The same dashboard. Two completely different meanings.

Walk onto any modern shop floor and you will find the same screen mounted above the line. To the plant manager who commissioned it, that dashboard means visibility — at last, a real-time view of what the line is actually doing. To the operator standing beneath it, the very same screen can mean something else entirely: I am being watched.

Same pixels. Same data. Two irreconcilable meanings.

Long before process knowledge has a chance to accumulate, most AI programmes fail a quieter test — the human feasibility test. They are technically sound and culturally dead on arrival.

The test no business case includes

Peter Checkland, who spent thirty years developing Soft Systems Methodology at Lancaster, made one observation that should be printed on the wall of every operations meeting: problems in human organisations are not objective things waiting to be solved. They are constructs — they look different depending on the worldview, the Weltanschauung, of the person looking (Checkland & Poulter, 2006).

A dashboard is a perfect example. SSM's CATWOE lens pulls the two worldviews apart cleanly:

Checkland's method demands two separate tests before any change goes live: is it systemically desirable, and is it culturally feasible? Manufacturers are fluent in the first and almost blind to the second. The business case proves the technology works. Nobody checks whether the people it lands on will let it work. That second test is where most industrial AI quietly dies.

What the cybernetics community just admitted

In July 2025, Andreas Slogar published a notable extension of Stafford Beer's Viable System Model: the Human-Centric VSM. Beer's classic five-system model — brilliant for diagnosing structure and information flow — has a blind spot: it tends to treat psychological safety, perceived fairness, trust and group cohesion as noise to be filtered out, rather than as live forces that decide whether the system survives. Slogar's fix adds a "System 6 — Observation": an explicit function for watching the psychological and cultural state of the organisation.

The deep point is the one that matters for AI: people are not a variable to be managed around. They are self-organising components of the system's viability. Bypass them, surveil them, talk past them — and you have not removed noise. You have removed the very capacity that lets the line respond to the unexpected. On a factory floor, the physical-world feedback loop that grounds every algorithm is the experienced operator (Farrell, 2025). Disengage her and you have severed the AI from the only sensor that understands what the data actually means at 2am, mid-shift, when something is subtly wrong.

The fastest way to destroy the process knowledge your AI depends on is to make the people who hold it feel surveilled.

What works, and what doesn't — on real floors

What doesn't work: the dashboard installed over people. A perfectly good OEE system, commissioned with genuine good intent, became a weapon within a fortnight — supervisors asking "why did you stop the line at 14:20?" rather than "what did the line need at 14:20?" The operators' response was rational and immediate: they stopped recording honest stoppage reasons, and the expensive new data turned to mush. Technically the system worked flawlessly. Humanly, it was dead. The numbers were worse than the clipboard it replaced — because the clipboard at least had the truth on it.

What works: the same technology, owned with the people it measures. Toyota's andon cord remains the cleanest example in the world — any operator can stop the entire line, and pulling it triggers help, not blame. The information flow is identical to a modern alert system; the worldview is the opposite. Siemens' Amberg electronics plant succeeded with AI inspection for the same reason: the analytics were embedded inside an existing Kaizen culture where data already belonged to the team.

The difference between the two is not in the hardware. It is entirely in who the screen belongs to.

Idealised design: build it backwards, build it together

How do you engineer for the human test rather than hope for it? Russell Ackoff's Idealised Design: instead of extrapolating forward from today's mess, gather the people who will live with the system and ask them to design the ideal version as if it could be built tomorrow — then work backwards to what is feasible now (Ackoff, 1981).

The power is not the elegant end-state. It is who is in the room. When operators help design the dashboard — choosing what it shows, who sees it, what an alert triggers — the ownership question resolves itself. The screen stops being something done to them and becomes something built by them. You cannot feel spied on by an instrument you helped specify.

Five things you can do before the next screen goes up

1. Run the second test. Before any AI go-live, ask explicitly: is this culturally feasible, not just technically sound? Name the operator's worldview out loud.
2. Decide who owns the screen. If the honest answer is "management," you have already chosen surveillance. Put the data in the operators' hands first.
3. Design it backwards, with the floor in the room. One idealised-design workshop with operators will tell you more about feasibility than a year of vendor demos.
4. Make the signal trigger help, not blame. The andon principle: an alert should summon support, never an interrogation. The first three uses of a new system set its meaning forever.
5. Watch System 6. Track psychological safety and trust as deliberately as you track OEE. They are leading indicators of whether your AI will climb the J-Curve or sink in it.

The bottom line

Your AI does not fail because the model is wrong. It fails because, on the day it goes live, the people standing under the screen quietly decided it was against them — and they were never asked. Build the screen with the people it measures. That is the whole game.