Why the bench stays with a person
Ask whether AI will replace clock repairers and the honest answer starts with the size of the parts. A repairer strips a movement down, cleans and oils components measured in fractions of a millimeter, fits a new mainspring, then adjusts the balance until the timing machine shows a steady rate. A model can read that timing trace. It cannot seat a jewel, free a bent pivot, or feel when a screw is about to shear.
The second reason is variety. Much of the work is old: a grandfather clock with a worn pinion, a pocket watch whose maker closed a century ago, a part that has to be turned on a lathe because none exists to buy. Every piece arrives in its own condition, and the first real job is diagnosis by hand and eye. New watches are assembled by machine in factories. Repairing a used one is a different problem, and automated assembly lines do not transfer to it.
Most of this occupation happens in the physical world, and the robotics panel above names the hardware class it would take: a dexterous humanoid. That tier does not exist as a product a repair shop can buy. Our guide to humanoid robots and physical jobs explains why fine manipulation lags behind language work by years, not months.
Scale matters too. The Bureau of Labor Statistics counts roughly 1,310 US jobs in this occupation, with median pay of $67,230 and a projected change of about -0.3% from 2025 to 2035 (BLS, 2025). That is a small, slow-moving trade. Who retires and who trains shapes it more than any software release.
What AI does, what it helps with, what stays human
Start with the paperwork. The share of task time our model treats as something AI can do on its own is 0%. That is the shop-floor admin rather than the repair: pricing a job and writing the estimate, logging what came in and what was done, looking up a movement caliber or a parts number. None of it touches the tweezers.
Next, the assist layer, at 18% of task time. Here software sits beside the bench. Timing and amplitude readings get interpreted faster. Reference material, service manuals and past repair notes are easier to search. Image tools can help flag a hairline crack or a worn tooth profile when you photograph a part under magnification. The judgment, and the hand that acts on it, stay with the repairer.
Everything else is human work: 82% of task time. Disassembly and reassembly, cleaning and lubrication, truing a balance wheel, fabricating a replacement part, regulating a clock in the customer’s home, and explaining to an owner what a repair will cost against what the piece is worth. Across all of it, the Can AI do it? score lands at 10 out of 100. That figure is a share of task time, not a prediction, and the coverage method page sets out how it is built.
Good to know: cost is not the limiting factor here, as the panel above shows software tools are cheap next to any hands-on alternative; capability is what is missing.
What the evidence actually shows
No one has tested a machine against a watchmaker at the bench. That is why the Is it better than a person? grade reads D, the level we reserve for work with no direct head-to-head measurement, and why no parity number appears on this page. We do not fill that gap with a guess.
What would settle it is specific: a published trial where a robotic system services a standard mechanical movement end to end, with rate and amplitude measured afterward against a trained repairer’s work on the same caliber. Benchmarks for office tasks say nothing about it. Until a trial like that exists, the fair description is untested, not unproven. Our full scoring method describes how evidence grades move when studies appear.
When this could change
Most likely after 2045 (8 in 10 of our scenarios). The replacement-year method explains what that window is measuring and how the range is produced.
Two things could pull it earlier. General-purpose manipulators could get good enough at sub-millimeter handling that a standardized service, such as cleaning and oiling a common modern caliber, becomes a machine task. And watch brands could push more repairs toward module swaps, where a sealed unit is exchanged rather than serviced, shrinking the hand work needed per job.
Two things hold it back. The first is parts: a shop that deals in vintage pieces often makes the component it needs, and no catalog covers that. The second is economics. With roughly 1,310 people in the occupation nationally (BLS, 2025), there is almost no market pull for a specialized robot. Hardware gets built for jobs with hundreds of thousands of workers, not hundreds. You can see how that pattern plays out in the list of jobs that mostly need a person.
How to stay needed at the bench
Lean into the parts of the work that keep arriving as one-offs. Diagnosis of unfamiliar movements, because it rewards experience no manual contains. Fabrication, turning or filing a part that cannot be ordered. And the customer conversation, where you weigh a repair bill against sentimental and resale value and give a straight recommendation.
Two skills compound. Lathe and machining work raises your ceiling, since restoration shops and museums need people who can make metal, not just fit it. Antique clock expertise, including house calls and installation, is work no remote tool reaches. Formal training still runs through watchmaking schools and factory service programs, and a visit to a working shop before you enroll will tell you more than any brochure.
Close trades are worth a look if you want to widen your options: Musical Instrument Repairers and Tuners, Camera and Photographic Equipment Repairers, and Locksmiths and Safe Repairers. All three sit in the same repair occupations family, and most of the shops hiring for them fall under the other services sector. To see how any two of them differ on our three questions, put them side by side with the job comparison tool.