An unplanned stoppage on an automotive assembly line costs far more than the failed part. It costs the idle production crew, components cooling on the line, customers rescheduling deliveries, and the maintenance manager explaining to the plant director why nobody saw the failure coming. According to a Siemens analysis published in 2024, unplanned downtime accounts for estimated losses of $1.4 trillion per year across the world’s 500 largest industrial companies (Siemens, 2024).
Predictive maintenance addresses this with a simple logic: replace the calendar with data. Rather than changing a bearing every 3,000 hours, you change it when vibration sensors detect real-world degradation — three to six weeks before failure. The difference: zero surprise stoppages, zero premature replacements.
In 2026, AI and IoT make this accessible to mid-market manufacturers, provided the chain is complete: sensors, CMMS (Computerised Maintenance Management System), and ERP all connected. This guide covers architecture, maturity levels, KPIs, and measurable ROI for any CIO or maintenance manager ready to act.
Why Move from Preventive to Predictive Maintenance?
The Hidden Cost of Unplanned Stoppages
Maintenance accounts for 15 to 40% of production costs in manufacturing. That figure covers very different realities depending on the maintenance regime in place.
In pure corrective maintenance (fix it when it breaks), the visible cost is the emergency repair. The hidden cost is the downtime: in discrete manufacturing (assembly, metalworking, plastics), industry estimates place the cost of one hour of unplanned downtime at between €5,000 and €22,000 depending on line complexity. In continuous process industries (chemicals, food and beverage, energy), that figure rises to €50,000–€150,000 per hour — higher still if the stoppage triggers batch losses or regulatory requalification.
The spare part rarely represents more than 10% of the total incident cost. The rest: emergency labour, lost production, customer penalties, and external expertise. McKinsey & Company has documented that reactive repairs cost on average four to five times more than planned repairs on the same equipment (McKinsey, Predictive Maintenance).
Corrective, Preventive and Predictive: The Three Regimes
Corrective maintenance: intervene after the failure. Low short-term cost, high risk, heavy production impact. Only acceptable for non-critical assets with short replacement lead times.
Calendar-based preventive maintenance: intervene at fixed intervals (every X hours or X weeks) regardless of actual equipment condition. You replace components that are still healthy (waste) or miss accelerated degradation between visits. This is the dominant regime in industry in 2026: it reduces failures but does not eliminate them.
Predictive (or condition-based) maintenance: intervene when data indicates the equipment will fail within a predictable window. Sensors continuously measure actual condition (vibration, temperature, motor current, oil analysis). AI detects anomaly signatures and predicts Remaining Useful Life (RUL). The intervention is planned in advance, at the right moment, with the right part available.
According to McKinsey, predictive maintenance reduces maintenance costs by 18–25% and unplanned downtime by 30–50% compared with a purely reactive strategy. Deloitte confirms similar gains, with equipment reliability improvements of 30–50% (Deloitte Insights).
The Technical Building Blocks of a Predictive Maintenance System Integrated with ERP
IoT Sensors and Edge Computing: The Data Foundation
Predictive maintenance rests on physical signals measured continuously. The most widely used sensing technologies in industry:
- Vibration analysis: accelerometer sensors on bearings, bushings and drive shafts. The vibration signature of a degrading bearing shifts weeks before failure. This is the most mature and widespread technique.
- Infrared thermography: thermal cameras or surface-temperature sensors detecting abnormal hot spots on motors, electrical connections, and heat exchangers.
- Oil analysis: periodic sampling or inline sensors measuring metallic particle concentration. Indicates gear and bearing wear.
- Motor Current Signature Analysis (MCSA): analysis of current harmonics to detect mechanical imbalance or winding defects.
- Ultrasound: detection of gas leaks, partial electrical discharges, and lubrication defects.
These raw data streams cannot flow directly into the ERP. An edge gateway installed on the shop floor collects readings, filters and aggregates them, and transmits only significant indicators to the ERP or IoT platform: trends, threshold alerts, anomaly scores. OPC-UA and MQTT are the industrial standards for this transit. For a detailed walkthrough of this architecture, see our IIoT and ERP integration guide: OPC-UA, MQTT and edge computing.
The Role of the CMMS in the Chain
The CMMS is the central repository for maintenance. It records the asset register, intervention history, preventive maintenance routines, and maintenance labour and spare parts resources.
In a predictive maintenance architecture, the CMMS plays two critical roles:
- Alert receiver: it receives alerts generated by sensors and the AI layer, maps them to a specific asset in its register, and creates the corresponding work order (WO).
- Historical data provider: AI models train on the historical failure records stored in the CMMS. Without a well-maintained history (failure type, root cause, failed component, intervention duration), predictive models cannot be trained reliably.
The quality of data labelling in the CMMS is a prerequisite that is often overlooked. A history that records “motor failure” without specifying the failed component or root cause is unusable for model training. This is a critical checkpoint before launching any predictive maintenance project.
