The plant had 30 years of batch records, a modern DCS, and a persistent yield gap between its best and average operators. The knowledge existed; it was not transferable.
The approach was not to replace operator judgement with a model. It was to make the plant legible: instrument the variables that mattered, model the relationship between trajectory and yield, and give operators a live view of where the current batch sat relative to the best historical trajectories.
Operators used the view to intervene earlier. Within two quarters, first-pass yield improved 4.6 points and batch-to-batch variance narrowed by a third. The model did not make the decisions; it made the consequences of decisions visible while there was still time to change them.
The critical design decision was ownership. The tool was built with the shift teams, launched as a recommendation rather than an instruction, and tuned weekly against their feedback. Attempts elsewhere in the group to launch the same capability as a centrally mandated optimiser stalled within months.
Industrial AI succeeds when it augments the judgement of the people closest to the process, and fails when it tries to relocate that judgement to a dashboard nobody trusts.
