Abstract
The debate about AI replacing maintenance labour is stuck on the wrong question. It assumes AI must first become good at prediction and diagnosis, which are the hardest things maintenance people do. This paper argues that prediction is not required. A system needs only to recognize that something is not working correctly, execute preventive maintenance, and replace the unit that is failing. All three are within technology that has existed for years. What is missing is not intelligence. It is architecture. If plants were built so that every functional element is a swappable module, the entire diagnostic layer becomes optional, and with it most of the labour that layer exists to support. The cost of replacing parts before they are fully used is real and is very likely smaller than the cost of the diagnostic apparatus it retires, including on expensive machinery.
The claim
AI does not have to get better to take these jobs. It has to be pointed at a different building.
Every serious argument about why industrial AI cannot replace tradesmen, including twenty-two papers of mine, runs through the same territory: prediction is not diagnosis, evidence is unreliable, the record does not describe the machine, verification is the bottleneck, the authority to act is not the same as the ability to answer. All of that is true, and all of it describes a plant that has to work out what is wrong.
Take that requirement away and the argument collapses.
Three capabilities, and only three:
- Recognize that a unit is not performing correctly. Not when it will fail. Not why. Just that its output is outside what it is supposed to produce.
- Execute preventive maintenance. Scheduled, repetitive, defined physical tasks.
- Replace the unit. Not repair it. Pull it and put in a new one.
None of that is at the frontier. Detection is comparison against a known-good band, which is the oldest trick in process control. PM is a calendar and a set of repeatable motions. Replacement is a defined procedure with a defined end state, which is precisely the kind of work robotics has been good at for decades.
The reason it has not happened is not that the machines are not smart enough. It is that our plants are built to be repaired, by people, in place, by someone who first has to figure out what is wrong.
Detection is not prediction, and the difference is the whole argument
This series has spent considerable effort on the distinction between prediction and diagnosis. There is an earlier and much cheaper distinction underneath it.
Prediction asks when a component will fail. It requires a model of degradation, history, mechanism and operating context, and it is genuinely hard. It is also what everybody has been selling.
Diagnosis asks what is wrong and why. It requires mechanism, evidence and the discriminating observation. Harder still, and the subject of half this series.
Detection asks whether output is outside the band. That is it.
A motor drawing more current than it did last week. A conveyor running at a different amplitude. A pump not making pressure. A gearbox running hot. None of that requires knowing anything about why. A first-year apprentice can be told what normal looks like and flag when it is not, and so can an instrument that costs very little, and so could an instrument that cost very little in 1995.
I once spent a call-out at an oil site on a pump that was not making pressure. The site was right. It was not putting out what a pump that size should. That reading was correct, measurable, and completely useless on its own, and I had to shut the machine down and watch the shaft slow to find out the thing was running backwards and had been for six years.
Here is the point that story usually gets used to make: detection told them nothing. Here is the point it makes in this paper: detection told them everything they would have needed, in a plant where the answer to a pump that is not making pressure is to pull the pump and put in another one. Six years of degraded performance ended the moment somebody replaced the unit. Nobody needs to know about reverse rotation and impeller clearance in a building where the response to out-of-band is swap.
The knowledge I was carrying up that hill was valuable because the alternative was expensive. It is only valuable while the alternative is expensive.
This is not speculation. Aviation has done it for decades.
The strongest evidence that this works is that the most safety-critical maintenance environment in the world already runs on it.
A line-replaceable unit is a modular component designed for rapid replacement at the operational level without specialized tools, extensive disassembly or removal of the surrounding system.[1] It is typically a sealed box, exchanged by opening and closing fasteners and connectors. When a fault is detected, the suspected unit is removed, a serviceable one is installed, and the aircraft returns to service. The removed unit goes to a repair shop for diagnosis at a lower level.[1]
That is two-level maintenance, and notice what it does to the skill distribution. Diagnosis has not disappeared from the world. It has been removed from the site and concentrated in a shop where conditions are controlled, the same unit arrives over and over, and the economics of expertise work. The line does detection and replacement. The shop does diagnosis and repair.
Aviation did not adopt this because line mechanics were stupid. It adopted it because aircraft downtime is catastrophically expensive and diagnosis at the gate is slow. A plant's economics are the same shape and the industry simply never reorganized around it.
Every objection a maintenance manager will raise to this paper has already been raised, answered and regulated in an industry where the consequence of getting it wrong is a hull loss.
And the process industries are already building this way
The second piece of evidence is that modular construction is not a proposal. It is an existing procurement pattern.
