Should you hire an agency, or just build your website with AI?

We use AI heavily, every day, to build websites. So take the following as a description of the tool by someone who relies on it, rather than a defence of a threatened trade.

The honest position is this. AI has genuinely collapsed the cost of producing a website. It has not touched the cost of knowing whether the website is any good. Those were always two different jobs, and for decades they were bundled into one invoice, which is why the bundle now looks overpriced.

What AI does genuinely well

Producing the thing. Layout, copy, structure, responsive behaviour, a contact form. Work that took a week takes an afternoon, and the output is often better than what a mid-range agency produced in 2019.

Not getting bored. Alt text on ninety images, consistent heading structure, a schema block on every page. Machines are better than people at the tedious work that actually moves rankings, because they do the ninetieth one exactly as carefully as the first.

Drafting. A blank page is expensive. A first draft you disagree with is cheap and much easier to fix.

If your need is a presentable single-page site for a business that gets its work by word of mouth, and you have an afternoon and reasonable judgement, you can do this yourself now. That is a real change and pretending otherwise would be dishonest.

What it does badly, and quietly

The failures are not the ones people expect. It rarely produces something obviously broken. It produces something plausible.

It does not know what your customers search for. It will guess, confidently, from the general shape of your industry. A local trade often finds that the phrase customers actually type differs from the industry term, and no amount of prompting surfaces that, because the information is in search data rather than in the model.

It cannot see the rest of the internet's version of you. Your Google Business Profile, the old listing with the previous phone number, the directory that has your address wrong. Those determine whether you show up locally far more than your homepage copy does, and a website generator has no visibility of any of it.

It has no idea whether it is wrong. This is the important one. Ask for a page and you get a page. Ask for a page about something you have misunderstood and you get a confident page about the wrong thing. The output looks identical in both cases.

It optimises for looking finished. Models are trained to produce complete-looking artefacts. Completeness and correctness feel the same from the outside, which is precisely the problem.

The failure mode nobody plans for

The expensive version of this is not a bad-looking website. It is a fine-looking website that does not work, undetected.

We see the same handful repeatedly. A contact form that posts into nothing, so the business is thanked for enquiries it never receives. A canonical tag pointing at the wrong URL, so Google indexes one page instead of thirty. A Google Business Profile that was never claimed, so the map listing shows the wrong hours. A site that renders only when JavaScript runs, so half of what reads the web sees an empty page.

Every one of these is invisible from the outside. The site loads, it looks smart, and nothing goes red. The owner concludes that the internet does not work for their industry, which is the most expensive conclusion available.

That is why we sell an audit as a separate thing before anything else: the question "is this working" is genuinely different from the question "does this exist", and almost nobody can answer it about their own site.

The actual decision

Forget the technology. The question is narrower than it looks.

Build it yourself with AI if: the site is a credibility page rather than a sales channel, you can tell the difference between a good result and a plausible one, and you are content to own it, including the day something breaks and you have to work out why.

Pay someone if: the site is where the money comes from, or you cannot judge the output, or, most commonly, you simply do not want this to be one of the things you are responsible for. That last one is not laziness. Time spent learning why a form is not submitting is time not spent on the work you are actually good at.

What you should now refuse to pay for

The honest consequence of all this is that some things people still charge for should be much cheaper than they were, and you should say so.

Building the pages is no longer where the effort is. If a quote is priced as though somebody is hand-writing HTML for three weeks, it is priced against a cost base that no longer exists. We wrote about why websites still cost what they do, and about our own decision to use AI and pass the saving on rather than keep it as margin.

What remains worth paying for is judgement, evidence and someone answering the phone: knowing what to build, proving it works afterwards, and being the person who fixes it in two years when you have changed your prices and cannot remember how anything is wired together.