Why this work stays split between a screen and the yard
Ask whether AI will replace petroleum pump system operators and you have to look at two different days of work. One part happens at a console: watching unit readings, answering alarms, lining up pumps and valves so product moves through pipelines and tanks on schedule. Software already does a lot of that math, and modern control systems hold more of it every year.
The other part happens outside. Gauging a storage tank, pulling a sample for the lab, checking a seal or a strainer, smelling and hearing a pump that is not running right. Those tasks need a body in a hard hat, in weather, near flammable product. A camera can see a leak. It cannot climb the stairs, open the hatch, and decide the batch is off spec.
There is also the accountability layer. Custody transfer figures, start-up and shutdown steps, and emergency response all carry legal and safety weight. Someone signs for them. That is why control systems keep getting smarter while the headcount in the plant falls slowly rather than all at once. The US Bureau of Labor Statistics counted about 33,160 people in this occupation with median pay of $96,710 (BLS, 2025), and projects employment to change by about -1.9% from 2025 to 2035 (BLS, 2025). That is erosion, not disappearance.
What software runs, what it assists, and what people keep
Start with the automated end. Our task split puts 5% of task time in work AI can handle today. The clearest examples are routine data work: logging gauge and meter readings, calculating flow and volume, and sequencing a transfer that follows a known recipe. Control systems have done versions of this for decades; machine learning mostly makes the alarms sharper.
Then the assisted middle, which holds 35% of the time. Here the operator stays in charge while software suggests. Unit optimization is one case: a model recommends setpoints to hold yield while an operator judges whether the unit will take it. Leak and equipment monitoring is another, where pattern detection flags a pump or a line early and a person goes to look. You can see how we score that in our coverage method.
The rest, 60% of task time, is still people. Manual gauging and sampling. Hands-on inspection and minor field maintenance. Emergency shutdown calls, where the right move depends on what you can smell, hear, and know about that unit’s history. That is the part robots would have to take, and the physical share of this job’s tasks falls into the mobile-robot tier rather than anything bolted to a bench.
What the evidence actually shows
Not much, yet, and that matters. This occupation carries an evidence grade of D on our scale, which means no study has tested an AI system against a qualified refinery operator on this job’s real tasks. So we publish no quality score for it. We would rather say that plainly than put a number on a guess.
What would settle it? A published trial where a control system runs a live unit through upset conditions without an operator intervening, with results reported against a human baseline. Or audited field data on automated custody transfer and tank gauging, including the exceptions. Vendor claims about advanced process control are not that, because they measure savings, not substitution. Our grades and what each one means sit on the methodology page.
Good to know: a low coverage figure with an unmeasured evidence grade means the work has barely been tested, not that it has been proven hard.
When it could change
Most likely after 2046 (8 in 10 of our scenarios). For what that window is measuring, see the replacement-year method.
Two things could pull that earlier. The first is remote and consolidated control rooms, where one center runs several sites and fewer operators cover more units. The second is instrumented tanks and lines: more sensors, automatic sampling, and metering that takes gauging rounds off the schedule.
Two things push it back. Safety regulation and process safety management put a person in the loop for start-ups, shutdowns, and abnormal operation, and rule changes move slowly. Capital is the other brake. Retrofitting an older refinery or tank farm with the hardware to remove field rounds costs far more than licensing software, and plants with thin margins defer it. Automating tasks is cheap; automating the yard is not.
How to stay needed
Lean into the tasks that only hold up with a person present. Abnormal-condition response, where you decide fast with partial information. Field verification and sampling, including the judgment call when a reading and the product do not agree. And unit start-up and shutdown, the sequences that go wrong in ways no model has seen.
Two skills raise your floor. One is instrumentation and control literacy: knowing why a loop is tuned the way it is, and when to trust the model’s setpoint. The other is process safety and incident command, the credential side of being the person who can authorize work and stop it.
If you want to see how nearby roles score, compare them with ours. Closest by work are Chemical Plant and System Operators and Gas Plant Operators, both console-and-field jobs in the same family. Upstream, Wellhead Pumpers shares the gauging and pump-tending side. You can also read the wider plant and system operators family, the mining, oil and gas sector, or our list of jobs expected to shrink to see where slow erosion shows up in official projections.
Put two jobs side by side if you are weighing a move between operator roles.