You've hit on a really important question that many plant managers and engineers struggle with! The unspoken challenges of integrating AI predictive maintenance with legacy PLC systems go way beyond just technical compatibility. Here's what people don't always talk about:

1. The 'data starvation' problem: Legacy PLCs were designed to control processes, not share data. They often lack the memory, processing power, or communication protocols to provide the rich, granular data AI models need. You might get basic I/O status, but AI needs historical trends, sensor readings, and operational patterns that these systems were never built to capture.

2. Hidden infrastructure costs: Everyone talks about the AI software costs, but nobody mentions the 'middleware tax' - you'll need additional hardware like data historians, protocol converters, and edge computing devices just to extract and process the data before it even reaches your AI system.

3. The 'production paralysis' risk: Boeing lost $12 million on an AI quality control system because their 25-year-old databases couldn't feed data in real-time. The fear of disrupting production while trying to extract data from mission-critical systems is a huge unspoken barrier.

4. Skills gap realities: You need people who understand both legacy PLC programming AND modern AI/ML - a rare combination. Most AI developers don't understand ladder logic or proprietary PLC protocols, and most PLC programmers don't understand data science.

5. The 'good enough' trap: Legacy systems often provide just enough data to keep things running, making it hard to justify the ROI for AI integration when 'if it ain't broke, don't fix it' mentality prevails.

6. Security blind spots: Adding AI connectivity to legacy systems creates new attack vectors that these systems were never designed to handle, creating cybersecurity risks that are often underestimated.

The real challenge isn't just technical - it's convincing management that the ROI justifies the technical debt and disruption. Many companies end up using modular AI 'agents' or cloud-based solutions that connect through APIs as translators, rather than trying to overhaul everything at once.