Some factories and refineries now let software change their equipment settings directly, and others let it only suggest. At Saudi Aramco's Yanbu refinery, optimization algorithms adjust production in a closed loop, with operators watching. At its Ras Tanura refinery, the algorithms recommend and operators decide. The software company behind both, AVEVA, expects most AI in plants to stay in the suggesting role for a good while, and to act alone only on narrow, repetitive jobs.
Software has controlled plants before, so the change is in how far it reaches
Software steering a plant is not new. Model predictive control, which holds a process steady while respecting set limits, has run in refineries and chemical plants since the 1980s. What a living digital twin adds is reach. It is a model of the plant that stays in step with live sensor readings, so an agent can act with a view of how the whole plant is behaving, not only one unit's loop. That widens the set of decisions that can be handed over, and so the number of times a person must decide whether to hand one over.
The control room of the 1970s shows where the strain moves
The operators of that era lost something along with the work. Handling the valves and dials gave them a feel for the plant. A senior operator who has never seen the plant shut down in an emergency becomes the person least ready when it finally does, and the usual remedy was practice on simulators and regular time on the plant floor.
AI repeats this at a larger scale. When an agent handles most adjustments, the human sees fewer of them, and the ones they do see are the odd cases the agent could not handle. The supervisor needs more judgment, not less, and gets less everyday practice using it.
What it looks like outside a plant
Take a warehouse shift manager who now approves every reorder of packing tape, pallets and labels that the system proposes. At about 150 a day and 20 seconds each, that is roughly 50 minutes a day of clicking yes, or four hours a week.
Now let an agent reorder anything under a set amount from approved suppliers. The manager writes the limits once: the amount, the suppliers, the stock levels at which it may act. The agent sends back only the five or six cases outside them. That gives back about four hours a week, but only if those five or six cases are worth reading. If the amount is set too low, the manager is back to 150 approvals. If it is set too high, a wrong order goes through and nobody sees it until the invoice arrives.
The same arithmetic applies to a finance director approving small supplier invoices, or a clinic manager approving shift swaps. In each case the agent takes over the work that was easy to judge, and the human is left with the hard cases and with the limit itself.
Before letting any agent act alone, write its limit in one sentence that a colleague could check. Then estimate how many cases it will send back each day. If that is more than one person can read carefully, the limit is wrong, whatever the agent is able to do.