Why the smallest parts stay with people
Timing device assemblers build and adjust the mechanisms inside watches, clocks, timers and precision instruments. Most of that work happens under a magnifier, with tweezers, at a scale where a fingerprint counts as a defect. Seating a jewel bearing, fitting an escapement, regulating a hairspring so a movement holds time: each step depends on touch, sight and small corrections made in the moment. Asking whether AI will replace timing device assemblers really means asking whether a machine can do that hand work, on varied parts, at a price a workshop would pay.
Software is not the hard part here. Hardware is. The class of robot that could match a bench assembler is a dexterous humanoid-grade hand with fine force control, not a bolted-down pick-and-place arm running one fixed motion. Cleaning and lubricating tiny components, inspecting them under magnification, and nudging a mechanism until it meets spec are all tasks where position has to be corrected by feel, not by a program written in advance. The cost panel above compares what a machine run would cost against a person doing the same month of work.
Scale matters too. The Bureau of Labor Statistics counts roughly 250 people in this occupation in the US, with median pay of $62,620 a year, and projects employment falling 6.1% between 2025 and 2035 (BLS, 2025). That decline is about demand, offshoring and consolidation in US manufacturing work, not about an AI system taking the bench. A job this small also gives nobody a business case for building a custom robot cell.
What AI does, what it assists, and what stays at the bench
A narrow slice of the work sits in the group AI can handle on its own: 5% of task time, by our scoring. That is paperwork and data work. Logging production counts, pulling readings off test gear into a record, and flagging a unit that drifted outside tolerance are all things software already does without a person in the loop.
A similar slice is assisted rather than automated: 0% of task time. Machine vision can grade a magnified image of a component faster than a tired eye. Automated timing testers can run a movement against a master standard and chart the error curve. In both cases the tool reports; the assembler decides what to change and makes the change.
Everything else is human work: 95% of task time. Assembling and fitting the mechanism itself, adjusting hairsprings and balance wheels to bring a device to spec, cleaning and lubricating parts measured in fractions of a millimeter, and reworking a unit that failed test for a reason nobody wrote down. The overall coverage figure, 5 out of 100, is built from that split; the coverage method page sets out how task time is weighted.
What the evidence actually shows
The evidence grade for this job is D, our weakest grade. It means no published study has tested an AI system or a robot against a trained assembler on these tasks, so we give no quality parity number at all. The grade is a statement about missing measurement, not a claim that machines failed.
General robot-manipulation research is improving, and so are language models that read drawings and specs. Neither tells you whether a machine can regulate a mechanical movement to a few seconds a day. What would settle it is a timed bench trial: a dexterous robot assembling and adjusting standard movements, with yield, accuracy after a week of running, and rework rate measured against human assemblers, published with its method. Until something like that exists, the parity question stays open. Our quality parity method explains why a D grade never gets a score, and the full scoring method covers how the three questions fit together.
Good to know: an old automation probability quoted around the web for this job came from a 2013 modeling exercise, not from any test of a machine doing the work.
When this could change
Most likely after 2046 (8 in 10 of our scenarios). The replacement-year method explains what that window is measuring and how the scenarios are built.
Two things could pull it earlier. General-purpose robot hands with fine force feedback could fall far enough in price that a contract manufacturer buys one for mixed small-part work rather than one product. And product design could shift: movements engineered for machine assembly, with fewer hand-fitted parts, move work out of the trade without any robot matching a human at the old task.
Two things hold it back. The occupation is tiny, so there is no volume to justify a bespoke automation cell, and fixing an error in a mechanical assembly still costs more than preventing one. Repair, rework and short runs also carry constant variation, and variation is what dexterous automation handles worst today. The guide to humanoid robots and physical work goes through where that hardware stands.
How to stay needed in this trade
Lean into the tasks machines are furthest from. Regulating and adjusting mechanisms to a tolerance is the core skill; keep it sharp on more than one movement family. Diagnosing why a device loses time, rather than just confirming that it does, is judgment nobody has automated here. Hand-fitting and repairing non-standard or legacy parts keeps you useful on work that automated lines cannot take.
Two skills travel well. Metrology and drawing literacy let you argue with a test result instead of accepting it. Tending and setting up automated inspection or timing equipment turns the new tools into your tools, which is how most of the assisted tasks above play out in practice.
If you are weighing a move, the closest work sits nearby in the same family. Look at electromechanical equipment assemblers, coil winders, tapers and finishers and electrical and electronic equipment assemblers, or browse the whole assemblers and fabricators family. You can put two of them side by side on the job comparison page, or see where hand-skilled work sits on the list of jobs that most need a person. The headline figure on this page, 86 out of 100 (higher is safer), is explained in full on the Still needs a human method page.