Few terms are currently used as generously as “AI-assisted web design”. It covers everything from an image generator to a builder that spits out a website at the push of a button. This article sorts out what holds up in practice, and what costs companies money without returning anything.
The starting point: AI in the result or in the process
There are two fundamentally different ways to use AI in a web project. The first produces the result: one prompt, and out falls a home page. The second speeds up the road to it: data preparation, cutting out images, text variants, translations, test coverage.
The first is loudly advertised, the second actually changes project costs. Confuse the two and you buy a promise instead of a service.
Where AI really saves time in a project
In our projects it is almost always the same places, and rarely the creative ones:
- Product data: bringing thousands of articles out of an ERP into a usable structure, harmonising descriptions, proposing categories. Manual work that takes weeks becomes days.
- Image preparation: cutting out, deriving formats, proposing alternative texts. The proposals get reviewed, but no longer written from scratch.
- Translation: first drafts for other languages that a native speaker then edits. Considerably faster than translating from zero.
- Migration: moving content from an old system into a new one, including structural and formatting changes.
- Recurring workflows: pre-sorting enquiries, preparing quotes, extracting data from documents. That keeps running after launch and saves hours every month.
Where AI fails in web design
A website has one job: turning visitors into enquiries. That requires a decision about what to leave out, and that is precisely what generative systems cannot do. They produce the plausible, not the pointed.
You recognise the result by its interchangeability: pages that look like every other page, with text that would fit any company in the sector. They are not wrong. They are simply replaceable, and in a competitive market, being replaceable is the most expensive outcome there is.
On top of that come practical problems: unclear rights on generated images, disclosure duties under Art. 50 of the EU AI Act, and content nobody has checked for accuracy that nonetheless goes out in the company's name.
What that means for choosing an agency
“We work with AI” is no longer a distinguishing feature. Everyone does by now. Three other questions are more interesting:
- At exactly which point? A concrete answer names tasks, not buzzwords.
- Who checks the result? If nobody is accountable by name, nobody is.
- Where does the data sit? For much of it, cloud processing is unproblematic. For client data, quotes and contracts often not, and then you need local models on your own network.
Data protection and disclosure
Two duties have not been negotiable since 2026. First: anyone having personal data processed by an AI system needs a legal basis for it and a data processing agreement under Art. 28 GDPR. Second: content generated by AI, or substantially edited by it, must be recognisable as such under Art. 50 of the EU AI Act.
Both are manageable if they are planned in from the start. Retrofitted, they become a project of their own.
Conclusion: a tool, not a replacement
AI genuinely changes web projects, but somewhere other than where it is advertised. It makes the dull work faster and cheaper. The decisions about what a brand says and what it leaves out stay handmade.
That is why we use AI in the process, not in the result. Concept, brand direction and accountability stay with people. This is not restraint on principle, but the experience from projects where the opposite was tried.