Manufacturing SMEs in Ghent, Antwerp, and across East Flanders have built legitimate operational discipline. ISO compliance is high. Lean adoption beats European averages. ERP penetration at the 10-50 employee level often exceeds what you'd find at firms three times that size. So when a production director says their operation is "data-rich," they're not lying—they're just describing a different problem than what AI tools actually need to function.

The Operational Discipline Trap

The trap is structural, not cultural. Flemish manufacturers collect operational data optimized for human operators and compliance audits: shift summaries, batch-level defect logs, machine state records every 30 seconds via SCADA. This looks like plenty of raw material for an AI quality control system or predictive maintenance model. It isn't. The demo vendor shows a compelling tool that detects surface defects in real time. The contract gets signed. Six months later the project is stalled because the AI needs clean, labelled, time-synchronized data tied to production batch metadata—and what the firm actually has is image archives on a NAS with inconsistent naming conventions, SCADA exports in proprietary formats nobody budgeted middleware for, and ERP records that don't share timestamp references with machine logs. The data exists. It is simply not AI-consumable.

Four Structural Gaps That Kill Manufacturing AI Projects

Based on patterns observed across industrial SMEs in the Benelux region, Dr Hernani Costa identifies four gaps that recur consistently: Timestamp fragmentation means ERP systems, SCADA platforms, quality management software, and operator logbooks run on independent time references. Merging these for a predictive maintenance model requires a unified event timeline—and even a 30-second offset between machine state logs and quality inspection records can corrupt a training dataset. Label scarcity is the second gap: defect records exist in QMS but are logged by shift or batch, not tied to specific machine states or upstream process parameters. Creating retrospective labels requires domain expertise that often lives in senior operators' heads rather than any system. Proprietary SCADA lock-in is the third barrier—many Flemish SMEs run SCADA systems from vendors who provide no open data export APIs, forcing reliance on OPC-UA connectors or middleware layers that don't appear in AI vendor proposals. The fourth gap is ERP data completeness: fields critical for AI like machine assignment per production order, operator ID, ambient conditions, and tooling state are often optional inputs operators skip or fill inconsistently.

What a Real Readiness Assessment Actually Looks Like

A credible AI readiness assessment for a Flemish manufacturing SME is not a maturity questionnaire. It's a technical audit against specific use case requirements that answers five questions: Can your data be extracted in structured formats on scheduled or real-time bases? Can it be joined—meaning timestamp alignment across systems confirms records from different sources can be reliably merged? Is it labelled for your target use case, with defect records tied to machine states rather than just shift summaries? Do you have enough history—typically 6-18 months of labelled production data for quality control models? And critically, who owns the data pipeline in production? An AI system requiring a dedicated data engineer is not operationally viable for a 25-person manufacturer. The assessment must determine whether proposed architecture can be maintained by existing staff.

The Right Sequencing for Ghent and Antwerp Firms

For most Flemish manufacturing SMEs, the correct AI adoption sequence is not "find an use case, procure a tool, implement." It starts with a data infrastructure audit against two or three highest-priority pain points—typically quality control yield, unplanned downtime, or supply chain lead time variability. Phase two addresses high-impact gaps: timestamp synchronization layers, structured defect labelling processes, confirmed SCADA data export. This unglamorous phase typically takes 8-12 weeks and is where most projects either succeed or quietly die. Only then does a constrained pilot make sense—single production line or product family with clearly defined success metrics like defect detection rate on a specific component. Firms often discover that the infrastructure work in phase two already delivers operational improvements before any AI model deploys, because cleaning and structuring data forces process discipline that was previously absent.

Key Takeaways

  • Having mature ERP systems does not equal AI readiness—operational data collection and AI-consumable pipelines are fundamentally different problems
  • Timestamp fragmentation is often the first red flag: most firms discover their systems don't share a common time reference when they try to join datasets
  • Proprietary SCADA lock-in creates hidden middleware costs that don't appear in AI vendor proposals
  • Quality control automation and predictive maintenance show strongest ROI at the 10-50 employee firm size
  • The correct sequence is audit first, infrastructure work second, constrained pilot third—never the other way around

The Bottom Line

This gap isn't a technology problem. It's an honesty problem. Vendors won't tell you about it until they're engaged because honest conversations kill deals. The manufacturers that will lead their sectors over the next three years aren't those with the highest AI ambition—they're the ones doing the unglamorous data pipeline work right now, building labelled datasets competitors haven't bothered to create.