The alarm fired at 2 a.m. Vibration on bearing 4 had crossed threshold. The dashboard turned amber, then the on-call engineer’s phone buzzed, then a decision that should have taken thirty seconds took three hours, because the alert told him what, not what to do about it.
This is where most predictive maintenance programs quietly stall. Plants spent a decade wiring up sensors and cloud dashboards, and the payoff was supposed to be foresight. Instead, many control rooms got alarm fatigue: more warnings, the same amount of human judgment required to turn a warning into an action.
The gap isn’t prediction. It’s the last mile.
Predicting a failure and preventing one are different problems. A digital twin that only displays a remaining-useful-life curve is still asking a person to diagnose the cause, simulate the fix, check it against safety limits, and dispatch the work order under time pressure, at 2 a.m., from a dashboard.
The shift underway now is closing that last mile. Instead of a passive mirror of the asset, the twin becomes an active participant: it runs the what-if scenarios itself, checks each option against the physics of the equipment, and hands the operator a recommendation instead of a symptom. Maintain current output and bearing temperature breaches its metallurgical limit in 48 hours. Throttle the drive 15% and you buy three weeks until the scheduled turnaround. That’s a decision an operator can approve in seconds, not reconstruct from raw telemetry.
The market is already pricing this in. The global Asset Performance Management sector - the category that includes this closed-loop capability is projected to grow from $2.40 billion in 2026 to $4.32 billion by 2032, a 10.3% CAGR. That’s not spend on better alerts. It’s spend on systems that act on them.
Autonomy without abdication
The natural objection is trust: nobody wants a plant that changes its own setpoints unsupervised. The honest answer is that almost none of this runs fully unsupervised, and it shouldn’t at most sites.
Three governance models coexist in practice. Human-in-the-loop, where the twin drafts the intervention and a certified engineer signs off before anything moves. Human-on-the-loop, where the twin acts and the operator has a window to pause or reverse it. And full closed-loop autonomy, reserved for a narrow set of unstaffed or remote assets.
Shell’s Groningen gas field is the reference case for that last tier: 29 production clusters and more than 900 square kilometers of pipeline network, monitored by two control-room operators. It didn’t get there by removing human judgment; it got there by proving, in shadow mode, that the system’s recommendations were reliable enough to earn delegated authority. Trust was built one verified intervention at a time, not declared on day one.
Verification is what makes it repeatable
The step people underestimate is the fourth one: after the twin acts - a setpoint change, a work order - it watches the sensors to confirm the asset actually returned to its predicted healthy state. If it did, the intervention is logged as verified. If it didn’t, the model recalibrates itself rather than the plant absorbing a silent miss.
That closing step is what turns a one-off save into a system that gets more trustworthy over time, and it’s the difference between “the AI made a recommendation” and “the AI made a recommendation, we know it worked, and it will make a better one next time.”
None of this requires abandoning the engineer at 2 a.m. It requires giving them a system that has already done the diagnosis, run the trade-offs, and checked its own homework against the physics, so the human decision left standing is the one that actually needs a human: approve, or don’t.
Where is your plant on that path? Is it still triaging alerts, or already approving twin-generated interventions?
Ferfier helps plants close this last mile by building digital twins that simulate the intervention, check it against equipment physics, and verify the outcome before the next recommendation gets trusted further. If you want to see where your own plant sits on the path from alarm to approved intervention, talk to us.
