EXPOSED
Almost nothing in this job is text or screen work — you drive a route, wrestle carts and bulk items, clear jams, and handle whatever residents actually put at the curb, which is why language AI barely touches it. The real threat is mechanical, not cognitive: automated side-loader arms and one-person cab operation have already eliminated the second and third crew members on many residential routes, and route-optimization software squeezes headcount further. CDL requirements and DOT/hazmat rules keep a licensed human in the seat for now, but that is a driving credential, not a judgment moat.
Tasks largely resist digitisation. Hooking a cart the arm missed, dragging a 90-gallon container off an icy verge, breaking down a mattress that won't fit the hopper, and back-blading a dead-end alley are physical improvisations no language model performs — automation here comes from hydraulics and route software, not from text prediction, which is why this sits at 16 rather than mid-band.
Hands-on in uncontrolled environments. Eight to ten hours in an unheated cab and on the step, 600-1,000 stops a shift, lifting bulky items in rain, sleet and August heat with traffic passing at 35 mph, plus climbing into the hopper area to clear jams — the only reason it isn't 20 is that ASL routes let some drivers stay seated for most residential collection.
Certification preferred, not legally required. A Class B CDL with air-brake endorsement, DOT medical card, and hazmat awareness training are hard legal prerequisites — no one puts you on that truck without them — but they license you to operate a vehicle, not to exercise protected professional judgment, so the shield stops well short of the 11+ band where a licensee personally answers for outcomes.
Anonymous artifact production. Residents notice you when service is missed, and a few know your name and leave water out in summer, but the customer relationship is with the hauler or the municipal contract; routes get reassigned quarterly and no household follows a particular collector, keeping this at the top of the anonymous-output band rather than into relationship territory.
Executes defined procedures on defined inputs. Decisions are calls like whether a container is overweight, whether a load is contaminated enough to tag and leave, and whether to back up in a cul-de-sac — real but bounded by company SOP, the acceptable-materials list, and DOT rules, with a dispatcher or supervisor deciding anything genuinely ambiguous.
Has AI actually changed your work?