Everyone tells you how fast you can build with AI now. They're right. That's also the problem.
I built Memolio on my own. It's a personalised illustrated life-story book for grandparents, and the whole thing runs on AI: the writing, the illustration, around 25 automated workflows, the invoicing, most of the regulatory work. The speed is real. I can have an idea at breakfast and a working version of it by lunch.
That is exactly where the bill comes due.
Scope creep is free now, so you order too much
When you have a team of developers, their time forces you to prioritise. Every feature costs something visible, so you argue about which ones are worth it. That friction was doing useful work you never noticed, right up until it was gone.
With AI the friction disappears. I get an idea, I write a brief, I get the brief built, because it's so easy. Nothing in that loop asks me whether I should. So I build things I didn't really need, and every one of them becomes something I now have to maintain, document, and stop from quietly breaking the things around it.
Decision fatigue is the real cost
The heavier tax is in my own head. I make decisions for hours, across tasks where I don't always have the full picture, because so much is running at once. Some days I make a hundred small calls and then genuinely lose track of what I decided. And the decisions I forget are the ones that come back to bite me.
The clearest example: I once spent an entire planning session rebuilding things that had already shipped, because the document I was working from was out of date. The plan looked right. The plan was stale. I spent an afternoon confidently solving problems I'd already solved.
That's the real shape of the problem. It isn't "AI can't code." AI codes fine. It's that the speed lets you outrun your own memory of what you've done.
The fix is boring, and it's the whole game
I started treating myself like the CEO of a company that happens to have one employee, who is also me. A company is mostly processes and documents. So I built the processes and documents.
An activity log of what I've done and when. A tracking system that knows when something is actually finished, not just claimed finished. Session prompts that make work happen in the right order. And an "architect" pass that audits the whole thing for drift, the silent gap between what I think is true and what's actually live.
I won't oversell it. The architect setup is days old. I don't yet know if it holds under real pressure. But progress already feels more linear. Less fixing the same bug three times, more moving toward something new.
The part nobody warns you about
You still need a human in the loop, and that human is you, and you will get tired. The skill that matters most isn't prompting. It's knowing which decisions you have to make yourself and which ones you can hand to the machine, then writing the difference down so you're not relitigating it at 11pm.
The speed is a gift. Just know it's a loan. You pay it back in attention, and the only way to afford it is to build the record-keeping a real business would have had anyway.