AI will draft a layout in an evening. So why is the site not three times cheaper

5 min readAIprocesspricing

Because it took away about a third of the time, not two thirds. And not because there are things it cannot do - because it rarely gets them right first time.

This comes up on almost every first call now. And it is a fair question: you watch a finished-looking layout appear in an evening and wonder why the invoice looks the same as it always did.

The short answer: AI really did take about a third of the time away. No more than that. Below is where that third comes from, and why the other two still belong to a person.

It is not that AI cannot do things

People like to write that it cannot, and it is not true. Give a model enough context and enough time and it will work almost anything out. It will happily account for your two warehouses and the one that shuts on Saturdays. Somebody who knows that has to tell it, is all.

Capability is not the constraint. The constraint is how many times you have to ask again, look, and say «not like that» - and who is able to see that it is not like that.

What generation actually hands you

A good-looking picture and bare markup. That is already something - but it is not a site, it is what a site looks like.

  • The markup is not semantic: instead of headings and buttons, identical blocks that merely resemble them
  • There is no accessibility at all: you cannot get through it by keyboard, a screen reader announces nothing
  • It is wired to nothing - no database, no payments, no mail
  • The copy is plausible but not yours - often with invented years and awards
  • The mobile version survives on luck, because nobody has checked it with a thumb

And the main thing: it looks finished. That is the expensive illusion in all of this, because it is the reason nobody checks. The page opens, the buttons are there, it all looks right - and nobody presses the one button with nothing behind it.

Then the work starts

The markup has to be fitted onto a system where you can change a price yourself without calling a developer. Then connected to whatever already runs the business. Then checked - by hand, on a real phone, not in a narrowed browser window.

  • Move it onto WordPress or Next.js so every block is editable, not only the ones somebody remembered
  • Connect payments, stock, CRM, mail - and see what happens when one of them drops out
  • Rewrite the markup properly: headings as headings, buttons as buttons
  • Do accessibility rather than tick it: keyboard, screen reader, contrast
  • Walk every path on a phone, including the ones where the item is gone and the payment failed

It does not come out right first time

This is probably the most important part and the least visible from outside. Claude rarely nails it in one go. On a serious page you sit with it for hours: ask, look, say «not like that», redo it, check again. Often it is faster and safer to write the bit yourself than to explain it a third time.

And to say «not like that», you have to know what right looks like. Someone who does not will only see that it is pretty. That is why generated sites usually fail not in the code but at handover: there is nobody to check them.

For AI to get it perfect, somebody beside it has to know what perfect looks like.

Where the third comes from

This is our own estimate across our own projects - not stopwatched, but not plucked from the air either. What disappeared is roughly what was already cheapest on the invoice: first layout options, routine plumbing, draft copy, test data, a rough translation pass.

Working things out, wiring them up, checking them and answering for the result stayed exactly where they were. So the estimate changed shape rather than total: fewer hours drawing, more on the part a person is actually worth paying for.

1/3of the time is what AI took. A person still does the rest - and for now that is the honest ratio

That number will grow

We are fairly sure of it, and that is fine. A year ago the share was noticeably smaller than it is today. A year from now it will be more than a third, and anyone pretending nothing is changing will simply work slower than the people who are not.

What will not change quickly is the need for someone who looks at the result and knows what is wrong with it. Nor the fact that a site keeps living after launch: we build a month of support in, because that is when the things development cannot show you turn up.

What to ask whoever builds it

Do not ask whether they use AI. In 2026 a «no» means you were either lied to or told, on principle, that your time is not worth saving. Ask about other things - you can hear the answer straight away.

  • Who reads the copy before it goes live, and how they know your business
  • Whether you can change a price and a photo yourself, or have to write to them for every edit
  • What the mobile version gets checked on - a real phone, or a narrowed browser window
  • What happens if the form stops sending mail: do you find out, or not
  • Whether you can open their previous sites right now and poke at them yourself

The last question is the easiest one, and almost nobody asks it. A layout can be generated. A screenshot can be generated. A testimonial, frankly, can be generated too. A working site that opens at its own address cannot.

Which is why every case on our home page opens in a new tab. The best way to check us is to go and look.