Machine state
Real cycle times, micro-stops and drift from PLC and SCADA, not the nominal capacity written in the master data years ago.
Practical guide
AI for production planning: what it changes on a real schedule
Most plants do not lack a plan. They lack a way to rebuild the plan when a machine slows down, a material arrives late or an order is pulled forward. This page explains how AI-based production planning works, where classic APS stops, and what a plant needs before it can trust an automated replan.
Short answer
AI production planning is not a better Gantt chart. It is a system that keeps reading machine state, order book, materials and maintenance at the same time, notices when the current plan has stopped being feasible, and proposes a revised sequence with the reason behind it. The plan only changes when a person signs it off, which is what makes it usable on a running line.
The four inputs a planning model needs
Order book
Due dates, priorities and changes coming from ERP. A plan that ignores a pulled-forward order is wrong the moment it is printed.
Materials
Stock, inbound deliveries and work in progress. Most infeasible schedules are not a machine problem, they are a material that is not there yet.
Maintenance
Planned stops and the early signs of an unplanned one. Planning and maintenance decided separately is how a line ends up stopped in the middle of a rush order.
Where classic APS stops and AI planning begins
Input data
Classic APSMaster data and nominal capacities, refreshed on a cycle.
AI planning layerLive machine signals alongside orders, stock and maintenance.
When it runs
Classic APSOn a schedule, or when a planner launches it by hand.
AI planning layerContinuously, and it raises its hand when the plan stops being feasible.
Output
Classic APSA new schedule to interpret and apply.
AI planning layerA proposed change with the evidence behind it and the impact on due dates.
Control
Classic APSThe planner owns every step, which is why replans are rare.
AI planning layerThe system prepares the replan, the plant signs it off. Every change keeps a trace of why.
Descriptions reflect how these approaches are commonly documented and deployed as of 2026. Every plant is different, so treat this as a starting point for your own evaluation rather than a verdict.
How a replan actually happens
01
It notices
A cycle time drifts, a delivery slips, an order moves. The model compares the live picture against the plan currently in force.
02
It explains
Before proposing anything it shows the signal, the baseline and the orders at risk, so the people on the floor can challenge it.
03
It proposes
A revised sequence with the trade-off stated: which orders stay on time, which slip, what it costs in changeovers.
04
It acts with your sign-off
Once approved, the change is written back into the systems you already run. Nothing moves in the plant without that approval.
Common questions
- What is AI production planning?
- It is a planning layer that reads machine state, orders, materials and maintenance together, continuously checks whether the current schedule is still feasible, and proposes a revised sequence with the reasoning behind it. Unlike a static schedule, it reacts to what the plant is actually doing rather than to what the master data says it should do.
- Does AI planning replace our APS or ERP?
- No. Atherya sits above SAP, Oracle, Infor, Dynamics, MES and SCADA, reads from them and writes approved changes back. The ERP stays the source of record, the planning layer supplies the decision.
- Do we need new sensors for AI scheduling?
- Usually not. If your PLCs, SCADA, MES and ERP already produce the signals, Atherya starts from those. Hardware retrofits add cost and months before the first result, so they should be a last resort, not a starting point.
- Can the system change the schedule on its own?
- It can prepare and execute a replan, but every replan waits for your sign-off first, and every change keeps a trace of why it was made. Autonomy is only useful when it is reversible and auditable.
- How long before we see a result?
- When no new hardware is needed, the limiting factor is data access. A typical start is one line and one month of history, replaying the days when the plan had to be rebuilt, to show what the model would have proposed.
Related guides
Try it on one line
Bring one line, one month of history and the reschedules that cost you the most. We show what the model would have proposed on those days, before you commit to anything.
Request a demo