Who Gets to Say It's Working?

In 1982, General Motors closed the worst car plant in America. Its factory in Fremont, California was a byword for dysfunction: absenteeism near one in five, routine wildcat stoppages, drinking on the line, grievances by the thousand. GM shut the doors.

Two years later the same plant reopened as NUMMI, a joint venture with Toyota. Same building. Same union. GM rehired most of the same workforce, laid-off troublemakers and all. Within a couple of years NUMMI was producing cars at quality near the top of GM's entire system, with absenteeism in the low single digits (Adler, 1993; Shook, 2010).

The technology barely changed. What changed was who was allowed to stop the line, name a problem, and be believed. Toyota gave every operator an andon cord and the authority to pull it. The productivity was latent in those workers all along — locked up not by a lack of tools, but by a structure of power that told them their knowledge did not count.

The blind spot in the model

Stafford Beer's Viable System Model is the sharpest map we have of what an organisation needs to stay alive. But its most persistent criticism is that it is functionalist — it tells you which functions must exist for viability; it is largely silent on who holds the power inside them (Espinosa, 2025).

This year the gap is finally being closed. Zeini and colleagues (2026), in Systems Research and Behavioral Science, read the VSM through a power-relations lens, borrowing John Gaventa's "power cube". Power wears three faces:

Almost every industrial-AI project negotiates visible power obsessively — access rights, sign-off, governance — and never once names the hidden and invisible kinds. Yet those decide whether an operator shares the fix she found at 2am or keeps it to herself.

Why this is your problem, specifically

When you roll out a new system, you are not deploying a neutral tool. You are making a political intervention — redistributing, whether you mean to or not, who gets to define what "good" looks like. Get it wrong and the pattern is grimly predictable: I have watched a technically-flawless OEE dashboard land on a floor as a weapon — used to interrogate operators rather than help them. The response was quiet and corrosive. People gamed the metric and stopped volunteering the truth. The system worked perfectly and learned nothing.

This is where AI raises the stakes. An algorithm trained on gamed data learns the game. An "AI overseer" bolted onto a low-trust floor doesn't just fail to capture process knowledge — it teaches people to hide it.

The systems view: who defines "viable"

In the VSM, System 5 holds identity and purpose — it decides what counts as "working". The power question is simply: whose voice actually reaches System 5? Where authority flows one way, the line is told, not heard; System 4's "environmental scanning" becomes surveillance; the homeostat ratchets down. Where authority is distributed — operators helping shape what "working" means, AI showing up as a coach not an overseer — the homeostat ratchets up. Same model. Different politics.

Your most dangerous blind spot is not a gap in the data. It is the assumption that "who decides" is settled, and that a new system won't disturb it.

What works: redistributing the right to decide

• The andon cord (Toyota / NUMMI). The most radical productivity device Toyota ever deployed is a piece of string that hands the lowest-status person on the line the authority to halt the highest-value asset in the building. Not a safety feature — a deliberate transfer of power.
• Operator-owned metrics. On the floors where AI has stuck, the operators own the screen. The morning huddle is run by the line, not at it. The moment a metric belongs to the people it measures, gaming stops and truth starts.
• One South African plant, three deliberate moves. A standing right for line leaders to halt a run on quality grounds without prior sign-off; the shift's real problems — chosen by operators — at the top of the daily meeting; a visible rule that no one is disciplined for surfacing a fault, only for hiding one. Modest, un-heroic, political to the core — and the fault reports his AI triage model had been starved of began flowing in.