Using AI to build client websites: the 3-day site in 3 hours

The build really did collapse. A small business landing page that used to take a day and a half of drafting and layout now comes out of a prompt in under an hour, and it is good enough to put in front of a client. That part of the promise holds.

A client project was never just the build. It was the brief, the decisions the client had not made yet, the copy that arrived late, two rounds of revisions, publishing, the domain, and the handover. None of those got ten times faster.

So "three days to three hours" is a shape, not a measurement. It describes what happened to one stage of the work and then quietly applies it to the whole project. Three hours is a real number for a real case — one page, content supplied, design decisions already made, domain already delegated, the agent connection authorized once, weeks ago. It is not the number for the project sitting in front of you today.

Here is which stage moved, which stages did not, and what the arithmetic does to a total when only one leg gets faster.

Where the three days actually went

Three days was rarely three calendar days. It was roughly three working days of billable attention spread across two or three weeks of waiting. Written out, a small landing page for a client used to look something like this:

• Brief and decisions. Two or three hours, plus the calendar time it takes a client to decide what the page is even for.
• Collecting content. An hour or two of your effort chasing copy, photos, a usable logo, and permission to use any of it. Days of waiting.
• Structure and first copy pass. Three or four hours turning what the client said into something a visitor would read.
• Design and build. Eight to ten hours. This is the leg everyone is talking about.
• Revision rounds. Three or four hours of changes, stretched across a week of replies.
• Publishing. An hour or two, more if the host wants a repository, a build command, and a pipeline it never needed.
• Domain and HTTPS. Half an hour of actual work, then waiting that is not yours to control.
• Handover. An hour or two explaining who owns what, where the enquiries go, and what happens when the client wants a word changed.
• The edits that arrive afterwards. Recurring, unbilled, and never in the original estimate.

Add it up and you are at twenty to twenty-seven hours. That is where "three days" came from. Design and build was the single largest block, but it was still under half of the job.

The 10× is real, and it lands on one leg

Take that eight-to-ten hour block. Give a coding agent a clear brief, a color palette, and the client's copy, and you get a complete, responsive landing page in well under an hour, including the passes where you tell it the hero is too tall and the testimonials look like a spreadsheet.

That is not a marketing claim with the edges filed off. It is the most reliable thing anyone can currently say about using AI to build client websites. The drafting leg genuinely fell by roughly an order of magnitude, and it fell hardest exactly where freelance work is densest: small commercial sites, landing pages, portfolios, one-page brochures for a business that already knows what it sells.

What made it possible was not a better website builder. It was that a description in plain language became a viable input: your brief is now the build instruction. That is a real shift in the craft, and it deserves the enthusiasm it gets. The trouble starts one line later, when the shift gets applied to the stages above and below it.

What a 10× on 40% of the job does to the total

Run the numbers on the list above. Call the project twenty-four working hours. Design and build was nine of them — a bit under forty percent. Make that leg ten times faster and it becomes one hour.

Twenty-four hours becomes sixteen. That is not ten times faster. That is about one and a half times faster.

This is the part nobody writes down. A tenfold improvement on forty percent of the work cannot produce a tenfold improvement overall, and no amount of better prompting changes that. Once the build leg is down to an hour, making it thirty minutes buys you thirty minutes. The remaining fifteen hours are not waiting on the model. They are waiting on people, on files, on approval, and on the publishing steps at the end.

So the honest version of the headline is this: the three-day site now takes about two days, and the way to reach three hours is not to speed up the build again. It is to delete stages that were never about building at all.

The stage nobody puts a number on

This is easy to check, because the pages selling the speed are all still online.

WordPress.com's AI Website Builder for Agencies promises "fully-designed, content-ready websites in seconds" and offers to "save hours on repetitive site setups." The single numeric figure in the whole post is "30 free prompts." How many hours, out of how many, is never stated.

Straydog Branding's guide to how creative agencies use AI for client websites says AI "dramatically speeds up the concept phase." Its own FAQ asks the question directly — do AI tools reduce agency costs? — and answers that they "can reduce production time and streamline workflows." Across fifteen hundred words there is not one hour, one percentage, or one before-and-after.

The most useful piece in that search result is the one that admits the least speed. Rosanna Bossom's post on how she uses AI in her website design business lists briefs, brainstorming, content repurposing, cleaning spreadsheets, and transcription — and then draws a line: "I personally DON'T use any AI tools in my actual design or website set up process right now; I design my client websites from scratch on Squarespace." A LinkedIn piece by Mark Hinkle sums itself up in the same spirit: AI will not build your entire website from a single prompt, so keep a human in the loop for the polish.

