I’ve been building websites for well over a decade, and for most of that time I’ve been told automation would take my job within two years. It didn’t. What changed instead was something quieter and bigger: the structure of the work itself. What used to take days now takes hours. What used to take hours now takes minutes.
It’s worth understanding exactly where that shift happened, because it determines what you’re actually paying for when you order a website today.
What the machine has taken over for good
The code skeleton. Writing an empty but correct page with a layout, a mobile version and basic styles used to be half a day’s work. Now it’s minutes. A language model produces that starting point without hesitation, and the developer just reviews it.
First layout proposals. “Give me three visual directions: classic, bold, light.” Working that out from scratch used to take a designer a week. Today the first attempts appear within minutes and serve as material for discussion, not as the finished product.
Draft copy. Nobody sensible publishes generated text without checking it. But a draft that a human then edits makes the work twice as fast. We’ve been writing this way for two years and we’re not going back.
Supporting images. Backgrounds, illustrations, fillers. Image generators make them more cheaply than stock libraries, with no licensing surprises.
What the machine hasn’t taken over
Taste. A model can generate a website. It doesn’t know which version is better for a hair salon in a small town and which for a law firm in a big city. That’s contextual knowledge you have to have, acquire or work out. The model has the average of everything; the client’s business is specific.
Responsibility for facts. Models make things up with complete composure: certifications the business doesn’t hold, awards it never won, reviews nobody wrote. On a business website that’s a material risk, not a detail. That’s why at Webomat the agent works only with content from the client’s old website and doesn’t add anything of its own.
Business decisions. What should be the first button on the homepage? Show prices up front or on request? Show the team or not? Those are questions for the business owner, not for a model.
Quality control. A generated site looks fine in a screenshot and falls apart on a phone. Or has dead links. Or all of these at once. Checking for that is work nobody will ever see in a portfolio, but everyone notices when it’s missing.
Why the shift is happening now
Three things matured at the same time.
First, models learned context. Five years ago, “make a website for a nursery school” produced a generic template. Today a model understands what role a photo plays, where to guide a parent and why the tone should be warm. Not always well, but well enough to serve as a starting point.
Second, they learned images. A website without photos is a flyer. Image generators closed that gap.
Third, the cost of getting it wrong dropped. Generating and discarding a proposal costs pennies. It used to be that every rejected design concept was a day of work down the drain. Now you can offer a client three directions to choose from, because two rejected ones won’t wreck the budget.
Where we stand
We built Webomat precisely around this division of labour. The machine reads the client’s old website, gathers the brand, content and photos, then designs and codes. Humans set the rules and check the result: at five screen widths, link by link, feature by feature. The client picks a direction and refines the rest in a chat.
The result: a new business website in under an hour instead of three weeks, for a fraction of agency rates. Not because it’s worse than an agency site. Because the days you used to pay for are no longer needed.
A week of work then and now
A concrete example from our own office. A business website, eight pages, two languages.
The version from five years ago. Monday: brief with the client, notes, structure. Tuesday and Wednesday: homepage design. Thursday: a call about revisions. Friday: revisions. Week two: the remaining pages, sign-offs. Week three: coding, content, testing. Launch at the start of week four. On the quote: about 60 hours of people’s time.
Today’s version with Webomat. The client pastes in their address. A minute later they have a snapshot and an audit. The agent reads the site, builds a visual direction from the logo and the industry, and generates proposals. Quality control runs in the background. After 15–25 minutes the client picks a direction; another 15–20 minutes later they’re looking at the whole website. Revisions go in through the chat and come back in minutes. The human work involved: a few hours, mostly decisions and checks.
The difference isn’t that someone learned to write code faster. It’s that the in-between stages no longer need people. The client pays for the result and the decisions, not for the process.
Will AI replace designers and developers?
This question comes up on every panel, and rightly so. Here’s the answer from where we sit:
Designers aren’t going anywhere, because a model has no taste. It has an average. Someone has to decide whether this square fits this brand, or no brand at all. A designer with a good eye moves from drawing squares to setting rules and judging results. It’s closer to an art director’s job than a craftsperson’s.
Developers aren’t going anywhere either, because someone has to understand what the machine spat out and take responsibility for it. Model-written code is sometimes correct and sometimes only looks correct. The difference shows up in the edge cases: an old browser, unusual hosting, an integration. Developers move from writing from scratch to reading, fixing and hardening.
What does disappear is the copy-and-paste work. Skeletons, routine layouts, repetitive forms. Those are exactly the hours vanishing from quotes, and that’s why websites are getting cheaper.
Things AI still gets wrong
To keep the picture honest, here are the slip-ups we see regularly:
- Invented facts. A model will describe awards, certifications and reviews the business doesn’t have, with total calm. No checking, no publishing.
- Drifting off-brand. Without plugging in real colours, logo and content, every generation drifts towards a “safe navy blue”.
- Blindness to mobile. A generated layout looks good in a screenshot and falls apart on a narrow screen. That’s why we enforce checks at five widths.
- Lazy language. Model-written text has a rhythm: sentences of equal length, stock phrases everywhere. An editor or very precise guidelines are a must, not an option.
If we had to boil it down to one sentence: a model is a brilliant intern who never sleeps and has no conscience. It gets a day’s work done, but someone has to sign it off.
What will change next
Briefly, because predictions are a great way to embarrass yourself:
- Quality control will be automated too. Today humans check the model’s output. Tomorrow models will check each other, and humans will approve the exceptions.
- Websites will maintain themselves. The current year in the footer, an updated price, a new case study: things someone has to remember to do by hand today.
- Price will stop being a barrier. When a business website costs about as much as a single enquiry is worth, the decision shifts from “can I afford it” to “do I want my business to be visible”.
How not to get fooled by “AI”
Now that everyone is adding “AI” to their offer, here are some test questions:
- Where does the tool get its content? (Your real content, or made-up content?)
- Who checks the result before you see it?
- What exactly does a human do in the process?
- Do you get the files, or just an account on a platform?
Treat “the AI does everything” the way you’d treat “full guarantee, trust me”. It sounds great and means nothing.
Not entirely, and not everywhere yet. But there’s only one direction this is going, and it isn’t backwards.