EXPOSED
The modal ag tech splits time between field plots/greenhouses/labs — laying out trial blocks, tagging plants, pulling soil cores, running tissue and seed assays, calibrating instruments — and a screen tier of data entry, spreadsheet cleanup, and routine report writing that AI plus sensor networks eats first. Sampling and instrument handling in muddy, variable field conditions still needs hands and legs, but no license protects the role and the scientist above signs off on the science. Drone imagery, in-line soil sensors, and automated plate readers shrink the headcount needed per acre of trial rather than eliminating the job.
Mixed — a routine tier and a judgment tier. Hand-thinning a variety trial, tagging individual plants for phenotyping, and pulling soil cores on a grid still need a person walking the plot, but the yield-data transcription, plate-reader output cleanup, and standard-format trial reports are already machine work — that split is what puts it at 11 rather than 15.
Hands-on in uncontrolled environments. You are in mud, greenhouses at 95°F, and grain bins on the same week, hauling augers and sample bags across uneven ground and handling live plant material and pesticide-treated plots — uncontrolled outdoor work, though the lab bench and instrument room half of the job keeps it off the 18-plus ladder-and-livestock tier.
No licence, no signature requirement. No state licenses agricultural technicians; a pesticide applicator card or a two-year ag degree may be listed as preferred, and even the GLP-relevant signature on a trial record belongs to the study director above you, so the 3 reflects credentials that help you get hired but shield nothing.
Anonymous artifact production. Your soil nitrate numbers and germination counts are read as data, not as your data — the seed company or extension scientist consuming them can swap technicians between seasons without any loss, and the 5 rather than 0 only covers the working familiarity with cooperating growers whose fields host the plots.
Meaningful discretion. Trials come with a written protocol specifying plot layout, sampling depth, and replication, so your discretion is real but bounded — deciding a plot is compromised by flooding and flagging it, or calling an instrument out of calibration mid-run, are the calls you own, and the protocol deviation still goes upstairs for approval.
Has AI actually changed your work?