Great question! Having worked with AI-PLC integration for predictive maintenance, I've seen both genuine ROI successes and some 'AI-washing' disappointments. Let me share what actually works versus what's mostly hype.

First, the ROI-positive scenarios I've implemented:

1. Vibration analysis on critical motors and pumps - By integrating AI algorithms with PLC vibration sensors, we've achieved 40-60% reduction in unplanned downtime. The key was focusing on high-impact equipment where failures cost thousands per hour in production losses.

2. Temperature trend analysis for hydraulic systems - Simple AI models analyzing PLC temperature data predicted seal failures 2-3 weeks in advance, reducing hydraulic system downtime by 70% in automotive plants.

3. Motor current signature analysis - This delivered solid ROI by detecting bearing wear and misalignment early, preventing catastrophic failures in conveyor systems and pumps.

Now, the 'AI-washing' scenarios that often disappoint:

1. Overly complex 'black box' solutions - When vendors promise AI that predicts everything but can't explain why, it's usually AI-washing. These systems generate alerts but no actionable insights.

2. Predictive maintenance for low-impact equipment - Applying AI to equipment where failures are cheap to fix often costs more than the problems it solves.

3. 'One-size-fits-all' AI models - Generic solutions that don't account for specific PLC data structures or plant conditions rarely deliver promised ROI.

The key difference? Successful implementations start with clear business problems, use existing PLC data effectively, and focus on equipment where failures have real financial impact. The AI-washing happens when technology drives the solution rather than business needs.