AI built the site in two days. The last ten percent took two weeks.

A couple of weeks ago I wrote about a hackathon build that went well: a live site for a childcare centre, shipped in two days, on a tool none of us had used before. What I did not write about is what happened when the hackathon was over and I looked properly at my own site, the one I had already built with AI months earlier and had been quietly living with.

What “done” had looked like for months

My site had been through several rounds of exactly this kind of AI-assisted build. It worked. The pages loaded, the structure held together, and for months it looked finished. If you had shown it to someone cold, they would have called it done, the same way the hackathon site looked done on day two.

Then I started actually using it the way a visitor would, page by page, link by link, and a different picture appeared.

A hero image was cropped tight enough to cut my face out of the frame on the list view. Blog thumbnails were rendering portrait instead of the 16:9 the layout expected, so every card looked slightly wrong without being obviously broken. A card that was not clickable had a hover animation on it anyway, quietly promising an interaction that did not exist. Links across the nav, the FAQ rows, the contact page, all had inconsistent hover feedback, so the site felt subtly less considered depending on which page you landed on.

None of these were bugs. The code ran. The pages loaded. Every one of them was a judgement call that the tool had no way to make, because making it requires knowing what a visitor actually needs to feel and see, not just what the layout technically requires.

Why the tool cannot see this gap

An AI tool optimises for what you asked it for. It does not know that a cropped face reads as careless to a visitor, because it does not have a stake in how you come across. It does not know that a hover effect on a static card is worse than no effect at all, because it cannot feel the small letdown of clicking something that does nothing. It will happily make a decision that is technically correct and quietly wrong, and it will not flag it, because from where it sits, nothing is wrong.

That gap is not a limitation you fix by using a better prompt or a newer model. It is the shape of the boundary between what AI can build and what a person has to check. The building part compresses. The judgement part does not.

What this means for what you build next

I hear a version of this from almost every business owner I work with once their first agent or automation is running. The output looks right for the first week. Then a client mentions something felt slightly off, or a colleague points out a mistake the tool made confidently and nobody caught, and the owner realises the tool never stopped and asked whether it had got it right. It cannot. That is not what it does.

This is exactly why the workflow you choose to automate first should be one you know intimately. If you cannot spot the ten percent that is wrong, because you do not know what right looks like in enough detail, the gap sits there indefinitely. Not because the AI failed. Because nobody with the judgement to catch it was actually looking.

The next step

Whatever you build with AI next, plan for two weeks of actually living with it before you call it finished. Use it the way your customer or your team would. Look for the thing that is technically fine but quietly wrong. That review is not overhead on top of the AI build. It is the other half of the job, and it is the half only you can do.

If you want a structured way to work through that process with guidance, the AI Agent Workshop is built around exactly this: bringing your own expertise to bear on where the tool needs a second pair of eyes.