Will AI replace web designers? Not the ones who write the brief

The short answer is no, and that answer is worthless on its own. It only holds if you first say what AI did take, and almost nobody wants to say it.

So here it is. If your offer, written out honestly, is "I will make you a tidy website from your description," that offer has been eaten. Not weakened. Eaten. A coding agent produces a competent, responsive, decent-looking page from a clear description in minutes, and the client can watch it happen.

What was not taken: deciding what the site is for and whose problem it solves; looking at five outputs and knowing which one is right and why; sitting with a client who cannot yet say what they want. And the plain fact that when the site breaks on a Friday evening, someone calls a person.

The job moved from making the thing to specifying it and judging it. You are the creative director now, and your prompt is the brief. Excellent news if you can write one, a genuine problem if the brief was always the part you skipped.

What AI actually replaced

An inventory, without the softening. Things people were paid for in 2023 and are largely not paid for now:

• Fitting a template to a client's content and making it look intentional.
• "Make it like this competitor's site, but ours."
• The first draft of the copy. Not the good copy. The draft that took an afternoon.
• Basic responsive layout: breakpoints, stacking order, the mobile menu. Real skill, thoroughly automated.
• Choosing a serviceable grid, type scale and palette when nobody has strong feelings.
• The third round of "move it up a bit."

That list is not marginal. For a lot of freelance web design it was most of the invoice. Pretending it survived is how you end up quoting for work the client already knows they can get for nothing.

One boundary, because the two questions get mixed constantly: this is web design, not web development. Whether AI replaces the person building an application with a database behind it is a different argument. Here it is the commercial site, the landing page, the one-page brochure for a business that knows what it sells. Which tool produces that page is a genuinely separate decision, and not the one that decides what you are worth.

What it did not touch

The surviving work is not a list of virtues. Creativity and empathy are not units of work and nobody invoices for them. These are units of work.

Deciding what the page is for

Every site starts with a decision that has nothing to do with layout: whose problem does this solve, and what single thing should a visitor do. That decision is most of what UX actually is. A model does not make it. It either receives it from you or substitutes the average of everything it has read, which is why so many generated pages feel like a competent answer to a question nobody asked.

Knowing which of five outputs is right

Generating five variations is free now. Choosing correctly is not, and the two were never the same skill. Anyone can ask for options. Far fewer can say which serves the visitor, and defend it to a client who prefers the other one.

The conversation with a client who does not know what they want

Half of any brief is extracted, not submitted. It comes from asking a builder why his enquiries dry up in January, or noticing that a physiotherapist's real competition is people deciding to wait and see. That happens in a conversation, not in the intake form.

Everything after "it looks done"

The page existing is not the project finishing. It needs an address, a domain the client owns, a way to be changed in March, and someone accountable for all three. Most of the remaining hours live here, and the speed conversation skips it. The arithmetic of what the 10× actually does to a project is its own piece.

Your prompt is the brief

The difference between a prompt and a brief is composition, not length. A brief says who the reader is, what one action matters, what is being sold, what must not appear, in what tone, what counts as done, what is out of scope. The average prompt contains none of that, which is why the average generated site looks like the average generated site. Ask for "a good title" and you get one of four different things that go by that name.

For that composition in copyable form, the six website prompts are the same idea, and the last two end at a live address. The five gaps a brief most often leaves are in AI website examples.

Here is one written out in full, for a solo physiotherapist's single page:

Reader. Someone in town with back or knee pain who has already decided to see someone. On a phone, usually in the evening, usually mildly anxious.

One action. Book a first appointment: tap the phone number, or open the booking calendar. Nothing else on the page competes with this.

What we are selling. An appointment this week, with evening slots. Not twenty years of experience. Everyone claims that and nobody is choosing on it.

Above the fold. What she treats, which town, the price of a first session, the booking action. In that order.

Must not appear. Stock photography of smiling strangers. A mission statement. A slider. The word "holistic."

Tone. Plain and calm, no exclamation marks. She is a clinician, not a brand.

Constraints. One page. No enquiry form, because booking goes to her existing calendar link. Her existing green, which she likes and is not changing.

Done means. A stranger on a phone can tell within five seconds whether she treats their problem, and can book without scrolling twice.

Out of scope. Blog. Testimonials, until there are three real ones.

