AI and leadership development get talked about as though the task is teaching managers to use the tools. The harder change is that a decent chunk of what a manager was for is being absorbed, and what’s left is a different kind of work that most management training doesn’t go near.
What’s left is judgement. Deciding which work should exist at all. Deciding whether an output can be trusted. Keeping a team steady while their jobs change underneath them. Judgement isn’t something you can be taught in a room. It’s built by making real calls that come back to you quickly, in front of people who will tell you whether they held.
Which is why I’d be careful with most of what’s being sold into L&D under this heading at the moment.
What managers were for, and what’s left of it
Strip the middle management job back and a lot of it is throughput and quality control. Allocating work. Reviewing it. Unblocking people. Chasing. Rolling status up so the layer above doesn’t have to ask.
Most of that is either automatable now or will be within a couple of years. Status reporting pulls itself out of the tickets. First-pass review happens before a human sees the document. Summarising a meeting is a solved problem.
What doesn’t go is deciding which work should exist in the first place, deciding when good enough is actually good enough, and noticing that someone on your team is quietly frightened about what all this means for their job. That shift in what the role is for is a large part of why AI leadership training has become a business priority rather than a nice-to-have.
The uncomfortable part is that the work being absorbed is usually the work people got promoted for being good at. We’ve spent twenty years selecting managers for reliability on exactly the tasks that are now cheap.
The capability that replaced it is judgement, not a skill
These get treated as the same thing and they aren’t.
A skill is a repeatable procedure. It improves with practice, you can watch someone do it, and you can tell them in advance what good looks like. Running a one-to-one is a skill. Giving difficult feedback is a skill. Delegation is a skill.
Judgement is a call made under uncertainty where nobody knows the right answer at the time, the feedback is slow, and sometimes it never arrives at all. Deciding whether this AI-drafted client summary is safe to send without someone senior reading it, given who the client is, what it’s worth and who wrote the prompt, is a judgement.
If someone can tell you the right answer in advance, it’s a skill. If the right answer only becomes visible afterwards, it’s judgement.
AI moves managers from the first column to the second. The tool runs the procedure. The person decides whether to trust it, and that decision has no procedure.
Why practice-based training doesn’t produce judgement
There’s a category of product being sold hard into L&D right now: AI-powered simulation and coaching platforms. Unlimited role-play, a safe environment, personalised feedback, scenarios built around your competency model.
They’re decent products. A manager who has rehearsed a redundancy conversation six times handles it better than one who hasn’t. What I’d push back on is the category they’re being sold under, because they’re being bought as the answer to “our managers need to lead through AI” and they can’t be.
Three reasons.
Nothing is at stake. The thing that makes a judgement stick is that the outcome was yours. A simulation can make you uncomfortable. It can’t make you accountable.
The feedback is the wrong shape. The simulation knows the answer. That’s what makes it a simulation. Real judgement calls are made in situations where nobody knows the answer, including whoever debriefs you afterwards, and learning to sit with that is most of the capability.
The scenarios arrive labelled. You’re told a decision is coming and roughly what kind. In the actual job, half the work is noticing there was a decision there at all. Nobody tags the moment where the team started trusting the output without checking it.
Practice builds skill. It’s good at that. The gap in front of us is judgement, and judgement needs decisions you still own afterwards.
What does work: real decisions, a short feedback loop, and people to compare notes with
The useful version of this is less dramatic than the one most people reach for. You don’t need managers to get something badly wrong and carry the scar. You need them making real calls that come back to them quickly, in front of people who will say whether they held.
Here’s how it runs in our own business, which is the only one I can describe in detail.
When we take on something new, the first move is to look at what tools we’ve already got and what we’re actually trying to get out of the process. Then someone tries something. Not a pilot with a steering group, someone just does the thing and sees what happens. The question that follows is whether this is an improvement to the process or whether the process itself needs reworking, and that’s the call the person who tried it has to make.
They don’t make it alone. Every weekly meeting we run has AI use in it. Delivery, operations, biz dev, SLT. Not as an agenda item about AI, but because the processes being discussed in those meetings are now partly AI processes, so you can’t talk about one without the other. Someone brings what they tried, says what it got them, and asks whether it’s enough. Other people tell them. That’s most of what running AI-native turns out to mean in practice.
That’s the whole mechanism. Short cycle, visible result, peers who will push back. Nobody waits a quarter to find out. Nobody finds out on their own.
Two things follow from it that are harder than they sound.
The first is that the calls have to be real and reversible, which means your people need enough authority to change a process without asking, and enough safety to say it didn’t work. We get the benefit because the sharing is honest. If “I tried this and it was worse” is a career risk in your business, you’ll get the meetings without the learning. That’s a structural problem rather than a training one, and it’s the same reason building AI-ready teams rarely comes down to tooling.
The second is that the rate of improvement tracks the amount of talking, not the amount of training. The more we share, the faster everything develops. That’s the bit I’d point at if you asked me what’s changed most about how we run, and it’s also the bit no platform can sell you, because it isn’t a product. It’s a meeting structure and a norm about admitting things.
So when I say judgement needs real stakes, this is what I mean. Not consequences severe enough to hurt. Consequences real enough to be yours, arriving fast enough to learn from, in front of people who’ll tell you the truth about them.
What this changes about where your development budget goes
Three shifts, if you take the above seriously.
Less of the budget on generic management programmes and more of it on structured exposure to real decisions. The courses aren’t worthless, they’re solving last decade’s problem.
Some of the budget has to cover the cost of being wrong. That’s an odd line to put in a training plan and it’s the one that makes the difference. If a manager’s first AI decision has to be right, they won’t make one.
Buy cohorts over platforms. The peer group is doing most of the work, and a platform can’t give you one. That’s the shape of our own AI leadership training: a small group of leaders working on their own businesses, meeting weekly, with coaching on what they tried in between. Same mechanism as the one above, run deliberately.
I should say the obvious thing about my own position. We sell training. The version we sell is cohort-based and project-led, which is the version I’ve just spent 1,700 words arguing for. Read it knowing that, and if you think the argument holds but we’re the wrong people to do it, that’s a reasonable conclusion to reach.
Frequently asked questions
Does AI mean we need fewer managers?
Probably fewer doing coordination, not fewer overall. The coordination load is what’s falling. The decision load is rising, and decisions don’t scale the way reporting does.
Can judgement actually be taught?
Not taught, developed, and the lever is cycle time rather than content. Shorten the gap between someone making a call and finding out whether it held, and make sure the finding out happens in front of other people. A manager who makes twenty small reversible decisions in a quarter and discusses each one develops faster than a manager who attends four days of training.
What’s the difference between AI leadership training and AI and leadership development?
Training usually means a defined set of things a leader should know about AI, delivered over a fixed period. Development is the longer job of changing how someone decides. Both are useful. Only one of them changes what a manager is for.
How do we let managers make real AI decisions without something going badly wrong?
Scope the decision rather than supervising it. Pick an area where a bad call is recoverable and visible quickly, hand over the whole decision inside that area, and agree up front when you’ll look at how it went.