A bolt worth roughly two cents stopped a high-volume production machine for the better part of a month. No sensor tracked it, no model predicted it, and no spare sat in a drawer. The part had to be ordered from a specific supplier, and the line stayed dead until it arrived and a technician installed it.

That account comes from Michael Podgortsev, a director of data and AI strategy and former CTO, who began his career in industrial engineering building predictive-maintenance systems from sensor data on industrial machines at a large manufacturer. He says the failure became the most expensive single event at that customer that year, despite being the cheapest and rarest component involved.

The system itself worked as designed. It watched wear-prone parts, learned the patterns that preceded common failures, and dispatched technicians to fix machines during planned downtime rather than losing days to breakdowns. For the failure modes it was built to see, it performed exactly as intended.

The bolt broke outside that envelope. Nobody had instrumented it, because putting sensors on a two-cent part that fails once in a blue moon made no engineering or budget sense when the team focused on expensive, predictable wear items. The failure never appeared as a warning. It appeared as a machine that was suddenly dead, with no alert ever fired.

Podgortsev argues the problem is structural rather than a mistake anyone made. Every predictive system is built around the failures its designers expect. Engineers instrument the parts they know wear out, train models on patterns they have seen before, and optimize for the common case. That approach is rational and productive — right up until the failure nobody modeled arrives.

The deeper trap, he writes, is that monitoring priorities tend to follow failure frequency. That instinct is backwards for the failures that hurt most, because the cost of a breakdown has almost nothing to do with how often it happens. A two-cent bolt failing once can cost more than a wear part replaced on a quarterly schedule, since the damage sits not in the part but in the downtime and the response readiness behind it.

"Frequency tells you what to expect. It tells you nothing about what will actually hurt you," Podgortsev writes.

He has watched the same pattern repeat across industries since moving from industrial engineering into data and AI, in systems that have nothing to do with printing presses or bolts.

The practical prescription is an audit of the uninstrumented. Before the next predictive-maintenance investment or monitoring overhaul, he recommends walking the machine and identifying parts with no sensors — not because they are unimportant, but because they seemed too small or too rare to bother with. The second filter is recovery speed: which of those parts, if they failed at the worst moment, the plant could not fix quickly because the component is specialized or the response is not staged.

That intersection — the failure you cannot see and cannot quickly recover from — is where the next month-long outage is already sitting, in his framing.

Podgortsev does not dismiss conventional monitoring. Catching known failure modes is necessary, and systems that do it well earn their keep. But a system optimized only for the average, expected failure leaves the operator exposed to the one nobody thought to watch, and on a factory floor that exposure is measured in weeks of downtime and an unplanned line on the profit-and-loss statement.

What to watch next: maintenance teams applying criticality analysis on downtime cost rather than failure frequency, and sensor coverage expanding to long-tail components with extended lead times. The plants that audit their unmonitored parts before the next rare failure — not after — will be the ones that keep the two-cent bolts from becoming month-long outages.