ERP as Orchestrator: Work Orders, Spare Parts, Costs
The ERP sits downstream of the CMMS, handling the business dimensions of maintenance:
- Spare parts inventory management: when the CMMS creates a preventive WO, the ERP checks component availability in stock, triggers replenishment if needed, and reserves the part for the planned intervention date.
- Labour planning: the ERP integrates maintenance resources into the workload schedule, accounting for required skills, certifications, and availability.
- Cost accounting: each intervention is posted to a cost centre, an asset, or an internal order in the ERP. The actual maintenance cost per asset becomes visible in management accounts.
- Maintenance budget control: the ERP consolidates forecast and actual costs, compares against budget, and flags assets whose maintenance cost exceeds the profitability threshold.
The AI/ML Layer: Anomaly Models and Remaining Useful Life
Two families of algorithms dominate industrial predictive maintenance:
Anomaly detection: a model is trained on “normal” equipment data (vibration, temperature, current under nominal operation). The model learns to recognise healthy behaviour. When real-time data deviates significantly from this profile, an alert is raised. Most commonly used algorithms: Isolation Forest, autoencoders (neural networks), and LSTM (Long Short-Term Memory) for time series.
Remaining Useful Life (RUL) prediction: the model does not merely detect an anomaly — it predicts how long the equipment can continue operating before failure. This prediction is what enables advance planning of the intervention. Random Forest, gradient boosting, and LSTM algorithms are used depending on historical data availability and the complexity of the degradation process.
These models must be trained on data from your plant, on your equipment, under your operating conditions. A generic off-the-shelf model will perform poorly without a fine-tuning phase on your data.
3 Integration Architectures Based on Your Maturity Level
Pattern 1: ERP with Native CMMS Module
This is the simplest architecture to deploy and the most functionally coherent. The ERP manages maintenance natively, with no third-party system.
SAP S/4HANA Plant Maintenance (PM) is the reference for large industrial enterprises. It covers asset management, maintenance orders, preventive routines, and intervention history. The SAP Asset Performance Management (APM) add-on, connected to SAP IoT, adds the predictive layer: real-time sensor data, anomaly detection, and automatic maintenance order creation. Owens Corning documented savings of $2 million per year per site using this SAP architecture, with a 30–50% reduction in unplanned downtime (SAP Community, 2024).
IFS Cloud Asset Management is a serious competitor, particularly well regarded in energy, transportation, and heavy manufacturing. IFS ships a native MQTT broker that receives sensor data directly, maps it to assets in IFS, and automatically creates work orders with no additional development. This native IoT strength is a significant competitive advantage over SAP.
Microsoft Dynamics 365 Field Service is a strong option for mid-market companies already in the Microsoft ecosystem. It integrates with Azure IoT Hub for sensor data ingestion and Azure AI for predictive models. The advantage: full coherence with the Microsoft 365 and Power BI environment already in place in many organisations.
Pattern 2: ERP + Integrated Third-Party CMMS
When the incumbent ERP lacks a sufficiently deep maintenance module, a specialist CMMS sits between the sensors and the ERP. ERP-CMMS integration is handled via REST API or EDI connector.
IBM Maximo Application Suite is the CMMS market standard for asset-intensive industries (energy, petrochemicals, utilities). Its 8.x release includes a native AI layer (IBM Watson) for predictive maintenance and connects to virtually every major ERP on the market.
Hexagon EAM (formerly Infor EAM) is a cloud CMMS designed for complex industrial environments, with strong coverage of condition-based and predictive maintenance.
Fiix by Rockwell Automation is a modern, cloud-native CMMS increasingly adopted by mid-market manufacturers in the UK and North America, with connectors for SAP, Oracle, and Microsoft Dynamics.
In this pattern, the flow is: sensors → IoT platform → CMMS (alert + WO) → ERP (spare parts + accounting + planning). Each boundary is an interface to develop and maintain.
Pattern 3: Dedicated IoT Platform + API to ERP
For companies whose sensor estate is already deployed on a proprietary or sector-specific IoT platform, the approach is different: the IoT platform remains central, and the ERP is called via API for operational actions.
Azure IoT Hub (Microsoft) and AWS IoT Core (Amazon) are the most widely used general-purpose cloud platforms. They ingest sensor data, execute rules or ML models, and call the ERP via webhooks or dedicated connectors.
Siemens MindSphere is Siemens’ industrial IoT platform, native for Siemens equipment but open to third-party assets. It includes pre-packaged predictive maintenance applications and SAP connectors.
This pattern suits manufacturers with multiple sites or multiple ERPs: the IoT platform is the centralisation point, distributing alerts to each site’s ERP via a standardised API.
Implementation in 5 Steps: From POC to Full Deployment
Step 1: Identify Critical Assets and Instrument Them
Do not start with the ERP. Start with a criticality analysis: which assets have the greatest impact if they fail (degraded OEE, line stoppage, safety shutdown)? For each of them, what are the known failure modes and which physical parameters precede them?