Compare that with "build a modern website for a physiotherapist," which is what most people type. Same model, same afternoon. Everything that distinguishes the two results was decided before either ran.

Notice how much of the brief is unguessable: the green, the evening slots, the banned word, the reason testimonials are excluded. None of it is inferable from the industry. It came out of a conversation with a person, and that is now the part that determines the output.

Taste is the deliverable now

Taste sounds aesthetic. In this job it is operational: being able to say what is wrong in words the agent can act on.

"I don't like it" is not a brief. "The headline promises efficiency and the page sells reassurance, so the visitor is being sold two different things" is a brief, and it produces a fix on the next pass. See the problem, name the cause, phrase the correction. That is what an art director always did, applied to a model instead of a junior designer.

If the diagnosis is merely "make it cleaner," use the five testable clean-design rules to name the hierarchy, spacing, type, colour or action conflict first.

This is also the whole defence against generic output. These tools produce the average of what they were asked for. We read thirteen rival blogs to find where the line falls, and it fell exactly there: published without a standard, not made with AI. A page looks templated when nobody upstream was willing to reject the first thing that appeared. Every workflow that works has the same shape, specify, generate, reject, re-specify, and the two steps that decide the result are the ones a person does.

Nobody pays a model when the site goes down

Here is the argument the other answers leave out, and the sturdiest one, because it is not about skill. Every article promising designers are safe argues from talent. The durable reason is structural.

A client is not buying a layout. They are buying somebody answerable for a live site. Whose domain is it. Where does it live. Who rolls back the change that broke the pricing section. Who replies when a phone number is wrong. Those questions resolve to a person, and no amount of capability on the generation side changes who.

That is also why the work does not end at a good-looking page. A published artifact in a chat is a link on somebody else's domain, not something a client owns. A domain has to be pointed at the site. And a year later someone has to change one line without republishing everything.

For most clients that is the part they will actually remember.

What the numbers say, and what they do not

You will be shown percentages. It is worth knowing where they come from. We did that exercise for the most-quoted vibe coding statistic, and it is worth doing here.

One of the pages ranking for this question opens with a key takeaways box listing, among other things, that eighty percent of repetitive design work is automatable, attributed to McKinsey, and that forty-five percent of designers expect demand for basic work to fall by 2030, attributed to a Creative Bloq survey. Another page ranking alongside it says ninety-three percent of designers use AI tools daily and that consumers will pay sixty percent more for human-intentional design. None arrive with a study title, a sample size or a date. They arrive as a number and a brand name, which is enough to be repeated and not enough to be checked.

We are not adding to the pile. The US Bureau of Labor Statistics does publish an employment projection for web developers and digital designers, and we could not open its page to quote the number exactly, so it is not in this article. Take it from the Occupational Outlook Handbook directly.

Nobody has a reliable measurement of this yet, and the pages implying otherwise are decorating an argument they had already decided. What you can observe without a statistic: what clients now expect to pay for a simple page, and how quickly they stop asking who made it. If the pay question is yours this quarter, the ranges freelancers report and the margin math behind them beat any percentage above.

What this changes about how you work

Four things follow, and none is "learn prompting."

Write the brief before the first prompt. Keep a template with the fields above and fill it in every time. It is the highest-leverage document in the project, and it doubles as what you show a client who asks what they are paying for.

Sell specification and judgement, not production hours. The hours you can defend are the ones spent deciding and rejecting. Quoting production invites exactly the comparison you do not want.

Own the last mile. Publishing, the domain, the ability to make a change later. That is what turns a folder of files into a delivered project. Handing over a zip file is how you become invisible.

Be able to explain the difference. When a client says "can't AI just do this," the answer is: yes, and here is the brief that made this version better than the one you would have got. Show them the brief.

Where Birta fits

Birta is only involved in the last item on that list.

A connected agent publishes the finished site straight to a live HTTPS address, and publishes again after each round of changes, with no repository or build pipeline between. Edits go file by file through the same connection, so changing one line does not mean republishing. Images and fonts travel with it; video and PDFs need an address elsewhere. Each publish is kept as a version and the last ten stay available to roll back to. You can point a client's domain at it with HTTPS. The connection itself is MCP, explained without the infrastructure vocabulary.

What it does not do is the rest of the job. It will not write the brief, sit through the approval round, or decide which of the five versions is right.

Frequently asked questions