Causal AI explains quality holds. PdM schedules maintenance. Agentic tools raise tickets. Soft sensors estimate process values. MES orchestrates orders. What still breaks delivery promises is the schedule: infinite-capacity plans that ignore changeovers, tool calendars, and real bottlenecks until the night run is already wrong.
AI-assisted finite-capacity scheduling is landing on top of APS and MES—re-solving feasible sequences as constraints move, not dumping a prettier Gantt that planners discard by 09:00.
The industrial point is a start time the plant can execute. An optimized plan that violates a furnace campaign is fiction.
What finite-capacity AI changes
- Live constraint packs — Tools, crews, materials, and changeovers in the solve.
- Bottleneck-first sequencing — Protect the scarce resource, not every idle minute.
- Reschedule loops — Partial replan when a hold or breakdown lands mid-shift.
What still fails
Black-box optimizers that cannot explain why an order moved. Systems divorced from MES genealogy and real setup matrices. This is not causal scrap RCA, not PdM work-order generation, not agentic shop-floor bots, and not soft-sensor APC loops.
What to watch next
- Sites that cut late orders without exploding changeover minutes.
- Whether planners trust and edit AI proposals inside the same system of record.
- Integration depth with MES/APS—not a side spreadsheet that fights the floor.
MES knows the order. Finite-capacity scheduling decides when it is honestly possible.