In 1984 GM and Toyota reopened a closed GM plant in Fremont, California, as NUMMI, and over 85 percent of the new workforce were old GM Fremont hands. Within a year, by the account of John Shook, who helped design the training on Toyota’s side, quality went from GM’s worst to GM’s best.
Part of the change hung in plain sight: the andon cord. A pull brings the team leader. If the problem can’t be fixed in time, the line stops. At the old GM plant, everybody knew the rule was never stop the line.
GM wanted that everywhere. Jeffrey Liker, a University of Michigan professor who studied NUMMI, remembers a GM vice president ordering a manager to photograph every square inch of NUMMI and make his own plant match. As Liker tells it, the manager knew at once that you can’t take a picture of employee motivation, or of a decent relationship between the union and management.
The first round of changes put andon cords into GM plants. In some, workers who pulled the cord got yelled at. A few plants cut the cords down.
The same rope, a different meaning
The cord was the easy part. What GM could not ship was what a pull meant. At NUMMI it brought a team leader to help. In some GM plants it got you yelled at.
You can copy what a culture looks like. What its people take for granted stays behind.
In a 1983 working paper, and in his 1985 book Organizational Culture and Leadership, Edgar Schein of MIT described culture at three levels. On top are artifacts: office layout, dress, visible behavior, and public documents such as charters and orientation materials. This level is “tricky,” he wrote in 1983, “because the data are easy to obtain but hard to interpret.”
Below them sit espoused values: what people say is the reason for their behavior, what they would like the reasons to be, and often their rationalizations. Underneath both are basic assumptions, so taken for granted that people stop noticing them. By Schein’s account they start as values that kept working, until each became an assumption about “how things really are” and stopped being debatable.
At GM the cord landed on top of an old assumption. A longtime GM manager once explained it: give workers the power to stop the line and they would stop it to sit and play cards. At Van Nuys, a GM plant that tried the Toyota system without Toyota as a partner, some managers also opposed stopping the line because their bonuses depended on how many cars rolled off it, This American Life reported.
Where the levels come from
Schein’s levels are a way to read a culture, and his own account of their basis is modest. In a 1988 paper he wrote that his material came from “my own empirical research and clinical experience as an organizational consultant,” a method that lacks “the methodological rigor of controlled empirical work.” He also argued that questionnaires mostly reach the middle level, because “it is too easy for the respondent just to provide the socially acceptable answer.” By his own argument, the level that matters most is the hardest to measure.
His way in was to look for places where what people do contradicts what they say. Shook, writing about NUMMI in 2010, came at it from the other end. His article’s summary says to start by changing what people do rather than how they think: “It’s easier to act your way to a new way of thinking than to think your way to a new way of acting.”
The first postmortem
Imagine a team lead who wants her engineers to report near misses. She writes “blameless postmortems” into the team handbook and adds a template to the wiki. The first report goes up, and in the review her director’s first question is who pushed the change.
The next near miss never gets written up. In Schein’s terms, the template is an artifact, the handbook line is an espoused value, and the director’s question showed the assumption underneath: when something breaks, find the person. What happens to the first person who uses a new rule tells the team which level is real.
The principles on the website
AI governance has its own top two levels: published principles, review processes, fairness toolkits. A 2023 study at the ACM FAccT conference looked underneath. Sanna Ali of Stanford and three co-authors interviewed 25 people, most of them workers tasked with putting AI ethics into practice at technology companies, and found an environment “where policies, practices, and outcomes are decoupled.”
Ethics workers described product teams as, by and large, “agreeable.” But power lay with product launches. Teams often consulted responsible AI staff at the end, after a model was trained but before it was deployed. Some product teams assumed any fairness fix would “degrade” their models; one worker who often heard that worry said he had not seen it happen in practice. Another described the bottom level: “No matter how much people want to do a good job, at the end of the day, quarterly goals will get in the way.”
It is a small qualitative study. Many people the authors contacted declined, some citing nondisclosure agreements, and it cannot say how common these patterns are. An AI policy can hang on the wall like a cord. The question is what happens to the person who pulls it when a launch date is close.