For most of my career, I've led change involving things I couldn't have done myself.

I designed transformation strategies for organisations whose services I couldn't have delivered. I've led technology implementations without ever configuring a system. I redesigned operating models with the support of people who understood parts of the business far better than I ever would.

During the Black Summer fires, when I was leading Australian Red Cross's response in Victoria, almost everything that happened on the ground - the logistics, the volunteer teams and the local response - was done by people who knew far more about emergency response than I did.

That's what leadership is.

You set the direction. You decide what matters. You ask questions, test the advice, hold the standards and back the people with the expertise.

And for a long time, that model worked. I don't think it holds for AI.

That's not to say leaders suddenly need to become technologists.

We don't.

But what is different is that we need enough first-hand understanding of AI to exercise judgement about it.

In the changes I've led before, if I didn't have the expertise, somebody else did.

They might have been clinicians, technology specialists or frontline teams. People who genuinely understood the thing we were changing. I could get briefed by them, ask questions, test their reasoning, notice when they disagreed and then form a view.

AI is different.

There are, of course, people who know a great deal about AI. There are specialists with deep technical expertise, and people building serious expertise in AI risk and governance. But no-one has ten years of leadership experience with this. The technology is literally changing whilst we are learning how to use it. Use cases and opportunities are rapidly emerging. The risks are emerging just as rapidly. And at the same time we are still working out what all of this does to our ways of working, expertise and professional value.

You can, of course, ask AI what to do. It will answer fluently - whether or not it is right.

I've been caught out by this more than once. Knowing that generative AI can do this does not make you immune to it. Something arrives beautifully written and entirely coherent. Even when you intellectually understand the risks of generative AI, it is hard to apply that knowing in every single moment. The problem is often small - a claim that is unfounded, a number that is not quite right, an argument more elegant than the situation warrants.

What catches it isn't always knowledge, although that's obviously the first line of defence. I now have a range of rituals and processes I use to stress-test and validate what comes back - a new way of working I developed because I needed to.

I know some of the places where I can be tempted to trust it too much. I know I need to be much more cautious when I am outside my own expertise. I know what happens when I give it much better context, or ask it to challenge me rather than agree with me.

I have used it often enough, and got it wrong often enough, to start building my own reference points. I think a senior leader in one of my AI Leadership Circles put it beautifully:

“AI is a DIY sport - you actually can't outsource the learning of this.”

I think she's right.

Which raises a slightly uncomfortable question.

If you are not confident enough to use AI on your own work, what reference points are you relying on when you support AI adoption across your organisation?

A briefing can explain the tools, the risks, the terminology and the latest capabilities. It can give you case studies and governance frameworks and tell you what everyone else is doing.

All useful.

But it can't give you the experience of looking at an answer that is eighty per cent right and working out whether the missing twenty per cent matters. It can't quite teach you how easy it is to accept something plausible when you are tired, rushed or outside your domain. Or give you a feel for when AI is genuinely extending your thinking - and when it is simply making your existing thinking sound better.

Those reference points get built by doing the work.

Which is why I think the first thing AI changes shouldn't be your operating model.

It should be your judgement.

Use AI regularly on work you are genuinely accountable for. And I don't mean only low-stakes emails, meeting summaries or the odd query because you were curious. Use it on something you understand well enough to assess the merits of the argument. Something where you know what good looks like or something where it matters if the reasoning is weak, the evidence doesn't hold or a confident fabrication slips through.

That is where the learning actually starts - because then you have to make the calls yourself.

Should I use AI here at all, and what role do I want it to play? What do I need to give it to get something useful back? What might be wrong with what it returns, and how would I even check? And am I prepared to put my name on it?

These are all judgement calls. And knowing what an LLM is doesn't help you with any of them.

There is another reason I think this matters.

People pay more attention to what leaders do than to what they say.

If you are asking your team to experiment with AI, rethink their work, get comfortable not knowing and make new calls about risk, it matters whether they can see you doing some of that too.

If you want them to trust you on this, let them see you doing it. Try things in front of them. Change your view out loud. Be able to explain why you trust AI for one kind of work and deliberately don't use it for another. That's what makes it safe for your team to be honest about what they haven't worked out yet.

The question I'd ask isn't "how do I get my organisation to adopt AI?"

It's: "how am I building the judgement to lead myself, and others, through the AI transition?".