Yesterday I drove to the top of Ilkley Moor and worked from the car. This is not as daft as it sounds. The signal up there is better than in half of London. Bars to spare, decent North Star coffee going cold on the dash, laptop on lap and the whole Wharfe valley laid out through the windscreen like something off a tea towel. Strong connection. Big view. Real perspective.

The thing about a hill (okay, it's technically a Moor). It gives you a view. It does not give you understanding. I could see for thirty miles and I could not have told you one new true thing about the weather I was sitting in. The altitude is a feeling that does a very good impression of insight.

That, to me, is a watchout for many senior people on all sides of the brand and marketing mix, the ones I work with who are "doing AI". The view from altitude.

I've watched how it happens at the top of an organisation. You read the McKinsey deck. You take the briefing. You nod along to the keynote and you come away with a view. A confident, well-sourced, beautifully rendered view of the AI landscape, seen from above, at speed, through somebody else's windscreen.

It feels like perspective. It is closer to sightseeing.

And the numbers are starting to show what it costs to govern from up there. PwC's 2026 CEO survey found more than half of chief executives have seen no financial return from AI yet. McKinsey's read is sharper still: senior leaders are now near-universally familiar with generative AI, and almost none of them are fluent in it. Familiar with. Not fluent in. DataCamp's 2026 work lands the same blow from another direction. The problem in most companies is not getting hold of the tools. Everyone has the tools. The problem is whether anyone can actually use them well.

The most expensive symptom is the quietest one. When you only know a thing from altitude, every real decision about it feels like risk. So the decision becomes no decision. The pilot stalls. The integration slips to next quarter. The view was lovely, and nothing got built.

I always held true the belief that you cannot integrate something you do not understand. Not with any confidence. I was the pain in the arse client back in the early social media days, sat in Meta's London HQ, wanting to speak to the devs, not our perfectly lovely account people just telling me how to spend it, grilling the CEO of the major social listening platform, who'd only popped in to meet his best-paying clients, and sidetracking him down a rabbit hole on the confidence in his sentiment model. (Fess up: I might have had one too many free Verves.) You do not understand it from the top of the hill.

The kid who took things apart

I was always the one who took things apart and put them back together. Usually back together. The radio, the bike, the cassette deck that was never quite the same afterwards. (Luckily, it was my sister's.) The understanding was never in the manual. It was in the guts.

Richard Feynman was that kid too, only better at it. (Significantly, compared to a guy who works in marketing, let's say.) The opening pages of his memoir are just stories about being eleven and fixing radios. There is a famous one where a man's set roars when he switches it on, and young Feynman paces the room thinking instead of touching anything, until the man snaps that he came to fix the radio, not walk about.

Feynman works out the valves are warming up in the wrong order, reverses them, and the thing plays clean. The man went round telling everyone, "He fixes radios by thinking."

Everyone loves that line because it sounds like thinking beat tinkering. It did not. The thinking only worked because by then he had already had his hands inside dozens of the things. The model in his head was built one stripped-down radio at a time. The pacing about was him surfacing. The taking-apart was the deep work that earned the surfacing. When Feynman died, the line still on his blackboard read: "What I cannot create, I do not understand." Mustafa Suleyman quotes it in The Coming Wave under a heading that is the whole argument in four words. To build is to know.

What you only learn with your hands in it.

This week I built a thing.

A daft little site called Slop-Eds.com. A community-judged museum of editorial opinion pieces that read like a language model wrote them and got a human byline anyway. The joke is the point. But the building was the education.

You do not learn what these tools can do by reading about what these tools can do. You learn it the way Ethan Mollick describes, by mapping what he calls the jagged frontier. AI is brilliant at things you would bet against and hopeless at things a child manages, and the only way to find the edges is to keep walking into them. His rule of thumb is around ten hours of proper hands-on use before any of it clicks.

