Premise

A lesson from GM’s billion dollar automation failure highlights the parallel between General Motors’ chaotic robotics investment in the 1980s and the current struggles of enterprise artificial intelligence.

Objective

Under the leadership of Roger Smith, General Motors attempted to automate entire factories before fully understanding their manual operations. The result was a catastrophe where robots welded doors shut and painted each other, generating massive amounts of Indirect Labor rather than the promised efficiency. This historical misstep mirrors the current corporate landscape, where MIT reports that a vast majority of enterprise AI pilots fail to return financial value. This failure is rarely due to model inadequacy, but rather Structural Unreadability—a state where organizational workflows rely on Tribal Knowledge, undocumented approvals, and "spreadsheet archaeology" rather than clear logic.

In knowledge work, the problem is compounded by a lack of Takt Time and Standard Work Sequence, creating what MIT terms The Learning Gap: the inability of workers to translate tacit, messy work into machine-readable steps. Companies often exacerbate this by deploying AI in Revenue Adjacent Areas like marketing—where changes are cosmetic—rather than deep Cost Centers where structural improvements are needed.

Toyota offered a contrasting methodology known as Jidoka, or "automation with a human touch." This philosophy dictates that one must refuse to automate any process that has not first been mastered manually. Shigeo Shingo, a leading engineer at Toyota, articulated that there are The 23 Steps between human work and true automation; GM skipped them all, and modern AI deployments are doing the same. Successful automation requires Jidoka Logic: a cycle of detect, stop, fix, remove root cause, and insert mistake-proofing. This was physically manifested in the Andon Cord, which allowed any worker to stop the line immediately to address quality issues. By prioritizing Small Local Experiments over massive capital expenditure bets, Toyota proved that automation is only effective when the underlying process is stable, transparent, and relentlessly simplified.

Subjective

The parallels between 1980s manufacturing and today's LLM deployments clarify a principle I first encountered through this concept by a different name when John Zdanowski taught me—and drilled into my head through repetition—The first time is handmade. It's the same wisdom that's being illustrated as the cause of GM's automation failure, the initial iteration of any capability must be executed manually before it can be systematized or scaled. Reading about Toyota’s refusal to automate unmatured processes, I realize that John was essentially teaching me a knowledge-work version of Jidoka.

Just as John hand-built over 150 financial models before codifying them into software for 📝Assembled Brands, true efficiency requires that I assume the 📝Individual Contributor (IC) role first. If I try to skip The 23 Steps and automate a workflow I haven't personally validated, I am simply behaving like Roger Smith—replacing bad work with faster bad work. The discipline to validate workflows hands-on is not just an entrepreneurial heuristic; it is the only way to close The Learning Gap before bringing in the heavy machinery.

The first time is handmade

Contexts

#lesson-in-automation
#on-my-desk