I gave AI a fortnight to bore me. Months on, it's gone midnight and I'm still down the rabbit hole.
It started as evening tinkering, a way to switch off from a particularly hectic implementation (yes, I know how geeky that sounds). My expectations were rock bottom; up to then, my experience of AI had been scrolling past the slop on YouTube and Twitter, and cringing at the robot voice on every chia-seed miracle short my mother-in-law sends me.
A few days in, the scepticism was gone. I fired a message off to my brother Ade; you'd be daft not to be looking at this, no exaggeration, it's game-changing, get in front of it now. (Ade, if you're reading, that was quantum leap one.)
For the first few months I wasn't chasing anything in particular. I was a kid in a sweet shop. I built small things; a document tool, some plumbing, a way to capture decisions, partly to be useful, mostly to learn how AI behaves when you hand it real work. Some evenings I pointed it at my own PDI; could it write me a sensible diagnostic script, could it make sense of an instance export. It could, mostly, though never left to its own devices. More on that pattern another time.
There was a catch, though. The same tools that let me build had done exactly the same for everyone else. When anyone can build anything, building stops being the hard part, or the valuable one. I'd already seen where that road ends, because I'd spent months scrolling past it; slop is what you get when capability gets cheap and judgement doesn't keep up.
Then the penny dropped, quantum leap two (I called it exactly that to my brother, which tells you how carried away I'd got). The most obvious thing to point AI at is your own itch. Of course it is. I'd walked straight past it for months; plonker of the year award goes to me.
The one thing I had that couldn't be downloaded, copied or vibe-coded was the stuff already in my head. The workshops. The data models. The scar tissue from the implementations that didn't go to plan. Knowing, before the import even runs, which single relationship in a workbook is going to fail quietly and cost someone a fortnight to unpick. None of that is sitting on the internet for a model to learn. So I stopped pointing the experiments at clever toys, and started pointing them at the work only I could point them at.
Which leaves the question I couldn't leave alone; how much of that work could AI actually carry? Not the coding, plenty of people were already proving that one. I mean everything around the build, the part that quietly fills an architect's week. That's the question this whole experiment has been chasing ever since.
Next post, the theatre. Auditioning imaginary people for imaginary roles in an imaginary company. Completely synthetic workshops with a cast. Not a single real client in sight.
What's your itch? If you've already scratched it with AI, or you're still circling it like I was, tell me in the comments.
First published on LinkedIn, 9 July 2026. The conversation lives there; the writing lives here.