste, and taste does not come out of a model. We use AI to widen the design search space, never to make the final choice.

Architectural decisions are also human. Choosing a stack, structuring a database, deciding when to extract a component, judging whether a feature should ship now or wait two weeks for a better implementation — these are judgment calls that depend on context the model does not have. The teams that try to outsource these to AI ship worse software, slower.

SEO content is an interesting middle ground. AI is useful for outlines, supporting paragraphs, and FAQ sections. AI is harmful when it is used to generate articles end-to-end and publish them without human editing — Google has gotten very good at detecting this kind of content, and ranking is suffering for sites that lean on it. The smart play is AI-assisted, human-edited; the dumb play is AI-generated, human-skipped.

Customer-facing AI (chatbots, assistants, copilots) is mostly still not great. The best implementations we have seen are narrow, transactional ones: order status, return tracking, sizing help. Open-ended 'ask anything' bots are a churn machine — they fail in front of customers, the failures are public, and the support team ends up cleaning up the mess. We recommend tightly-scoped bots only.

Our internal rule: AI is allowed in the workflow when a senior human reviews the output before it ships. AI is not allowed to be the last set of eyes on anything that goes to a customer. That rule has kept the quality bar where we want it while still capturing the speed benefit.

If you are a small brand wondering whether to add AI to your stack: start with the parts where AI obviously wins — first drafts, code scaffolding, image work — and stay skeptical of the parts where it has not earned trust yet. The teams that do this win the speed benefit without paying the quality cost.