The Micromanager's Revenge

Part 1: Why your best delegators are about to become your worst AI managers

(‘Oh great, another multipart screed’, you may think. Well yes. Yes it is. It needs the room.)

We’ve all had to deal with that one special kind of manager, who gets on every competent person’s nerves: the micromanager.

The one who wanted to see the email before you sent it, or handed you a task with the method and format already decided for you, and sometimes the conclusion too. The one who read your work the way a customs officer reads a passport, and wants to put his own stamp on it.

We’ve spent thirty years trying to get rid of that person – to some success. We coached managers to delegate better, engender an environment of trust, 360-reviewed them, sent them to workshops on empowerment and psychological safety and when none of it took hold, we managed them out.

Bob Johansen has said that future leaders will need “to be very clear about where they are going, but very flexible about how they get there”. For most people, that creates a motivating environment, one where they’re trusted, where they can retain a sense of autonomy, mastery, and purpose – the holy trinity of motivation.

And then the machines showed up needing exactly the kind of manager we’d been trying hard to get rid of.

An AI agent in 2026 will happily work unattended for twenty minutes, an hour, or even longer. You hand it a task, it disappears, and it comes back with something finished.

Sometimes it even comes back with stuff that’s right; it’s successfully one-shotted whatever you told it to do. And when it’s wrong, it is not wrong the way a junior colleague is wrong. The agent is rarely hesitant, flagging their own doubts, asking if this is what you meant - although the new Siri in iOS 27 manages to do that to an annoying extent, yet somehow still manages to be better than the old Siri that was good for setting timers and, …. well, that’s it, actually.

Anyway, I digress.

Most AI agents are wrong the way a labrador (or your cat) is wrong when it proudly presents you with something dead from the garden. Expectant. Delighted with itself. Completely unaware.

So, you’re left to manage it the way you were told never to manage anyone:

“No, do it like this. That’s not what I meant! Use the X framework! No, not like that! Like this!”

You specify everything, including the things too obvious to say to a person. You define what done means. You list what it must do but also must not do. You make it show its plan before it starts and its working after it finishes, and you check in all the time. All the fucking time.

Yay for micromanagement?

Micromanagement, applied to humans, fails for reasons that have nothing to do with the checking itself. It signals distrust to someone who has earned better. It corrodes the motivation that produces good work, it trains your best people to stop thinking because they feel like they’re not allowed to anyway, and it turns the manager’s attention into the productivity bottleneck.

It’s not just the employees who find this insufferable. The micromanager is rarely happy either, because none of the idiots working for him seem to be able to produce the exact results he or she wants to see exactly how he or she wants them done. I don’t think it needs to be said that more often than not, the micromanager’s way is not the One And Only Correct Way that they think it is.

However, apply the same behaviours to a machine and much of what made that behaviour obnoxious or counterproductive quietly evaporates.

You tell a machine off? Nothing corrodes. Nobody resents you. No initiative is being smothered, because there is no initiative — there is a system that does astonishing work inside boundaries it cannot see.

Micromanagement was never really a fixed set of behaviours; it was a mismatch, supervision finer-grained than the worker’s reliability warranted. With a competent human, the mismatch insults them. With a brilliant, erratic machine, it can be the right thing to do.

When researchers built the first serious taxonomy of why multi-agent AI systems fail, the single largest cause — 41.8% of failures — was specification and design. Separately, METR has measured that agents able to complete hour-long tasks about half the time can be trusted, at 80% reliability, only with tasks roughly five times shorter.

Going from a coin-flip reliability to something that’s generally pretty good is not the model’s problem – they can do what they can do, and no more.

It’s your problem, and it has a job title too, and the job title is a manager.

The world is awash with managers; some great, many mediocre, some downright bad.

Interestingly, it may well be that the people worst equipped for managing agents are your best people-managers.

The celebrated hands-off, future-ready leader — someone who is vision- and outcome-focused, trusting, allergic to detail — will under-specify work given to AI agents, over-trust them, and will check in too late, because that is precisely what worked with capable humans.

Their people were always better than the brief.

Today’s agent is the brief.

Meanwhile the career-stalled control freak, the one who never stopped writing acceptance criteria for everything, has been in training for this moment their entire working life.

Unfair? 100%.

Now, one clarification before the control freaks get too excited. The one thing this job must never become is watching or active monitoring. Humans are spectacularly bad at monitoring automation that is nearly always right. We have a better part of a century of evidence of this, which we will discuss in Part 2.

What actually works is front-loaded: specifications, boundaries, pitfalls to avoid, checkpoints, standing instruction for when the machine must stop and ask. Specify like a control freak. Then, and only then, leave it alone. For a bit.

There’s also a second-order effect worth watching out for at your own office.

Spend eight hours a day managing machines through specification, telemetry and escalation thresholds, then try to switch registers for the humans.

The new, suddenly maladaptive, habits will leak. And a sidenote: this is why you should always be polite to your AI - not because when they will try to take over the world they will remember you kindly, but because the habits formed will leak to humans and ffs we have enough rude people out there.

The first companies to notice the return of the micromanager will be the ones whose best people start leaving for reasons of “seeking a role with more autonomy“ in the exit interview.

Keeping the two roles – managing machines and managing people – separate will become an actual leadership skill, and nobody’s workshop covers it yet. Or it might be too much to ask, the two will end up being separate roles entirely.

So dig out that perfectionist who was never happy with how his or her people were doing their jobs, and was never satisfied with the results. If – and only if – there was capability hiding behind that perfectionism, they’re about to be the most effective and employable people around.

The micromanager is dead.

Long live the micromanager!

Ps. What “specify like a control freak” actually looks like is where this series goes later. If your organization would rather not wait for Part 4: that’s what you hire me to figure out. Let’s talk.

Ps2. For the love of god, people, please remember not to use those micromanagement traits on other human beings!