Post caption
Corporate America is repeating a familiar mistake. In the 1980s, General Motors spent billions on robots that promised efficiency but delivered chaos because the underlying work was never understood, let alone simplified. Today’s enterprises are doing the same with artificial intelligence. MIT now estimates that the vast majority of AI pilots produce no measurable return, not because the models are inadequate, but because the organisations deploying them remain structurally unreadable. Workflows still depend on tribal knowledge, undocumented approvals, and assumptions inherited from another era.
Toyota’s old lessons remain the clearest guide. Automation only works when the process beneath it is stable, transparent, and relentlessly improved through small, local experiments. Companies that skip this step end up multiplying dysfunction at scale, replacing bad work with faster bad work. AI is not the problem. The refusal to redesign the work is.
#economics #ai #business #casestudy
Transcript
This is a lesson from GM's billion dollar automation failure. America seemingly reenacting a $45,000,000,000 mistake from the nineteen eighties. GM thought robots would rescue it from Japanese competition. AI is playing the same role today. Roger Smith automated entire factories before he understood how they worked. The robots welded doors shut, painted each other, and created more indirect labor than they eliminated. And MIT now reports that 95% of enterprise AI pilots produce zero Jidoka or automation, which is automation that refuses to automate anything you haven't mastered manually. Shigeo Shingo said that tier are 23 steps between human work and true automation. GM skipped all 23. Most companies deploying AI also skipped them. Today, we see LLMs get deployed into workflows that still run on tribal knowledge, undocumented approvals, and spreadsheet archaeology, meaning more often than not the processes are opaque and the models are failing. And knowledge work has no takt time, no and uncord, no standard work sequence. Most of the real job happens in hallway conversations, Slack threads, and undocumented exceptions, which is what MIT calls the learning gap. Workers do not know how to translate their tacit work into machine readable steps. Managers think the model needs to be fine tuned when the workflow needs an autopsy. Even worse is when AI is deployed in revenue adjacent areas like marketing instead of cost better, where Toyota would tier, leading to many cosmetic changes rather than structural ones. So the societal risk is not that AI replaces worse, though that is important, it is also the risk that companies never redesign the work, then AI amplifies the dysfunction with GPU budgets attached. So the GM lesson is simple. Do not automate the wrong work. Fix the work in question until it is boring, predictable, and almost insulting in simplicity, then let machines handle it. Start with judoka logic. Detect, stop, fix, remove root cause, insert mistake proofing. In software, this means instrument the workflow before you automate the worse. Because if you train the model on sloppy processes, the model becomes a high speed replicator of said sloppy processes. That was GM's entire decade in one sentence. Toyota proved that small local experiments outperform massive capex bets. GM spent billions while Toyota pooled outperformed massive capex bets. GM spent billions, while Toyota pulled every good every time something looked off. Guess who won?
