In early 2024, I decided to rebuild my corporate website. Wix was slow, painful to manage, and useless for SEO. I wanted control. I also wanted speed. AI seemed perfect for this.
I used Cline with OpenRouter, picked Next.js, shared my content, gave some guidelines, and let it run.
At first it felt like progress. Then I ran PageSpeed. The SEO score came back at 55. Page load was poor. The sitemap showed errors everywhere in Google Search Console. Blog posts had landed inside service pages. New Next.js features like the Image component were not even used — the model did not know they existed.
This was not a cleanup problem. The foundation was wrong.
What I Thought I Had Delegated
I thought I was delegating execution. What I had actually done was hand over a core decision: decide the right website structure for SEO so I do not have to think about it.
That was the mistake.
The AI created a structure that did not match how my business actually works, did not reflect how I want customers to understand me, and optimised for generic good SEO rather than my specific positioning. It made reasonable choices. They were just wrong for me.
AI optimises for the middle of the bell curve. Operators rarely live there. If your business sits even two steps away from average, the model will not know that unless you force it to. I did not force it to. I let it infer.
The Operator Realisation
AI does not fail at execution. It fails when you ask it to infer your intent, your boundaries, and your non-negotiables.
Those cannot come from prompts. They must be decided and frozen before execution starts. Once I understood that, I stopped prompting and started designing control.
What Actually Fixed It
I scrapped the build and started again. This time I did three boring things before writing a single line of code.
First, I wrote a PRD for AI, not for humans. I defined the why and the what before any execution began. Second, I locked the constraints. Pages, hierarchy, SEO intent, and non-negotiables were decided by me, not the model. Third, I stopped asking for the whole site at once. I built one small feature per session, and every session ended with a summary that I fed back into the PRD.
The PRD became context, memory, and project manager in one document. AI stopped deciding what mattered. It only executed what was already decided.
The “AI will just do it” phase cost 3 weeks and produced a broken site. The controlled system took 1 week to design and 5 weeks to execute. Longer on paper. Faster in reality, because I stopped fixing surprises.
The Boundary
If you are using AI to build anything important and you have not documented your why, your how, and your what, stop. You are not delegating. You are gambling.
Open one AI-built system in your business today and ask one question: what decision have I handed over here without first freezing my own judgment about it? If you cannot answer that in one sentence, the system is already drifting. That is not an AI problem. That is an operator problem.