Kinder Morgan's own description of modular skid-mounted treating plants states that the plant arrives fully assembled and tested for faster speed to market, that the ease of future modifications and the ability to move skids to new locations maximizes capital efficiency, that ease of installation and reduced fieldwork save time and money through smaller field crews, and that leased or refurbished modular skids can be returned after the contract period, disconnected at the interconnecting pipes and moved or shipped back.[2]
Read that last part again as a maintenance thesis rather than a commercial one. Disconnect the interconnecting pipes and the unit moves. That is a line-replaceable unit the size of a building. The commercial arrangement, leasing a functional block and returning it, is the replace-do-not-repair model applied at plant scale, and it is being sold today for reasons that have nothing to do with AI.
The pieces exist. Nobody has assembled them with the intention of removing the trade.
Why a plant cannot do this today
Two stories, and between them they contain most of the real barrier.
The replacement conveyor came the wrong length. I do not know if anyone measured it, or measured it wrong. I had to retrofit both the infeed and the discharge to make it fit, and it took the whole line down for an entire week. A sixteen hour scheduled window turned into an unplanned week because the part was wrong.
That is the swap-first model failing, and it failed for a purely architectural reason. The machine was not designed to be exchanged. Its interfaces were not standardized, its dimensions were not guaranteed, and the surrounding structure had been built around that specific unit rather than around a defined envelope. A robot could not have saved that week. The building did it.
And then there is the gearbox problem. A water bottling plant broke down with no spare gearbox. I asked why we could not pull one off another conveyor. Because it is a different gearbox. I looked at the two conveyors and they were basically identical. Why do we have different gearboxes on identical conveyors? Nobody knew. This is what they were sent with. So we waited, with the line down.
Afterward I went through the whole plant. Everything was five-horse motors and gearboxes, and there were something like twenty different kinds. I put in work orders to consolidate the lot down to one type, or a limited set. Instead of hundreds of different spares we kept several. It worked like gangbusters. Less inventory sitting there, and we always had the spare, because everything was the same thing.
The variety was never a decision. It accumulated one shipment at a time and nobody owned the standard.
Standardization is the precondition for everything in this paper. A plant with twenty gearbox types cannot be run on replacement because there is no unit to hold. A plant with one gearbox type can be run on replacement by a system that does not know what a gearbox is. And notice that consolidating cut inventory value and raised availability at the same time, which is the same economic shape this whole thesis rests on.
There is a third barrier worth naming because it is cultural rather than physical. Nobody verifies the part. The tradesman says he needs it, the planner looks it up, orders it, and builds the schedule around a supposed delivery date, and nobody in that chain verifies a thing, not the part, not a measurement. In a repair culture that produces a bad day. In a replacement culture it is fatal, because the entire model assumes the unit in the crib is the unit that fits.
The economics, stated honestly
His claim, and I think it is right, is that the money saved offsets the premature replacement, even on expensive machinery. It is worth being precise about what is on each side, because the argument is usually lost by people who only count the part.
What a replacement-architecture plant stops buying:
- Diagnostic labour, which is the most expensive hour in the building.
- Troubleshooting time, which has no upper bound. A fault that takes twenty minutes to find and a
fault that takes two shifts cost the same to schedule and wildly different amounts to endure. - Unplanned downtime, which in most plants dwarfs every other line in this list. - The scarcity premium on experienced tradesmen, and the risk of losing the one man who knows. - The coordination overhead that WP-20 argues is where the real money already goes: the permits, the sequencing, the shared iron, the crew standing around waiting on a crane. - The readiness debt that makes half of this series necessary. There is no work package to get wrong if the work is pull and replace. - The repair supply chain: the shop, the specialist tooling, the rebuild kits, the skills to use them.
What it starts buying:
- More units, consumed earlier in their life.
- A larger and more disciplined spares holding.
- Design cost, once, to get standardized interfaces and envelopes.
- A shop, or a supplier's shop, to take in the failed units.
The asymmetry is that the first list is recurring, unbounded and paid in the scarcest resource the plant has, while the second is recurring, bounded and paid in dollars for objects. That is the whole argument and it is a good one.
I am not going to put a number on it, and neither should anyone else without a real case. The arithmetic is plant-specific and depends on the cost of the unit against the cost of an hour of downtime, and anyone who quotes a general percentage is guessing. What can be said structurally is that the more expensive the downtime, the better this trade looks, which means it is strongest exactly where the machinery is most expensive. That is counterintuitive and it is the part of the thesis most worth defending.
Where this gets expensive, and it is not where people expect
The honest cost is not the wasted remaining life in a discarded unit. It is storage.
