Abstract
A work order describes a task. The job is that task plus everything required to reach it, everything discovered on arrival, and everything needed to leave the asset supportable. The difference is invisible in every system of record in common use, which is why estimates are wrong in a consistent direction, and why anything trained on those estimates inherits the error. This paper describes the three categories of unrecorded work, shows how method choice moves large amounts of it, and proposes recording it as a first-class field.
Three categories nobody has a field for
Access work is everything required to reach the job. Scaffold, guarding removal, insulation stripping, draining, purging, and the walk to a crib on the other side of the building. On many jobs it exceeds the recorded task, and none of it appears on the record.
Discovery work is what the job becomes once it is opened. The scope written at a desk describes the symptom. The scope found at the machine describes the fault, and the two differ often enough that planners quietly pad every estimate to survive the gap.
Restoration work is everything after the repair. Reinstatement, alignment checks, guarding, permit closure, cleanup, and the paperwork. It is the most reliably omitted category, because the job feels finished before it is.
Nobody decided to exclude any of this. The record was designed to answer what was done to the asset, and access, discovery and restoration are not done to the asset. They are done to the building, the schedule and the permit system. A field nobody designed a home for does not get recorded, and work that is not recorded cannot be estimated, planned or improved.
The alignment you never have to redo
Here is the part most people miss. The size of the hidden work is not fixed. It is a function of method, and a good method can delete most of it.
I changed a lot of couplers over the years. The standard objection to how I did it is that you cannot skip the alignment, and doubters raise it every time.
I did not skip the alignment. I never disturbed it.
Leave the bearing housings bolted in place. Clean them where they sit. Drop the new bearings straight in. The alignment never moves, so there is nothing to redo. That is not a shortcut, it is the removal of a restoration step by refusing to create the disturbance that would have required it.
As far as I know that machine ran flawlessly for about six years, right up until the plant burned down. Six years is the proof. If the method were quietly wrong, it would have told you inside a month.
Now the same job done the other way. On a different coupler, contractors slammed it on using a forklift counterweight as a press. I found out later they had not seated it all the way, and it caused problems downstream.
That is the whole argument about hidden work in two paragraphs. A shortcut that skips the right tool does not save time. It moves the cost downstream, into a period where nothing will connect it back to the decision that caused it. The forklift saved an hour on a Tuesday and spent it, with interest, on a day when nobody was thinking about couplers.
There is a smaller example inside the same job that I like because it is so mundane. That coupler takes about fifty-two litres of oil. The recorded task is the coupler change. The actual work includes filling it, checking it, topping it, and checking it again, over and over, and none of that lives anywhere except in the time it takes.
The window that became a week
Discovery work has a signature: a scheduled window turning into an unscheduled outage.
A vibrating conveyor was replaced and the replacement arrived the wrong length. I do not know whether 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 line down for an entire week. A sixteen-hour scheduled window became an unplanned week, because the part was wrong.
Where does that week get booked? Almost always against execution. The maintenance department ran long. The record shows a job that took a week, and the next estimate for a similar job will be built from that number, which describes a failure of verification rather than a duration of work.
The error does not just cost a week. It poisons every estimate downstream of it.
Making access work disappear by moving it
The counterpart to the alignment story is what happens when access work is done deliberately in advance rather than absorbed into the window.
The same heavy lift took twenty-six hours on day shift with more than twenty people on it, and eight hours on graveyard with three of us. Me, another millwright, and a welder. The contractor crew called in for the job walked in to find it already running.
The difference was entirely front-loaded access work. The job was planned ahead instead of organized the morning of. The area was cleared before anyone needed it. The lockout was done by one man before the crew arrived, instead of twenty people queuing for a permit. One man was on signals instead of four people shouting.
None of that is task time. All of it is access work, and doing it the night before rather than the morning of is worth more than eighteen hours on a single job.
Which raises the question the metric cannot answer. On the eight-hour version, was the plant more productive, or did it simply move four hours of access work outside the measured window? Both, and the record shows neither.
What to record instead
Split the duration field into four: access, task, discovery, restoration.
Do not attempt precision. Ranges are enough, and a range is more honest than the false precision it replaces. Within two quarters a plant holds something almost no plant has, which is an evidence base for how much of its maintenance labour is not maintenance.
Three things become visible immediately.
Which assets are expensive to reach, as distinct from expensive to fix. Those are two different problems with two different fixes, and only one of them is solved by better spares.
Which jobs are actually discovery, meaning the scope was never known and the work order was a request to go and find out. Those should have been discovery tasks, scheduled small and separate, rather than full jobs scheduled on a guess.
Which methods create restoration work and which avoid it. Nothing in a conventional record distinguishes a coupler changed without disturbing the bearings from one pressed on with a forklift. Both close. One of them comes back.
Where AI genuinely helps here
This is a good early application, and a rare one where the risk is close to zero.
Extracting a probable access and restoration profile from historical notes, drawings and permit records is retrieval and pattern-matching over documents the plant already owns. The output is a question for a planner, not an instruction to a tradesman. Nothing is written, nothing is released, and no authority question has to be settled first.
Vendors are already circling the adjacent ground. IAMTech markets AI-generated work packs built from drawings, photographs and engineering documents, and Prometheus promotes planning automation on time-to-produce. Both are vendor claims rather than independent validation, and it is worth noticing what the claims are actually about, which is the speed of producing a package rather than the completeness of its content. A package generated in four minutes that still omits access and restoration has automated the wrong half.
The honest test for any of these tools is whether the generated package tells you how to get to the machine and what condition to leave it in. If it only tells you what to do when you are standing there, it has automated the part that was never the problem.
Conclusion
Plants do not have a labour productivity problem so much as a measurement problem that presents as one. The hidden work is real, it accounts for most of the variance in every estimate, and it is absent from the record by design rather than by neglect.
It is also not a constant. The alignment you never disturb does not need redoing. The lockout done the night before does not consume the window. The coupler pressed on with a forklift will cost somebody a day, months from now, and nothing in the system will ever connect the two.
Recording access, discovery and restoration is unglamorous and costs nothing. It is also a precondition for any scheduling system, human or automated, that expects to be believed.
References
- Li, Feng, Chen, Tsai, Sun, Das, El Maghraoui, Lin & Patel, “PHMForge: Evaluating LLM Agents on Industrial Prognostics through MCP-Native, Algorithm-Grounded Tools,” 2026. https://arxiv.org/abs/2604.01532
- IBM Research (O’Donncha et al.), “Evidence-Driven Reasoning for Industrial Maintenance Using Heterogeneous Data,” ACL 2026. https://research.ibm.com/publications/evidencedriven-reasoning-for-industrial-maintenance-using-heterogeneous-data
- IAMTech, “iPlanSTO / AI-powered planning & work pack generation.” https://go.iamtech.com/iplan-turnaroundinsurance/
- Prometheus Group, “From Hours to Minutes: How AI Is Transforming Maintenance Planning and Scheduling,” June 26, 2026. https://www.prometheusgroup.com/resources/posts/from-hours-to-minutes-how-ai-is-transforming-maintenance-planning-and-scheduling