- AI integration in legacy factories is a connectivity problem first, an AI problem second — the path to intelligence runs through data unification, not model sophistication.
- The right approach is additive, not replacement: connect existing PLCs and SCADA to a data layer, then deploy AI on top. Rip-and-replace is expensive, disruptive, and rarely necessary.
- Most legacy factories can reach meaningful AI capability in 12-24 weeks without replacing any existing control infrastructure.
The starting point for 90% of manufacturers is legacy infrastructure, not a greenfield
Industry 4.0 content is dominated by greenfield factory case studies — new plants built from scratch with integrated IIoT, cloud connectivity, and AI from day one. This is the experience of a tiny minority of manufacturers.
The other 90% have production lines installed in the 1990s and 2000s, SCADA systems running Windows XP, PLCs communicating on proprietary protocols, and ERP systems that have never seen real-time shop floor data. AI integration in this environment doesn’t start with algorithms — it starts with connectivity.
Before scoping an AI integration project in a legacy factory, do a one-day data audit: can you answer these three questions from existing systems? (1) What did each machine produce in the last shift, with downtime reasons? (2) What quality results exist per production order? (3) What maintenance work was done on each asset in the last 12 months? If you can’t answer all three, close those gaps before deploying AI — the AI will only be as good as the data feeding it.
The four-layer integration approach for legacy factories
Layer 1 — Connectivity: Establish data paths from PLCs and legacy SCADA to a modern data layer. OPC-UA gateways, MQTT brokers, or direct historian connectors depending on what protocols your equipment supports. This doesn’t touch control logic.
Layer 2 — Unification: FactoLake collects data from all sources — PLC outputs, SCADA historians, manual entry, ERP feeds — and normalises it into a unified time-series and relational model. Data from a 1995 PLC and a 2020 IIoT sensor are queryable together.
Layer 3 — Context: FactoMES adds production context — linking machine signals to work orders, products, shifts, and operators. Raw machine data becomes production intelligence.
Layer 4 — AI: MaintIQ, ProdIQ, and QualIQ run on top of Layers 2 and 3, delivering predictions and recommendations from unified, contextualised data.
Sub-guides in this series
- How to Connect PLCs to AI
- Retrofitting Brownfield Plants with AI
- Industrial Edge Computing
- Using OPC-UA for AI Integration
- Common AI Integration Challenges
Frequently Asked Questions
At minimum: a connectivity path from at least your most critical equipment (OPC-UA gateway or PLC data logger), 6+ months of historical data in some form (even CSV exports from a historian), and production records (even from ERP) that can be linked to machine data by timestamp. This is enough to start a predictive maintenance pilot.
Almost never. The integration layer reads from existing PLCs via gateways or SCADA historians — no hardware replacement required. The exception: PLCs too old to communicate via any supported protocol (pre-1990 equipment sometimes) may require a hardware data logger at the machine.
Underestimating the data quality problem. Historical data that looks complete often has gaps, inconsistent timestamps, tag naming inconsistencies, or incorrect units — any of which degrade model performance. Budget for a data cleaning and validation phase before model training.