Simon Willison, who has built dozens of small tools this way, puts it plainer. The best way to learn these systems is to play with them until they almost work. He even recommends learning on the weaker models, because they fail more often, and the failures are where the knowledge lives. Slop-Eds.com is doing okay, by the way. A dozen or so submissions, 400+ site visits in a day. It broke yesterday, which I didn't notice for four hours, while trying to automate some more functionality as I feel my way around integrating two new SQL techs (Supabase and Postgres) I had never even touched before Tuesday. I managed to accidentally wipe the server of all files except the home page index. Lesson learnt. I now have Claude Code self-monitoring for outages. It's at this point you're probably reading this and maybe thinking I'm a natural technical geek, doing what they do, some sort of software shizzle. But I'm not. I'm a senior creative brand marketer who enjoys coming down the hill to get my hands dirty and build shit.

That is the bit the hill cannot give you. You go down into the workings, you feel where it gives and where it snaps, then you come back up to see how the benefit actually lands for a real person doing a real job. Down and back up. The round trip is the understanding. Sit at the top the whole time and all you have is a view of a thing you have never touched.

The tempting answer that is still the hill

The obvious move, if you are running a business, is to send everyone on a course. Buy the licences. Write the policy. Mandate the training and tick the box.

It is not enough, and the 2026 data is quite rude about why. Most organisations already offer some form of AI training. The capability still is not landing, because watching a video about a tool is not the same as building something with it that has to work for someone. Courses are better binoculars. They sharpen the view from the hill. They do not get you down off it. The thing that builds the muscle is the unglamorous, hands-in, break-it-and-find-out work that does not fit neatly on a slide. Which is precisely why most people at the top will not do it. The edge nobody will climb down for. The scarce thing, the one that does not commoditise, is not access to AI. Everyone has that now. It is earned understanding. The judgement that only exists in someone who has had their hands inside the machine and felt how it actually behaves.

You can buy the same tools as everyone else. You can read the same decks on the same hill. What you cannot shortcut is the model in your head, and that gets built one stripped-down radio at a time.

The reason this is a real edge and not a motivational poster is that almost nobody senior will pay for it. Climbing down off the hill is slow, uncomfortable, and faintly undignified for somebody with a corner office. The view up there is genuinely good. So most leaders keep governing AI from altitude, forming views, deferring decisions, and wondering why the transformation never quite arrives. When a leader actually uses the tools, teams follow far faster than they ever follow a policy. But that means going down where the signal is worse and the work is messier, and most won't.

Which is the whole opportunity, if you are the sort who climbs down.

Coming off the Moor

I drove back down in the late afternoon, still not having managed to spot the three Adidas sheep that ran off after Leeds United shot their new kit drop video up there.

Best signal in the parish at the top. Best view too. And not one useful thing learned up there that I had not carried up with me. Everything I actually understood this week, I understood at the bottom. Hands in. Site half broken. Reversing the valves to see what happened.

Funny thing about a hill. Brilliant for a signal. Brilliant for a view. Useless for understanding the weather you are sitting in.

You don't learn AI sat on a hill. I think you already knew that. The question is whether you are willing to drive back down.

Asked at the counter

How should small businesses learn AI? By building something with it that has to work for a real person, not by watching a video about it. Fluency lives in the guts of the machine, in feeling where it gives and where it snaps, then coming back up to see how the benefit lands for someone doing a real job. The round trip is the understanding. Pick a small, genuine job and make the tool almost do it.

Do you need a course to use AI at work? A course helps, but it is not enough on its own. Most organisations already offer AI training and the capability still is not landing, because watching a video about a tool is not the same as building with it. Courses are better binoculars: they sharpen the view from the hill, they do not get you down off it. The muscle is built by the unglamorous, hands-in, break-it-and-find-out work that does not fit neatly on a slide.

What is the fastest way to build AI fluency? Get your hands in and walk into the edges. Ethan Mollick's rule of thumb is around ten hours of proper hands-on use before it clicks, and Simon Willison recommends playing with the systems until they almost work, even on weaker models, because the failures are where the knowledge lives. The scarce, uncommodifiable thing is earned understanding, the judgement that only exists in someone who has had their hands inside the machine.