Time horizon
ERP / MRPWeeks and months. The plan is recalculated on a cycle, typically nightly or weekly.
AI decision layerMinutes and hours. The current shift is re-evaluated as machine state changes.
Integration guide
An ERP records what was planned and what was consumed. It was never designed to decide what to do when a machine slows down at ten in the morning. This page explains how an AI decision layer works with SAP, Oracle, Infor or Dynamics, what it reads, what it writes back, and what has to stay under human control.
Short answer
AI does not replace the ERP. It sits above it as a decision layer: it reads orders, materials and routings from the ERP, reads real machine state from PLC, SCADA and MES, works out what the plan should become, and writes the approved change back into the ERP. The ERP stays the system of record. The AI supplies the decision the ERP cannot make on its own.
ERP / MRPWeeks and months. The plan is recalculated on a cycle, typically nightly or weekly.
AI decision layerMinutes and hours. The current shift is re-evaluated as machine state changes.
ERP / MRPNominal capacity and standard times held in the master data.
AI decision layerMeasured cycle times, micro-stops and drift as they are happening.
ERP / MRPReports the deviation. Deciding the new sequence is left to people in a meeting.
AI decision layerProposes a revised sequence with the reasoning and the trade-off it accepted.
ERP / MRPHolds orders, stock, costing and the official record of what happened.
AI decision layerHolds no record of its own. Approved changes are written back so the ERP stays authoritative.
This describes how these systems are commonly configured as of 2026. ERP landscapes differ widely between plants, so treat it as a starting point for your own assessment rather than a verdict on any product.
Open orders, due dates, bills of materials, routings, stock levels and purchase lead times.
Machine state, real cycle times, alarms and quality events from PLC, SCADA and MES.
Checks whether the current plan is still feasible against real capacity, and builds the alternative when it is not.
Revised sequence, moved work orders and maintenance windows go back into the ERP and MES with a trace of why.
01
The first connection only reads. Nothing in the ERP changes while the model is being validated against your history.
02
Take the days the plan had to be rebuilt by hand and compare what the model would have proposed.
03
Go live in advisory mode: proposals reach the planner, who accepts or rejects them. Trust is built on real cases.
04
Only then enable write-back, scoped to specific order types and always behind an approval with a full audit trail.
A read-only connection to your ERP and one line of machine data is enough to show what the decision layer would have proposed on the days that went wrong.
Request a demo