Over the first quarter we ran observation-only pilots in three Tier-2 stamping plants in Selangor and Johor. Different MES stacks (Wonderware, FactoryTalk, a homegrown C# system that should not exist), different shift patterns, different OEE definitions. The pattern that repeated was uncanny.
Pattern one: Monday is the bottleneck
In all three plants, the Monday morning OEE meeting was the single largest engineering time sink of the week. Engineers spent Sunday night or Monday early-shift exporting MES data into spreadsheets, building pivot tables, and arguing over which deviation came first. By the time a corrective action shipped, three more shifts had run with the same loss.
“We don't run OEE. We document it. Every Monday I have a meeting where we describe what already happened to people who can no longer change it.”
Pattern two: the data was always there
Not once did we find a plant that needed more sensors, more historians, more dashboards. The MES knew. The historian knew. The downtime codes were tagged. The data was sitting in the same systems that the engineers were exporting on Sunday night.
The gap was not capture. The gap was attention. No human can watch fourteen lines, eight signals per line, in real time, and notice when station 12's torque drifts six standard deviations off baseline three minutes before it stalls the cell.
Pattern three: trust is earned per shift
- Week one: engineers ignored every agent recommendation, watched it from a distance.
- Week three: engineers checked the agent's reasoning when something they'd already noticed showed up in it.
- Week six: engineers were forwarding agent recommendations to maintenance without re-checking.
- Week ten: engineers asked us to add a second use case.
Earned autonomy is not a slogan. It is what happens when an agent says 'here is the cycle-time drift on station 12, here is what changed in the last 14 days, here is what I would do', and is right enough times that the engineer stops re-doing the math.