The stranders in my plant ran motors so big the mill had to phone BC Hydro every time they started them. Easily six figures apiece. A spare came in, and some genius decided a good idea would be to park it right underneath the stranders, the highest vibration point in the entire mill.
I agreed to it at the time, because there was a mechanism built to keep the shaft up so the rotor did not ride on the bearings. It worked. It actually held the shaft up. With the rotor's weight off the bearings the vibration had nothing to chew on.
Then somebody removed the mechanism. No record, no rationale anyone could point to. The motor sat there, full rotor weight on stationary bearings, vibrating, for at least three years. Then another mill needed it. They shipped it, installed it, and it lasted about three months.
That is false brinelling. Three years of the stranders hammering the rollers against one spot on the races, lubricant squeezed out of contact points that never moved, wear divots pounded in at exact roller spacing. The bearing was destroyed years before it ever turned, and the failure showed up at a different mill where nobody could see the cause. I never found out whether they worked it out. My guess is they blamed manufacturing, because that is what usually gets blamed.
For three years the storeroom said spare motor on hand, condition good, while reality said paperweight, failing in place.
A replacement-architecture plant holds more spares, and a spare is only a spare if it would survive being installed. Scale the holding up and this failure mode scales with it. Storage becomes a maintenance activity in its own right: shaft supports, scheduled rotation PMs, storage location treated as a reliability decision rather than a floor-space decision.
Which is fine, and is the point. That is a defined, repetitive, schedulable task. It is exactly the kind of work the system in this paper is already doing. The thesis survives its own worst objection, but only if somebody designs for it, and the industry's track record on designing for storage is the motor under the stranders.
What this means for the trade
I am not going to dress this up.
The knowledge that makes a good millwright valuable is diagnostic knowledge. Knowing that a bearing can be destroyed by not running. Knowing that a pump can run backwards for six years because tight impeller clearance compensated until it wore away. Knowing which variable a ceiling belongs to. That knowledge is expensive to acquire, it does not travel in an SOP, and it is worth what it is worth because the alternative, replacing everything, has historically been unaffordable.
The thesis in this paper is that the alternative is becoming affordable, and that when it does, the value of diagnostic knowledge does not fall gradually. It falls the way any premium falls when a substitute arrives.
The work does not vanish. It moves, the way aviation moved it: off the line and into the shop, where fewer people do it on units that arrive in batches. That is a smaller trade and a differently distributed one, and the people who go to the shop will be the diagnosticians. Everyone whose job is to find out what is wrong on site is doing a job the building was designed to require.
That is not a prediction about ten years from now waiting on a breakthrough. Every component of it works today.
What this does not claim
It does not claim every plant can be rebuilt. Brownfield sites carry decades of accumulated non-standard iron, and the conveyor that came the wrong length is what retrofitting looks like. This is an argument about how the next plant gets built, not about converting the one you have.
It does not claim repair disappears. It relocates.
It does not claim the transition is free or fast. Standardization is a multi-year discipline that most plants have never sustained for a single equipment class, let alone all of them.
And it does not claim AI is driving this. The architecture is the thing that matters. A modular plant is cheaper to run with people too. AI simply makes the resulting work, detect and swap and do the PMs, almost entirely automatable, which is the part that ought to concern the trade.
Conclusion
The industry has spent a decade trying to make software smart enough to do what a good tradesman does. It is hard, it is going slowly, and this series has spent twenty-two papers on why.
It was the wrong target. You do not need software that can work out what is wrong. You need a building where nobody has to.
Detection is easy and always was. PMs are easy and always were. Replacement is a defined procedure. The only genuinely hard thing in industrial maintenance is diagnosis, and diagnosis is a cost imposed by how we choose to build plants, not a law of nature.
Twenty kinds of gearbox on identical conveyors was never a decision. Nobody owned the standard. The plant that gets built the other way, one gearbox, one interface, one envelope, does not need anybody to be clever at three in the morning. It needs somebody to notice the number is out of band and bring another one.
We have been able to do that for years.
References
- "Line Replaceable Unit (LRU)," EE-Aero glossary, and IFS Maintenix documentation, "Line and shop replaceable units." Accessed 21 September 2026. https://ee-aero.com/glossary/lru/ and https://docs.ifs.com/maintenix/24r1/Operator/enduser/Engineering_TechRecords/part_numbers_and_groups/c_line_shop_replaceable_units.html
- Kinder Morgan, "Benefits of Modular Skid-Mounted Plants," KM Treating. Accessed 21 September 2026. https://www.kindermorgan.com/Operations/KM-Treating/News/Benefits-of-Modular-Skid-Mounted-Plants