A production-grade industrial AI deployment spans five layers: physical process and PLC at the bottom, SCADA for supervision, MES for production management, edge computing for local inference, and cloud or enterprise systems for training, analytics and generative AI. Understanding where each AI model runs, how data flows between layers, and where the boundaries between deterministic control and probabilistic AI sit is essential for any architect or technology leader building AI into a manufacturing environment.
This architecture reference maps the standard layers of an industrial AI deployment — from PLC and SCADA through edge and cloud to generative AI — and explains where each type of AI model should run, how data flows between layers, and what the integration points with existing plant systems look like. It is written for architects, CTOs and project leaders who need to design or evaluate a production-grade industrial AI stack.
Industrial AI does not replace existing plant systems — it extends them. The PLC controls the physical process deterministically. SCADA supervises and logs. MES manages production orders, recipes and traceability. Edge computing provides local compute for real-time AI inference. Cloud and enterprise systems handle model training, large-scale analytics and generative AI workloads. Each layer has a distinct role, and the architecture must define what crosses the boundary between them.
The key design principle is that deterministic control stays on the PLC and SCADA layers. AI adds intelligence — predictions, optimizations, anomaly detection — but it does so through supervised interfaces, not by replacing safety interlocks or deterministic logic. Generative AI, in particular, operates at the application layer: it helps humans interpret data and make decisions, but it does not write directly to physical equipment.
For guidance on moving AI through these layers from pilot to production, see our pilot-to-production guide. For planning the overall initiative, see our industrial AI roadmap.
| Layer | Role | AI models that run here |
|---|---|---|
| 1. PLC and physical process | Deterministic control of motors, valves, conveyors and sensors. Safety interlocks and real-time logic. | No AI models run directly on the PLC. AI recommendations reach the PLC through a supervised write interface, never bypassing safety logic. |
| 2. SCADA | Supervision, alarming, real-time visualization and data logging from PLCs across a line or cell. | Lightweight anomaly detection models can run alongside SCADA for real-time alerts. SCADA data is a primary input for AI models at higher layers. |
| 3. MES | Production order management, recipe execution, traceability, quality records and OEE calculation. | MES provides production context (batch ID, recipe, work order) that enriches AI predictions. Predictions and recommendations can be written back to MES for operator visibility. |
| 4. Edge computing | Local compute appliances or industrial PCs placed near the line. Low-latency inference without cloud dependency. | Computer vision models for defect detection, predictive maintenance inference, process optimization models. Edge is where most production AI inference happens. |
| 5. Cloud and enterprise | Model training, large-scale analytics, data lakes, generative AI services and integration with ERP. | Model training and retraining, digital twin simulations, generative AI for operator assistance, shift report summarization and enterprise-wide analytics. |
Cloud inference introduces network latency and a dependency on connectivity that a plant cannot afford for real-time decisions. Any inference that affects safety or line speed must run on edge, not in the cloud.
AI models are probabilistic. Safety interlocks on PLCs are deterministic. Never allow AI to write directly to a PLC output that bypasses the safety logic. AI recommendations must pass through the existing control hierarchy.
Large language models can hallucinate and produce non-deterministic output. Generative AI belongs at the application layer — assisting operators, summarizing data, generating procedures — not in the control loop of physical equipment.
The Purdue model exists for a reason. Connecting AI systems directly from enterprise cloud to PLC without proper segmentation, firewalls and DMZs creates security vulnerabilities that can compromise the entire plant.
Some AI teams build a separate data pipeline that duplicates sensor readings, creating a second source of truth that drifts from SCADA and MES. AI must consume data from the existing systems of record, not from a shadow pipeline.
Edge devices have limited compute, memory and thermal envelopes. A model that runs on a cloud GPU may not fit on an industrial PC in a cabinet at 50 °C. Validate model size and inference time on the actual edge hardware before deployment.
It depends on latency, safety and connectivity. Real-time control and safety-critical inference belong on edge devices close to the PLC. Batch analytics, model training and generative AI tasks that tolerate seconds of latency can run in the cloud. Most production deployments use both, with edge for inference and cloud for training and heavy processing.
Generative AI is not a replacement for SCADA or MES. It sits at the application layer, where it can assist operators with natural-language queries about plant status, generate maintenance procedures from documentation, or summarize shift reports. It does not control physical equipment directly.
MES is the system of record for production orders, traceability and quality data. AI models consume MES data for context — batch numbers, recipes, work orders — and can write predictions back to MES so that operators see AI output in the interface they already use.
Direct AI-to-PLC writes are possible but should be approached with caution. Safety interlocks and deterministic control logic must remain on the PLC. AI recommendations should pass through a supervised layer or a human-in-the-loop approval before affecting physical equipment, unless the use case has been fully validated for autonomous control.
The Purdue Enterprise Reference Architecture defines network levels from Level 0 (physical process) to Level 4 (enterprise). Industrial AI deployments must respect these levels: edge inference at Level 1–2, MES integration at Level 3, and cloud analytics at Level 4. Crossing levels without proper network segmentation creates security risks.
Related resources: our consulting service includes architecture advisory for manufacturers designing their AI stack, and our pilot-to-production guide covers how to move AI through these layers in practice. For sector-specific architecture considerations, see our energy and chemical industry pages.
Tell us the plant environment, the use case and the existing systems. We propose architects who have designed production-grade AI deployments with MES, SCADA, edge and generative AI.