AI Is More Horse Than Plough
It was a Friday afternoon in May, and Derek McCurdy, Ute Fiedler, Steven Gordon, and I were supposed to talk about AI governance.
We didn’t stick to the agenda. We never do. That’s kind of the point.
Steven said it almost as an aside – “AI is more horse than plough” – and then kept talking, and I wrote it down and didn’t say anything because I needed a minute with it.
He was right.
A plough is a tool. Passive. You pick it up, you push it, you put it down. A plough doesn’t care. It doesn’t have energy of its own. It’s just shaped resistance.
A horse is different.
A horse has energy you don’t. It can pull weight you can’t. But it also has its own motion, its own judgment about the ground underfoot, and its own opinion about whether today is a good day to work. You can’t just pick it up and use it. You have to show up for it.
Horse + harness + driver + implement = a system that does things at a scale neither the horse nor the human achieves alone. Take any one of those four things away and you’ve got a mess, or nothing. The harness is the interface. The driver brings judgment and direction. The horse brings power.
“AI is just a tool” treats it like a plough. You invoke it, you apply force, you put it down. I’ve argued this myself. It’s not wrong, exactly. But it’s not the whole picture.
Here’s what changes when you start thinking about AI as a horse instead of a plough.
You have to show up. A plough sits in the barn until you need it. A horse needs tending whether you’re using it today or not. The people who get the most out of AI aren’t the ones who open the tab when they have a question – they’re the ones who’ve put in the time to calibrate it. They know how it moves, where it drifts, where to trust it and where to push back. That’s not the plough dynamic. That’s the horse dynamic.
The harness matters as much as the horse. Most AI failures I see aren’t about the model being bad. They’re about the harness being wrong. Same model, same task – radically different output depending on how you frame the question, what context you give it, whether you’re asking the right thing in the first place. You can have an excellent horse and a broken harness and go nowhere. The interface is load-bearing. Treat it like it is.
The driver has to know the direction. A horse will go somewhere. It needs you to know where. This is the skill that matters most and gets the least attention in every AI training session I’ve ever seen. Not “what can this tool do” but “where am I going and how do I use this to get there.” That’s your job.
Roughly 3% of Microsoft 365 commercial seats globally are Copilot-licensed. Not active usage – just seats that have the license. Microsoft’s own numbers put it around 15-16 million Copilot seats against 415-450 million M365 subscribers. That’s a procurement signal, not an adoption one. And among the people who do have access, only about a third use it regularly.
Nobody in that Thursday meeting was surprised. Copilot – in its current form, for the range of needs people actually have – often doesn’t work well enough to bother.
A horse that won’t go is just a very expensive plough.
The usual response to low adoption is better training or communications – “people just don’t know how to use it.” Sometimes that’s true. But more often the horse really isn’t going, and blaming the driver doesn’t change that. The question is whether the tool is matched to the work, and whether the harness is set up right. That’s a different intervention than a lunch-and-learn.
Institutions are trying to govern AI the way you’d govern a plough.
Write the policy. Define appropriate use. Issue the guidance. Done.
But you don’t policy your horse. You train it. You build a relationship with it. You make decisions about what work it’s suited for and what it isn’t. You don’t issue guidance – you develop judgment. And that judgment lives in the people doing the driving, not in the document.
This doesn’t mean policy is useless. You still need a corral. But the corral is not the point. The work is the point. The corral exists to make the work safe, not to replace the driver’s judgment.
The governance conversation has been backwards. We’ve been designing the corral. We should have been training the drivers.
Driver training looks like what we were doing that Thursday afternoon.
You’re the driver.
It doesn’t do whatever you push it to do. It moves on its own, with its own momentum, and it needs direction. Your direction. Or it’ll find its own.
Drive it or get dragged. Your call.
Steven would say the model’s wrong, because all models are. But this one’s useful.
Horses are willful beasts.
