Why this work stays in the building
General maintenance and repair work happens in places that were never designed for machines. A clogged drain in a 1970s apartment block, a jammed door closer, a leaking roof flashing, a motor that smells hot but reads fine on the meter. Each job starts with walking to it, looking at it and deciding what it actually is. That decision rarely comes from a clean data feed.
Two tasks show the problem clearly. Repairing machines, plumbing and building structures with hand and power tools needs grip, balance and judgment in tight spaces. Inspecting drives, motors, belts and fluid levels needs a person who can hear a bearing, feel a vibration and notice the thing nobody logged. Software can read a sensor. It cannot crawl behind a water heater and find the valve someone taped shut three years ago.
The second reason is variety. One worker may patch drywall in the morning and reset a rooftop unit in the afternoon. Building a robot for one of those jobs is hard. Building one that moves between them, in occupied buildings, costs more than the pay of the person doing it today. US median pay for the occupation was about $49,590 a year (BLS, 2025), and the robot does not yet exist at that price.
What AI does, helps with, and leaves to people
The paperwork side is where software already works alone. Logging completed repairs, tracking parts and supplies, drafting work orders and pulling a manual for a model number are all text and records tasks. That slice of task time is 0% of the job.
A larger block is assistance. Diagnosing a fault from symptoms, reading schematics or blueprints before a job, and estimating time and materials all go faster with a system that has seen the model before. Sensor-based monitoring can flag a failing bearing or a drifting pressure reading before it fails. The worker still decides what to do about it. Tasks where AI helps rather than acts make up 17% of the work.
Everything physical stays with people: dismantling and reassembling equipment, replacing belts and seals, climbing ladders, painting and patching, and the safety calls that come with working around live power and water. Those tasks account for 83% of task time. Overall, the share of task time AI can handle today sits at 10 out of 100 on our coverage measure.
What the evidence shows
No study has tested an AI system against a qualified maintenance worker on this job’s real tasks. That is why the evidence grade for quality parity is D, and why we publish no parity number here. A letter grade is not a verdict; it describes how much direct testing exists.
What would settle it is specific. A field trial where a robotic system completes a mixed repair list in an occupied building, timed and graded by licensed inspectors against a human crew. Or a controlled comparison on fault diagnosis, where a model and a technician both work from the same symptoms, photos and meter readings, and both get checked against the actual fault. Until something like that is published and repeatable, the honest answer is that the hands-on half has not been measured. Our full approach is on the methodology page, and quality parity explains how grades are set.
When this could change
Most likely after 2046 (8 in 10 of our scenarios). The method behind that window is set out under replacement year.
Two things could pull it earlier. First, general-purpose robots that can walk, climb a short ladder and use ordinary tools; most of this job’s physical demand falls in the dexterous humanoid tier, so progress there matters more than progress in language models. Second, buildings that are instrumented from the start, where sensors and standardized equipment turn diagnosis into a lookup and cut the hardest part of the call-out.
Two things hold it back. The cost gap runs the wrong way for hardware: software assistance is cheap per seat, but a capable mobile robot with service and insurance is not. And the buildings themselves resist change. Old wiring, odd fittings, tenants in the room and local codes mean any machine has to handle exceptions all day. Employment is projected to grow about 4.2% between 2025 and 2035 (BLS projections), against roughly 1.53 million workers in the occupation (BLS, 2025).
What to do: get fluent with the diagnostic and work-order software your employer uses, because that is the part of the job changing first.
How to stay needed
Lean into the tasks that stay with people. Fault finding on equipment nobody documented. Repairs that involve water, power or structure, where a mistake is expensive. Safety judgment on ladders, in crawl spaces and around live systems, including knowing when to stop and call a licensed trade.
Two skills raise your floor. One is a license or certification in a regulated area, such as HVAC refrigerant handling or electrical work, because sign-off needs a named person. The other is reading and challenging machine output: when predictive maintenance flags a part, being the one who can confirm or reject the call is worth more than following it.
If you are weighing a move, look at nearby trades. Industrial machinery mechanics work on heavier plant with more documentation. Maintenance workers, machinery sits close to this role on the equipment side. Heating, air conditioning, and refrigeration mechanics and installers adds licensing and higher pay. You can see the wider group on the installation, maintenance and repair family page, the industry view under other services, and put two of these roles next to each other with the job comparison tool. Our guide to humanoid robots and physical jobs covers the hardware question in more detail, and the list of jobs that mostly need a person shows where this role sits among them.