Why this trade stays with people
Ask will AI replace millwrights, and the task list answers before any model does. A millwright installs, aligns, dismantles and moves heavy machinery inside working plants. That means rigging a gearbox through a door it barely fits, leveling a bedplate with shims, and bringing couplings into alignment with dial indicators or a laser kit until the readings hold. Software can plan the lift. It cannot feel the shim pack seat or hear a bearing complain on the first run-up.
The second reason is the setting. The machine sits where it sits, often in a mill, mine, power plant or food line, with pipework, guards and other trades in the way. Every job is a one-off layout. A robot that could do this would need to climb, crawl, carry tools, and improvise when a bolt shears. Our robotics read for this occupation puts the physical share of the work at 100% and the hardware needed in the dexterous humanoid tier, which is the hardest tier to buy today.
Pay and headcount matter too. The Bureau of Labor Statistics counts about 40,330 millwrights in the US with median pay near $65,700, and projects employment up roughly 0.9% between 2025 and 2035 (BLS, 2025). That is a small, scattered workforce. Nobody builds a specialized machine for a job that changes shape at every site.
What software handles, what it assists, what stays manual
Every task on this job’s list sits in the group that needs a person: 100% of task time. Aligning new equipment, dismantling machinery before a move, replacing worn gears and bearings, and testing a unit under load after installation all land there. So does signaling a crane operator, where the cost of a misread is measured in tons.
No task on this job’s list currently sits in the group AI can do on its own: 0%. That is not a claim that nothing will ever move. It is what the current task-by-task read shows for this trade.
The assist group is also empty on this list: 0%. In practice, plants still use plenty of software around millwrights. Vibration and thermal monitoring flag a failing bearing earlier. Scheduling systems bundle the work into a shutdown window. Those tools change when you are called, and sometimes what you are told before you arrive. The wrench work does not change hands. Coverage, our Can AI do it? score, reads 3 out of 100 for millwrights; the coverage method page explains how that share of task time is built.
What the evidence actually shows
There is no direct head-to-head test of an AI system against a working millwright. Our quality-parity grade for this occupation is D, which is the grade we use when nothing credible has measured the comparison, so no parity number is published here. We do not estimate one from adjacent work.
What would settle it is specific: a documented trial where a machine performs a shaft alignment to a published tolerance on an unfamiliar drive train, a rigging task completed without a human rigger on the floor, or a post-install test run signed off without a tradesperson present. Field data from plant maintenance departments would count. Demo videos in a controlled cell would not. Until something like that exists, this page reports an untested comparison rather than a flattering one. Our full scoring method sets out how grades are assigned and when a number is withheld.
When the picture could shift
Most likely after 2046 (8 in 10 of our scenarios). The replacement-year method explains how that window is built and what it does and does not cover.
Two things could pull it earlier. General-purpose humanoid hardware that handles stairs, ladders and awkward loads would remove the biggest barrier, since the physical share of this work leaves nothing to route around. And a shift in how plants buy machines, toward modular skids that are swapped rather than aligned in place, would cut the hardest tasks out of the job itself.
Two things hold it back. Cost is one. Running AI software against this work prices out at roughly $10 to $520 a month, while the human cost of the same work runs about $1,160 to $2,340 a month, and that comparison only matters if the machine can actually do the task. The other is safety and liability. Lockout-tagout, confined spaces, and suspended loads sit under rules written around a competent person on site. Plants do not sign off a rebuilt drive line on a model’s word.
How to stay needed in this trade
Lean into the tasks that the list keeps with people. Precision alignment and shimming is the first, because tolerance work on rotating equipment is where plants lose money fastest. Rigging and machine moves is the second, since planning a lift inside a crowded building is judgment under real constraint. Commissioning and test runs is the third: you are the one who decides a machine is fit to turn.
Two skills compound on top of that. Read and act on condition-monitoring data, so you arrive at a fault with a hypothesis instead of a hunch. And learn to specify and verify work for others, which is what supervision and contractor oversight pay for.
What to do: ask your plant for access to the vibration and thermal data you are already responding to, and start tracking which alerts proved real.
If you are weighing nearby trades, the closest work sits with industrial machinery mechanics, maintenance workers, machinery and riggers. You can put any two side by side on the job comparison tool, read the wider installation, maintenance and repair family, or see how the manufacturing sector breaks down. For the hardware question behind all of this, the guide on humanoid robots and physical jobs covers what the machines can and cannot do yet, and the list of jobs that most need a person shows where this trade sits among the rest.