None of these are dishonest. They are just all measuring the same leg, in adverbs, and none of them follows the project to the end. Search for this and the first result Google returns is not a product page at all. It is a forum thread asking whether anyone has actually done it.

The last mile is where the days still hide

Look again at the fifteen hours that survived the 10×. A chunk of them are not craft. They are logistics, and they are the reason a finished folder of files still is not a delivered project.

Publishing. The files exist and nobody can open them. Most hosting was designed for applications, so the shortest published route to a public address tends to start with a repository and a build configuration a landing page does not have. That is the whole argument behind vibe deployment: a five-page site and a product with accounts and background jobs should not share a path to production. If the site currently lives inside a chat, it is worth being precise about what you actually have. A published artifact is a link on someone else's domain, not a site your client owns.

The domain. Half an hour of work, and then a wait that belongs to the internet rather than to you. DNS records carry a TTL, and what you are waiting on is the old answer expiring from other people's caches. Squarespace tells customers 24 to 48 hours. GoDaddy is more honest and says a few minutes to several hours. The mechanism, and what actually determines which of those you get, is in our piece on connecting a custom domain in one record. This is the one stage of a client project that no model will ever compress.

The edits that come later. A client will ask for a changed phone number three weeks after sign-off. Whether that costs you five minutes or an afternoon depends entirely on whether your host lets you change one file on a live site or expects you to rebuild and republish the lot.

The handover. Who owns the domain. Where the enquiry form sends its submissions. Whose payment account is behind the buy button. Who makes the next change. This stage did not shrink at all, and pretending otherwise is how freelancers end up doing unpaid support for two years.

What a three-hour client site actually requires

The headline is achievable. It is just conditional, and the conditions are worth stating plainly, because they are what separate the demo from your Tuesday.

• It is one page, or a very short site.
• The client has supplied the copy and the images, and has the rights to them.
• The design decisions are already made before the first prompt. Color, tone, what the page is for.
• The address is either the platform's or a domain that is already delegated and resolving.
• Your agent is already connected and authorized. That is a one-time step, and if you are counting it in every project you are counting it wrong.
• There is no enquiry form, no checkout, and nothing behind the page.

The page still needs a coherent commercial sequence. Our review of five one-page website examples shows how offer, proof, objections, and one next action can share a URL without turning the scroll into an archive.

Meet all six and three hours is not optimistic. Miss two of them and you are back to a week of calendar, most of it spent waiting for someone else.

What still takes days, and always will

The stages AI has not touched are not technical, which is why no tool announcement will fix them.

Client decisions still take as long as the client takes. Content is still the single most common reason a site sits unfinished. Approval rounds still run at the speed of email. If the domain needs its nameservers changed rather than a record added, the wait is genuinely slow and always was.

And most things with a backend behind them, a checkout, a login, are still real work with real consequences, and it is still yours to arrange. A working enquiry form is the one exception: Birta receives those directly now, no separate service to wire up. The rest is the honest remainder of every client project, and generating the page faster does not touch it.

What changes is your week, not your price list

The useful conclusion is not that you now build the same site cheaper. What changed is the shape of the week: the largest single block of focused production work stopped being the largest block. Whether that turns into more projects, more attention on positioning and content, or simply fewer evenings, is a decision rather than an outcome.

What it should not turn into is a promise of ten times faster delivery to a client whose copy has not arrived. Pricing is a separate question with a separate answer, and this piece is not it. How much to charge for a website when the AI did the building is.

Where Birta fits

Birta claims the publishing stage, plus one more most hosts leave out: a form that actually receives something. Worth being exact about both.

A connected agent publishes the finished site to a public HTTPS address and publishes it again after each round of changes, without a repository or a build pipeline in between. Edits go file by file through the same connection, so a one-word change does not mean republishing everything. Images and fonts travel with the site; video and PDFs still need a public URL somewhere else. Every publish is kept as a version, and a bad one can be rolled back. You can point a domain at it with HTTPS, and each project has its own analytics. The same agent can declare an enquiry form and get its submissions landing in your panel, without a separate form service. The agent is authorized in the browser once, not per project.

For a client site whose whole point is bringing them enquiries, the lead-gen page walkthrough covers what belongs on the page and which plan limit the client hits first.

The mechanism behind that connection is MCP, explained without the infrastructure vocabulary, and the exact steps for one tool are in the Claude Code deployment guide.

What Birta does not do is the rest of the list. It will not collect the client's copy, sit through the approval round, make the DNS wait shorter, or run the checkout for you. Those are still the freelancer's hours, and they are now the majority of the project.

Frequently asked questions