A client asked us this on a call last month: if AI can build a website, why am I paying you?
Fair question. And the answer is not that AI is bad at building websites, because it is decent at it now. The answer is that building was never the expensive part of the job.
Developer adoption sits at 84% using or planning to use AI tools, and GitHub reports Copilot users finishing routine coding tasks around 55% faster. Those numbers are real, and we see them on our own builds. What they measure is typing speed, not project cost.
Where it genuinely saves time
Four parts of a build where the time saving shows up in the invoice.
Boilerplate and repetition
Custom post types, ACF field groups, form validation, responsive CSS for the fifth card layout this month. Work that has been done a thousand times and has one obvious correct answer. This is where the 55% figure holds up.
Debugging
Pasting an error and a stack trace into a model gets you to the cause faster than searching for it. Not always right, but a good first hypothesis, and a first hypothesis is most of the work.
First drafts
Outlines, alt text, meta descriptions, a starting version of a services page. All of it needs rewriting by someone who knows the business, but rewriting is faster than staring at a blank page.
Testing and review
Generating test cases, catching accessibility problems, and sweeping a diff for obvious mistakes before a person reads it properly.
Where it falls over
The failures get less coverage and cost more.
- Code that looks right and is not. AI-generated code fails in a particular way: it reads as plausible, survives a glance, then breaks on an edge case or opens a security hole. Anything touching payments, authentication or personal data gets read line by line.
- Working out what the client actually needs. People ask for a redesign when the real problem is that nobody can find the booking form. Getting from the stated request to the real one takes a conversation.
- Design direction. AI produces competent, familiar layouts. Competent and familiar is fine for a dashboard and useless for a business trying not to look like its four nearest competitors.
- Anything containing a fact. AI writes confident, invented statistics, dates and sources. For a business publishing under its own name that is a credibility problem rather than a proofreading one.
- Accountability. When the site goes down on a Friday afternoon, a tool does not answer the phone.
What the tools do not take on
Back to the question at the top. What costs money on a web project is deciding what the site should say, structuring it so it can rank, handling the case where the contact form silently fails, working out which of twelve client requests actually matter, and being there in eighteen months when a WordPress update breaks something.
AI compresses the typing. It does not compress the judgement, and it carries none of the risk.
A prompt-built site from Wix ADI or Lovable is genuinely fine for testing an idea or getting something live this week. It stops being fine when the business needs custom functionality, real search visibility, or someone answerable for uptime.
What this means if you are buying a website
- Ask what your agency uses AI for. Anyone claiming they avoid it entirely is either behind or not being straight with you. Anyone who cannot tell you where they stop trusting it is worse.
- Expect it in the price. AI has taken hours out of builds. If quotes have not shifted at all in two years, ask why.
- Find out who reviews the output. The value is in what a person does with the draft.
- Do not accept AI-written content unedited. Google does not penalise AI content as such, but it rewards content showing real experience. Generic output has none, and your customers hear it too.
How we use it
On a typical build, AI writes a good deal of the repetitive code, drafts documentation, and produces first passes at content that we then rewrite. It does not choose the site structure, write final copy, make design decisions, or ship anything to a client site unreviewed.
That split has held for about a year. The tools have improved a lot in that time. The line has moved less than we expected.
