The Machinery is the Tell
Every large system that got built wrong tells you so the same way: by the size of the apparatus assembled to cope with the mistake. The machinery is not evidence of sophistication. It is evidence of a foundation laid wrong, and of an industry that found it cheaper to build on the error than to correct it.
This is usually filed under sunk-cost fallacy, which is the wrong diagnosis. Sunk-cost fallacy is an individual failing to ignore money already spent. What happens at the scale of an industry is stranger and more durable, because nearly every actor involved is behaving rationally. The wrong architecture does not survive because people are foolish about it. It survives because it manufactures a set of individually sound incentives to preserve it, and those incentives, summed, produce a civilization that cannot move.
Consider three.
Boeing froze the geometry of the 737 in the 1960s. The airframe sat low to the ground because that is what airports needed when passengers boarded by rolling staircase. Sixty years later, the only way to hang a modern high-bypass engine under that low wing was to mount it farther forward and higher than the aircraft was designed for, which changed how it pitched, which Boeing corrected with a piece of software called MCAS. The software, trusting a single sensor, pushed the nose down against pilots who did not know it existed. It killed three hundred and forty-six people across two crashes.
The instructive part is not the software. It is why Boeing would not simply design a new aircraft. The true sunk cost was never the aluminum. It was the type certificate and the pilot training built on top of it. “No simulator retraining” was the sales pitch that sold the MAX to airlines. The constituency defending a 1967 decision about staircase height was every carrier that had trained its pilots to the old type. MCAS is the archetype of the whole phenomenon: a layer of software laid over a structural flaw that cannot be removed without abandoning the thing everyone is already trained on.
Telecom did the same with copper, and it repays watching because it is the rare case now being undone. The copper last mile was never just billions of dollars of wire in the ground. Copper was a regulated, depreciating asset with an accounting treatment, a maintenance workforce, and a regulatory framework all constructed to keep sweating it. Overbuilding fiber meant stranding all of that at once. So the better architecture waited behind “good enough” DSL for the better part of two decades, in the country that invented the internet.
It is being undone only now. AT&T is retiring copper across 2025 to 2029, and in March 2026 the FCC finally cleared the regulatory underbrush that had made the retirement slow. The lesson runs in both directions. Undoing a generational architecture is possible, which is the hopeful half. It became possible only when the constituency defending copper had eroded, the workforce aging out and the economics of fiber becoming undeniable, which is the warning. The undoing took twenty years and a regulator finally willing to remove its own rules.
Then there is American healthcare administration, where the machinery is most naked. There is no canonical patient record and no unified payer, so the industry built an enormous apparatus to manage the resulting chaos: medical coding, clearinghouses, prior authorization, eligibility checks, the interoperability standards written to paper over the absence of a shared record. Administration now consumes an estimated 25 to 35 percent of all US health spending, on the order of a trillion dollars a year. In 2023, American hospitals spent roughly twice as much on administration as on direct patient care.
That apparatus is routinely defended as the necessary complexity of a complex domain. It is nothing of the kind. It is the compounding cost of a foundation laid wrong, and it has become a profession, an industry, a set of publicly traded companies whose existence depends on the underlying chaos persisting. When the machinery required to run a system exceeds the system it runs, the machinery is not a feature. It is the tell.
The pattern is not confined to three industries. It is closer to a law of large systems. The internet ran out of addresses because a 1970s research experiment used a thirty-two-bit address space, and the fix, IPv6, has taken twenty-five years to arrive, because engineers invented a workaround called network address translation that made the shortage survivable and therefore permanent. NAT is the internet’s MCAS, a patch that relieves the symptom of a foundational limit and, in relieving it, removes the pressure that would have forced the cure. The world’s banks and governments run their most critical ledgers on COBOL written in the 1960s, and when the pandemic hit in 2020, states discovered they could not update unemployment systems because the people who understood them had retired. The civilian nuclear industry standardized on the light-water reactor, a design chosen in the 1950s for the expedience of powering submarines quickly, and better paths, including the molten-salt work Alvin Weinberg championed and then watched cancelled, were foreclosed less on the merits than by momentum. Economists took the name for the whole phenomenon from the keyboard under your hands, a layout whose only remaining justification is that everyone learned it.
Read across the cases and the mechanism is always the same, in three moves. A foundational choice is made for a reason that is sound at the time: staircase height, a submarine schedule, a research shortcut. The choice is then built upon, and the superstructure recruits a constituency, trained pilots, a regulated asset base, a coding profession, a generation of programmers, whose interests fuse to the mistake. And somewhere along the way a piece of software appears, MCAS, NAT, DSL, the billing apparatus, that makes the flaw tolerable enough to live with. That patch is the decisive event. It converts a problem that would eventually force a rebuild into a condition the system can carry indefinitely. The patch is what turns a wrong foundation into a permanent one.
Which brings this uncomfortably close to home, because enterprise software is running the same play, and is about to run it again.
Enterprise resource planning is the healthcare administration of the corporate world. Because there was never a governed, canonical record of what an operation is actually doing, the industry built a superstructure to cope: custom fields that fork on every upgrade, integration middleware to reconcile systems that disagree, an entire consulting economy whose revenue depends on the disagreement persisting. The machinery is vast, and it is held up as sophistication.
And now artificial intelligence is being bolted on top of that foundation, and the wrapper vendors flooding the market are selling the next MCAS. They are laying a probabilistic software layer over a structural flaw, the absence of a trustworthy operational record, and calling it a solution. The danger is not that it fails. The danger is that it works well enough. If AI makes the broken foundation tolerable, it will do to enterprise data what NAT did to the internet and DSL did to copper: relieve the pressure that would have forced the fix, and recruit a fresh constituency of tools and vendors and trained people with an interest in the foundation staying broken. Once the AI layer is load-bearing, it will be as hard to remove as a type certificate.
There is a narrow window, and telecom illustrates it by counterexample. The foundation can be replaced. Copper proves it. But copper also proves the price: a generation, and only after the constituency defending the error had died. The lesson for anyone building on enterprise data now is not to admire the machinery. It is to fix the foundation before the next superstructure sets on top of it, which means putting a governed operational record at the base, where the truth of what happened is deterministic and auditable, before AI becomes the patch that makes doing so feel unnecessary and therefore renders it impossible.
The machinery is always the tell. The only question an industry gets to answer is whether it reads the tell before or after it has spent a generation building more of it.