I was watching my daughter play the latest capybara-themed game that I got Claude to vibe code.

These little games are a combination of some problem solving, route finding, and elements of maths and spelling - in three languages - all wrapped up in an incredibly cute package of adorable and faintly ridiculous gameplay. We already limit her screen time, and apart from Minecraft I hadn't found many games out there that fit the combination I wanted: fun, healthy, non-scary, capybara-themed (by her request, non-negotiable), and carrying some element of her school work without being contrived about it.

I shared one of the games with a couple of parents in our group. They went: I didn't know AI can do this!

Then the chat turned into a discussion about how to use AI. I suspect I have now accidentally become the AI Guy of that parents group.

The AI-crazy fella

To be fair, I earned that title once before, and rather more publicly.

A year or so back, early in my exploration, I could not shut up about it. At the time it was ChatGPT and Gemini - Claude came later - and it was already helping me with proposal brainstorms, keeping track of the commercial side of my work, first-pass redlining on agreements, and a dozen other things. On the side I was, like millions of other people, using it as a sort of counsellor, vent-catcher, office peer, fellow gossip, and of course the thing you ask at 11pm whether your daughter's small mysterious symptom warrants a doctor.

And at the same time I was designing the lead management system for a client - which is now not just rolled out but sits among their critical daily systems. I found it genuinely startling that I finally had a someone who would brainstorm and argue back with me about features, architecture, data structure, workflow, people, data protection, all of it, at 1am, without getting tired of me.

That excitement leaked. During the Phase 2 and Phase 3 part of a roadmap presentation, I got so carried away about the predictive and data-sanitisation possibilities that one of the Sales Managers gave me a long sideways look and called me "the AI-crazy fella" for the remainder of her tenure there.

I laugh at that version of me now, because the me-of-today would be the first to tell the me-of-that-year to sit down. AI is a tool. An extraordinary one. But it is the hype right now, and too often it gets treated as the answer, without much interest in the journey you have to make before the door to that answer will even open.

A map of somebody else's business

A couple of days ago I shared a rough map of a business, which admittedly, I had only one thorough conversation about so far. From that, I plotted the chain - design, specification, certification, manufacture, packing, logistics, distribution, retail, consumer, returns - and beneath every stage sat the same five layers: process, data, people, systems, and the third parties who hold much of it. Running the full length of it: finance, governance, regulatory, knowledge.

The point was never to describe the business. It was to ask, at each node, what information enters, what moves on, what stops there - and what would become possible if it didn't stop.

And then to attach to every one of those possibilities the thing it actually rests on. Agreed definitions. Records that hold time, so you can tell what was true then and not just what is true now. A contractual right to data that somebody else is holding. Batch identity that survives a handover between two parties who use different names for the same thing. A named owner.

That last part is the whole exercise. A list of what technology could do, stripped of what it depends on, is a brochure. The dependencies are where the work is.

The person I sent it to got there pretty quickly, and described it back to me in a way I really found brilliant: What you're building, he said, is a taxonomy - so the knowledge sits in the organisation rather than in the heads of a few individuals. He'd studied knowledge management twenty-five years ago as part of an MBA, back when the course materials arrived as cassette tapes you played in the car on the way to work. The core hasn't changed, he said. Only the tools.

He was right. The core hasn't changed in twenty-five years. What changed is that the tools finally got good enough to make much of the tedious part possible.

Why AI gets mistaken for the answer

Here's what I think is actually going on with the hype beyond the excitement.

AI gets conflated with the solution because it genuinely does turn up at every single point of the work.

It's the researcher who reads forty pages while you make coffee. It's the note-taker in the meeting who never zones out. It's the council I've written about before - three models arguing with each other so I can find the edges of a decision I have to make alone. Some days it's a development team. Some days it's the QC team poking holes in what that development team just built. It's the executive assistant that keeps the thread from unravelling, the monitor that flags the thing you'd have missed at 4pm on a Friday, and - as of last week - the person who builds a capybara game because a six year old asked for one and nothing on the App Store was quite right.

That's an absurd range. And when one thing shows up in every room, it starts to look like the reason the house works.

But every one of those is an assisting role. Researcher, note-taker, drafter, checker, builder. None of them decide anything. The reason my map was worth sending wasn't the technology layer - it was the dependency layer underneath, which is entirely a question of process, contracts, ownership and human agreement. AI can help you see that layer much faster than you could alone. It cannot build it for you, because most of it isn't a data problem at all. It's people agreeing on what a word means, and somebody's name being written next to it.

Which is why the groundwork isn't the boring bit before the AI bit. It is the bit. Point a very capable tool at an estate nobody has audited in years and you don't get transformation. You get your existing mess, executed faster and with better formatting.

The part that stays yours

So: know when to use it. Know what it's good at. Get the environment ready before you expect much of it.

And stay in the chair. You are still the one who decides what's worth doing, what a good outcome looks like, what to discard, and what you'll put your name to. The tool is astonishing and it is still a tool, and the responsibility for the output doesn't transfer just because something else did the typing.

Speaking of which - it's time I told her the screen capybaras are done for today, and we went out for a walk, in real life, on the streets, hand in hand.