If ERP and MES Don't Talk, AI Won't Reach the Shop Floor

Many manufacturing AI pilots start with data but fail because of missing context.

A model can read a number. A factory does not run on numbers alone. It runs on work orders, routings, recipes, machines, shifts, operators, material status, quality results, maintenance history and customer delivery dates.

Without this context, AI may produce a prediction. But it will not produce action.

This is one of the most common reasons why AI pilots in manufacturing look promising in a demo but do not become part of daily operations.

The problem is not always the model. Often, the problem is the missing connection between ERP, MES and the real shop-floor process.

ERP knows the plan. MES knows the reality.

ERP systems usually know the business and planning side of manufacturing.

They contain customer orders, delivery dates, materials, purchasing, stock, finance, invoices and planning information.

MES systems are closer to what actually happens on the shop floor.

They capture when a work order started, which machine ran the operation, where a stoppage occurred, how many parts were produced, how many were rejected, and which quality result was recorded.

For AI, this distinction is critical.

ERP is close to the question: What should happen?
MES is close to the question: What actually happened?
AI becomes valuable when it can connect both and support the next question: What should we do now?

Why ISA-95 still matters in the AI era

ISA-95 may sound like a traditional industrial standard, but it is highly relevant for AI projects.

The ISA-95 framework defines the integration between enterprise systems and manufacturing control systems. It describes technology and business processes in layers and helps companies define interfaces between these layers. In the ISA-95 model, Level 3 includes manufacturing operations management systems such as MES, while Level 4 includes business planning and logistics systems such as ERP. The standard focuses heavily on the interface between these levels.

For AI projects, this becomes a practical architecture question:

Where does the AI system read data from?
Which level owns the decision?
Where is the AI output written?
Which system remains the source of truth?
Which actions require human approval?

If these questions are not answered early, the project is not ready to scale.

Data without manufacturing context creates weak recommendations.

A sensor value alone is rarely enough.

Imagine a machine temperature of 82 °C. Is that normal? It depends.

Which product was being processed? Which recipe was active? What was the normal range for this operation? Was the machine recently maintained? Did the scrap rate increase during the same shift? Is the work order linked to a critical customer delivery?

The same value can be normal for one product and risky for another. The same downtime category can indicate a maintenance problem on one machine and a material supply issue on another.

Without context, AI finds correlations. With context, AI can support operational decisions.

"We'll integrate it later" is expensive technical debt.

A fast prototype is useful. Not every early experiment needs full ERP-MES integration.

But in manufacturing, integration is not a technical detail that can always be added at the end. Integration is the backbone that connects AI output to operational reality.

If the model predicts a quality risk, which work order is affected?
If it suggests a maintenance priority, which system creates the task?
If it detects a delivery delay risk, which planner receives the alert?
If it classifies downtime, how does that classification map to existing MES reason codes?

These questions should be part of the pilot design, not a phase after the pilot.

McKinsey's work on digital manufacturing "pilot purgatory" describes the same scaling challenge: many manufacturers run pilots, but fewer move critical use cases into large-scale rollout. One of McKinsey's recommendations is to work "bottom-line-value backward" instead of technology-forward and to develop an early view of a scalable, analytics-enabled technology stack.

A better SME starting point: the minimum context model

SMEs do not need to replace every system before starting with AI.

A practical first step is to create a minimum context model for one decision.

For example, if the goal is to predict delay risk for critical customer orders, the AI system may only need a focused dataset that connects:

work order,
product or part number,
routing step,
machine or production line,
planned start and finish time,
actual start and finish time,
downtime reason,
produced quantity,
scrap or rework quantity,
quality result,
shift,
customer delivery date.

This does not require a huge data lake on day one.

It requires a clear decision, reliable data ownership and a simple integration design that connects ERP planning data with MES execution data.

Five AI use cases unlocked by ERP-MES context

Once ERP and MES context is connected, several practical AI use cases become much easier to implement.

Delivery delay prediction. AI can identify which work orders are likely to miss a customer delivery date based on current shop-floor execution.
Downtime classification. AI can help standardize downtime reasons and reduce unclear or inconsistent shop-floor entries.
Quality risk detection. AI can connect process data, operation history and quality results to highlight risky batches or parts.
Maintenance prioritization. AI can support maintenance teams by combining stoppage patterns, machine history and production urgency.
Planning support. AI can help planners understand which orders, machines or operations need attention first.

The common point is simple: each use case becomes valuable only when the AI output is connected to an operational action.

Read-only, assisted action or automated write-back?

One of the most important design decisions is the AI system's permission level.

Will AI only read data and create a dashboard?
Will it suggest actions for human approval?
Will it create maintenance tasks?
Will it update a status in the ERP system?
Will it write a downtime category into MES?

For many manufacturing SMEs, the best first step is not full automation. It is assisted action.

AI recommends. A human approves.
AI classifies. A human reviews exceptions.
AI detects risk. A planner decides the action.

This approach reduces risk, builds trust and creates a learning loop. It also makes the system easier to audit.

Deloitte's 2025 Smart Manufacturing and Operations Survey shows that manufacturers are investing in smart manufacturing foundations such as clean data analytics, sensors, cloud and AI, but also face challenges around operational risk, complex transformation and human capital.

Integration is trust architecture.

In manufacturing, trust is practical.

If operators do not trust the alert, they ignore it.
If planners do not trust the data, they keep using Excel.
If maintenance teams cannot see why AI suggested a task, they do not act on it.
If quality teams cannot trace the recommendation, they reject the system.

This is why ERP-MES-AI integration is not just about APIs.

It is about source of truth, traceability, responsibility and action.

A production AI system should answer:

Which data source was used?
When was the data updated?
Which system owns the status?
What did the AI recommend?
Who approved the action?
What was the result?

Without this traceability, AI will struggle to become part of production operations.

A 90-day pilot structure for manufacturing SMEs

A practical AI pilot can start small.

Weeks 1–2: select one decision and define the business KPI. For example: reduce late orders, reduce unclear downtime codes, improve first-time-right quality, or prioritize maintenance tasks.
Weeks 3–6: map the required ERP and MES fields and build the minimum context model.
Weeks 7–10: create the AI-supported workflow with human review.
Weeks 11–12: measure business impact and decide whether the use case should scale to another line, machine group, location or process.

The key is not to prove that AI can generate a prediction.

The key is to prove that AI can improve a real manufacturing decision.

Conclusion

If ERP and MES do not talk, AI will not reach the shop floor.

The model may still generate a dashboard, an alert or a prediction. But without business context, shop-floor context and system ownership, the output will not reliably become action.

For manufacturing SMEs, the best AI projects start with one operational decision, one measurable KPI and one integration design that connects the planned world of ERP with the actual world of MES.

At OzyCore, we design AI automation projects around this principle: start with the business decision, connect the right data context, keep human approval where needed, and scale only when the pilot proves measurable value.

Ready to map your ERP-MES-AI integration?

If your factory runs on disconnected ERP, MES and Excel realities, start with an AI readiness check. We help you pick the right operational decision, define the minimum context model and design a controlled 90-day pilot that connects planning and execution.

Sources

ISA, "ISA-95 Series of Standards: Enterprise-Control System Integration" — isa.org
MESA International, "AI & Machine Learning Hub" — mesa.org
Deloitte, "2025 Smart Manufacturing and Operations Survey" — deloitte.com
McKinsey, "How digital manufacturing can escape 'pilot purgatory'" — mckinsey.com