This analysis drives sensor selection and instrumentation strategy. An asset with three distinct failure modes may require three different sensor types.
Step 2: Connect Sensors to Your ERP/CMMS
Deploy the edge gateway, configure the protocols (OPC-UA from PLCs, MQTT from wireless sensors), and wire alerts to the CMMS. Do not underestimate this step: in existing plants, OT and IT networks are often physically separate, PLCs run on proprietary protocols, and some assets have no digital output at all.
The IEC 62443 standard governs OT/IT separation and the security of flows between the two networks. It must be respected from the architecture design stage.
Step 3: Train AI Models on Historical Failure Data
This is the most underestimated step. Anomaly detection models need large volumes of “normal” data and small but precisely labelled volumes of “degradation” data. If the CMMS does not contain well-documented failure history spanning at least two to three years, models cannot be trained reliably.
Auditing CMMS data quality before launching the project is essential. If historical data is insufficient, a six- to twelve-month data collection and enrichment phase must precede predictive model deployment.
Step 4: Configure Alerts and Automatic Work Orders
The end-to-end pipeline must be tested and validated: sensor detects anomaly → anomaly score crosses threshold → alert sent to CMMS → work order created with the asset, intervention type, and required spare part → ERP checks part availability → if out of stock, purchase order triggered → WO scheduled in the maintenance planning calendar.
This end-to-end test in a live environment (not in simulation) is the condition for project success. It surfaces latency issues, data quality problems, and system integration gaps.
Step 5: Measure ROI and Refine Alert Thresholds
A predictive model is not deployed and forgotten. Alert thresholds must be adjusted regularly to minimise false positives (alerts with no actual failure, which demotivate maintenance teams) and false negatives (undetected failures). This ongoing calibration work is often neglected in projects, with the consequence that organisations revert to calendar-based preventive maintenance after twelve months.
KPIs to Track and Expected ROI
MTBF, MTTR, OEE: The Three Pillar Indicators
MTBF (Mean Time Between Failures): average time between two failures on an asset. Predictive maintenance aims to extend it by anticipating failures. Doubling MTBF on a critical asset means a production line that stops once a year instead of twice.
MTTR (Mean Time To Repair): average time to return an asset to service after a failure. Predictive maintenance reduces MTTR by planning the intervention in advance: the part is ordered and on hand, the technician is trained, and the downtime slot is chosen to minimise impact.
OEE (Overall Equipment Effectiveness): a composite indicator combining availability, performance, and quality. An OEE of 75% means the asset only generates productive value 75% of the theoretical available time. Each OEE point gained on a critical line translates directly into additional production capacity.
ROI by Sector: Documented Ranges
Cross-sector experience consistently points to several ranges:
- Reduction in unplanned downtime: 30–50% according to McKinsey, up to 70–90% at high-maturity deployments according to Deloitte.
- Reduction in maintenance costs: 18–25% according to McKinsey, up to 40% according to Deloitte.
- Extension of asset service life: 20–40% according to McKinsey.
- Payback period: 27% of manufacturers deploying predictive maintenance reach payback in under 12 months. The median payback falls between 12 and 24 months (Oxmaint Global Report 2025).
These figures are not uniform: they depend on the maturity of historical data, the criticality of instrumented assets, and the quality of CMMS-ERP integration.
Which ERP/CMMS for Predictive Maintenance in 2026?
| Solution | Type | Predictive Maintenance Strengths | Target |
|---|---|---|---|
| SAP S/4HANA + SAP APM | Native ERP | Deep functionality, SAP IoT connectors, AI agents H2 2026 | Mid-market / Enterprise |
| IFS Cloud Asset Management | Native ERP | Native MQTT broker, automatic work orders, strong asset coverage | Heavy industry mid-market |
| Microsoft Dynamics 365 + Azure IoT | ERP + Cloud | Microsoft coherence, integrated Power BI, Azure ML | Microsoft ecosystem mid-market |
| IBM Maximo Application Suite | Third-party CMMS | High maturity, Watson AI, multi-ERP | Large industrial enterprises |
| Fiix by Rockwell Automation | Third-party CMMS | Modern UX, cloud-native, SAP/Oracle/Dynamics connectors | UK & North America mid-market |
| Odoo Manufacturing + OCA modules | Generalist ERP | Low cost, flexibility | SMEs with limited budget |
For mid-market manufacturers starting their predictive maintenance journey, the recommended path is pragmatic: begin by instrumenting three to five critical assets with vibration sensors, connect them to a CMMS (or a native ERP module), and measure ROI before scaling. A three-month POC on a target process typically costs €15,000–€40,000, including sensor infrastructure and CMMS-ERP integration. Budget depends heavily on the number of assets instrumented and the condition of the existing OT network.
To go deeper on the technical architecture, read our ERP and CMMS integration guide for industrial maintenance and our manufacturing ERP, MES and IoT Industry 4.0 guide.
