Manufacturing Execution Systems (MES) have been around for decades. They track production, manage workflows, collect data from the shop floor. Most manufacturers I work with have some form of MES, even if it’s a patchwork of different systems.

AI needs data. MES has data. Seems like a natural fit.

But connecting the two isn’t always straightforward. Here’s what I’ve learned from implementations across Australian manufacturing.

Why MES-AI integration matters

Your MES is probably your richest source of production data. It knows what was made, when, by whom, on which equipment. It might capture quality measurements, cycle times, material consumption.

AI applications—predictive maintenance, quality prediction, process optimisation—need this data to function. Without MES integration, you’re either working from incomplete data or manually exporting and importing constantly.

The goal is getting MES data flowing to AI systems automatically, reliably, and in real-time.

The integration landscape

Standard interfaces

Modern MES platforms generally support industry standards:

OPC-UA: The dominant standard for industrial data exchange. Most MES systems can expose data via OPC-UA, and most AI platforms can consume it.

REST APIs: Web-style interfaces for accessing data. Increasingly common and often easier to work with than older protocols.

Database connections: Direct queries against the MES database. Powerful but can create performance issues if not managed carefully.

File-based exports: The simplest approach—regular exports to CSV or XML. Limited but sometimes sufficient for less time-sensitive applications.

Legacy system challenges

Older MES systems—and there are plenty still running in Australian factories—don’t always support modern interfaces.

Options include:

Middleware: Products like KEPServerEX or Ignition can bridge older protocols to modern standards.

Custom integration: Building specific connectors for your situation.

Database replication: Copying MES data to a separate database that AI systems can query without affecting production.

Vendor upgrades: Some MES vendors offer modernisation paths that add API capabilities.

Real-time vs batch

Some AI applications need real-time data (predictive maintenance monitoring current conditions). Others work fine with batch updates (monthly production analysis for planning).

Real-time integration is more complex and expensive. Make sure you actually need it before paying for it.

Common integration patterns

Pattern 1: Data lake approach

MES data flows to a central data lake (cloud storage or on-premise data warehouse). AI applications read from the lake.

Advantages: Clean separation between operational systems and analytics. Multiple AI applications can share data. Historical data readily available.

Disadvantages: Not truly real-time. Adds infrastructure. Data quality issues can accumulate.

This works well for analytics and planning applications where latency of minutes or hours is acceptable.

Pattern 2: Event streaming

MES publishes events (production starts, quality measurements, equipment alerts) to a message queue. AI applications subscribe to relevant events.

Advantages: Near real-time. Scalable. Multiple consumers possible.

Disadvantages: More complex architecture. Requires reliable messaging infrastructure.

Good for applications that need to respond quickly, like anomaly detection or real-time quality monitoring.

Pattern 3: Direct integration

AI application connects directly to MES APIs or database.

Advantages: Simple. Minimal infrastructure.

Disadvantages: Tight coupling between systems. Can impact MES performance. Each AI application needs its own integration.

Suitable for single-purpose AI applications in smaller operations.

Data quality issues

Getting data out of MES is one challenge. Making it useful for AI is another.

Missing data: Sensors fail. Operators skip entries. Equipment offline. AI systems need to handle gaps gracefully.

Inconsistent data: The same measurement might be recorded differently across shifts, machines, or time periods. Normalisation is essential.

Contextual information: Raw production data often needs context. What product was running? What recipe? Which operators? This contextual data is crucial for meaningful AI analysis.

Timestamps: Clock synchronisation matters when correlating data from multiple sources. I’ve seen analyses fail because equipment clocks were minutes apart.

One chemical manufacturer I worked with spent three months cleaning and normalising their MES data before AI applications could use it effectively. Frustrating, but necessary.

Security considerations

MES systems are operational technology. They’re on your factory network, connected to equipment. Opening them up for AI integration creates security implications.

Network segmentation: Don’t expose MES directly to cloud AI services. Use DMZ architecture or secure gateways.

Read-only access: AI applications typically need to read data, not write it. Enforce this at the integration layer.

Authentication and encryption: Standard IT security practices apply.

Monitoring: Track integration connections. Unusual patterns might indicate problems.

Practical implementation steps

Step 1: Inventory your data

Before integrating anything, understand what data your MES actually contains.

• What production events are captured?
• What quality measurements exist?
• What contextual information is available?
• How far back does historical data go?
• What’s the data quality like?

This assessment often reveals gaps that need addressing before AI can be useful.

Step 2: Define AI requirements

What data does your intended AI application actually need?

Be specific. “Production data” is too vague. “Cycle times for product X on machine Y with associated temperature readings” is actionable.

Work backwards from the AI application’s requirements to identify exactly what needs to flow from MES.

Step 3: Design the integration

Based on data requirements and system capabilities, choose an integration pattern.

Consider:

• Latency requirements (real-time vs batch)
• Number of AI applications that will use the data
• Technical capabilities and infrastructure
• Budget and timeline

Step 4: Build and test

Start with a subset of data. Prove the integration works before expanding scope.

Test data quality, not just connectivity. Does the data that arrives at the AI system match what’s in MES? Are there transformation errors?

Step 5: Operate and maintain

Integrations need ongoing attention:

• Monitoring for failures
• Adjusting to MES changes
• Scaling as data volumes grow
• Addressing data quality issues as they arise

Vendor considerations

If you’re purchasing AI solutions, ask vendors about integration capabilities.

Pre-built connectors: Do they support your MES platform specifically?

Standards support: OPC-UA, REST, database connectivity?

Implementation support: Will they help with integration, or is that on you?

Data requirements: What specific data do they need? How granular?

The answers reveal a lot about how mature the vendor’s industrial AI capabilities really are.

For complex integrations, bringing in specialists helps. AI consultants Melbourne and similar firms have done this before and can avoid common pitfalls.

A realistic timeline

MES-AI integration isn’t a weekend project. For a typical mid-size manufacturer:

• Assessment and planning: 4-8 weeks
• Infrastructure and architecture: 4-6 weeks
• Integration development: 6-12 weeks
• Testing and validation: 4-8 weeks
• Deployment and stabilisation: 4-6 weeks

Total: roughly 6-9 months for a comprehensive integration. Individual AI applications can pilot faster with limited data, then expand.

The payoff

Done well, MES-AI integration creates a foundation for multiple AI applications.

The building materials manufacturer I mentioned in a previous article—they invested significantly in integration infrastructure. But that investment now supports predictive maintenance, quality prediction, and energy optimisation from a shared data foundation.

The first AI application carried the integration cost. Subsequent applications got a head start.

If you’re planning multiple AI initiatives, invest in integration infrastructure early. It pays dividends across projects.

Working with AI consultants Brisbane or other experienced partners can help design integration architecture that scales, rather than building point-to-point connections that become technical debt.

Looking ahead

Integration between operational systems and AI will only become more important. As AI capabilities mature, the bottleneck increasingly becomes data access, not algorithms.

Manufacturers who invest in solid integration foundations now will be able to adopt new AI applications faster than those who keep patching together ad-hoc connections.

It’s infrastructure work—not glamorous, but essential.