The mess is the point

In 1982 General Motors closed its car plant in Fremont, California. It had a reputation as one of the worst factories in the company, known for absenteeism, disputes on the floor and cars that left the line with the faults built in. Two years later it reopened as NUMMI, a joint venture with Toyota, with many of the same workers back on the line and far less of the automation GM was spending heavily on elsewhere. Within a couple of years it was turning out some of the best quality cars in the GM system.

The usual telling credits the Toyota Production System, and fair enough. The detail I keep coming back to is smaller: the cord. Anyone on the line could pull it and stop production when something looked wrong. Through that decade GM tried to engineer human variation out of its factories with robots, while Toyota, in a plant GM half-owned, built a system that treated one worker noticing something odd as the most valuable signal on the floor.

I think a lot of the leadership teams I work with are about to make GM’s choice again, this time with better software.

We built organisations to make people predictable

For a hundred years we’ve designed organisations to make people more predictable. Job descriptions, competency frameworks, hiring for fit, performance curves, and the meeting where the person who asked the strange question quietly learns not to ask it next time. Most of it was done for good reasons. When the work ran through people, consistency was quality, and the person who didn’t fit was a cost to be managed.

That logic has flipped while nobody was looking. The predictable work is precisely what AI now does well and cheaply, and it’ll only get cheaper. What it can’t supply is the thing we spent all that effort sanding off: the person who asks the question everyone else thought was too basic, or the analyst who’s a bit of a pain in meetings because she won’t let a lazy assumption through. I’ve started calling that the mess, on purpose, because it’s the word most organisations reach for when they’re trying to get rid of it. The mess is how a learning system finds out where it’s wrong. Take it away and the system can only get faster at being what it already is.

The design always wins

I’ve said in a lot of rooms that AI is a force multiplier, and that the force we should want multiplied is you, whatever makes you curious, odd or interesting. I still believe it. But I’ve come to think it lets leaders off lightly, because it’s addressed to the individual. You can tell every person in a business that their difference is a superpower and then send them back into an organisation designed to flatten it, and the design will win every time. I say that knowing I’ve done it myself. In September I sat down with my own team and handed out the disciplines each person would own, rather than asking where their particular kind of mess would be most useful. It was quick and tidy, and it’s exactly what I’d argue against in anyone else’s organisation.

There’s a finding from the research on AI itself that sharpens this for me. A 2024 paper in Nature showed that models trained over and over on their own output degrade: generation by generation the rare cases drain away and the model settles on a narrower, blander picture of the world. In one test, text about medieval architecture had turned into a list of jackrabbits by the ninth generation. An organisation that pushes its people through enough standard process, and then trains its AI on what those people produce, is running a slower version of the same experiment. The rare cases were the mess, and the mess was where the new information lived. Feed your systems nothing but your own conformity and conformity is what you’ll get back, at speed.

First generation

Ninth generation

Mess needs direction

I want to be careful here, because “celebrate the mess” can sound like a case against structure, and Toyota is the evidence that it isn’t one. NUMMI ran on standard work. Every task was defined, and that’s exactly what made the odd thing visible, because you can only notice something’s wrong when everyone knows what right looks like. The line was disciplined so the cord could do its real job, which was getting the mess through to someone who could act on it. So the real design question for leaders is where the structure goes. Put it in the work and the processes AI can run consistently, and be equally deliberate about protecting the mess in the people around them.

This is where I think the conversation about AI and people has had the question the wrong way round. We keep asking which human tasks AI can take. The better question is what each brings to the way an organisation learns. AI systems already reason in some ways we can’t, holding more evidence in view and following it further without getting bored or cutting corners on a Friday afternoon, and that gap will widen. What they won’t do well is be unreasonable. People will. We’re emotional, creative and messy, and we’ll back a hunch the data doesn’t support yet or go after something that’s never been done. Shaw saw it in 1903, arguing that “all progress depends on the unreasonable man”. Learning loops run on experiments, and most experiments start with someone being unreasonable. Humans are the messy edge that keeps the loop supplied, and AI is what lets an organisation test what comes back and act on it faster than it ever could. Get that division right and you stop having to choose between an efficient organisation and a human one, and I think that’s the only version of this future most of us would want to work in.

The unreasonable idea

The loop that tests it

The unreasonable idea

The loop that tests it

“Learning loops run on experiments, and most experiments start with someone being unreasonable.”

What leaders build now

Mess without direction

Mess with direction

That changes what leadership is for. The job becomes narrower and harder at once: set the direction so clearly that people know what they’re aiming at, then build an organisation where the strangest useful thing anyone notices reaches someone who can act on it, quickly. Direction is what turns mess into learning. Without it you get a lot of interesting people pulling different ways; without the mess you get a very efficient organisation heading somewhere a competitor will reach first.

In practice it starts with some uncomfortable habits. Not long ago I took my own hypothesis for where a business should point its AI into a session with its leadership team, and then withdrew it so they’d write their own. Theirs is the one they’ll be held to, and that matters more than whether mine was right. It also means looking hard at hiring. When organisations cut graduate and entry-level roles in the name of AI, they’re removing the least standardised people they have, the ones who haven’t yet learned which questions not to ask. And it means changing what gets rewarded, so the person whose awkward question changed a decision gets more credit than the person who delivered the template on time. None of that is soft. It’s the one part of the organisation a competitor can’t buy off the shelf.

So think about the most unusual person in your organisation, the one who doesn’t quite fit the competency framework. Is the way you’re redesigning work around AI making them more valuable this year, or more normal?