You've hit on a really important question that doesn't get enough attention! From what I've found, vendor presentations often gloss over these gritty realities. Here are the hidden challenges nobody wants to talk about:

1. The 20-year-old PLC speaking Modbus problem - AI models might be great, but they can't communicate with legacy PLCs using ancient protocols like Modbus. One engineer shared that their most expensive AI pilot failed not because the AI was bad, but because it couldn't handle a 20-year-old PLC speaking Modbus.

2. The 'last mile' integration nightmare - Integration is never "just an API" as vendors claim. The real work happens in translating modern AI inference outputs into legacy protocols without breaking existing production lines.

3. Safety hazards from AI hallucinations - When AI vision systems make mistakes (hallucinate), it's not just an error - it's a safety hazard in industrial environments. This trust gap is the hardest bridge to cross.

4. Real-time latency issues - Legacy PLCs weren't designed for the data volumes and speed requirements of modern AI vision systems. The communication protocols just can't keep up.

5. Hidden middleware costs - You'll need "wrappers" or edge gateways to translate between systems, adding complexity and cost that's rarely mentioned upfront.

The successful projects seem to be the ones that treat AI as a 'human-in-the-loop' advisor first, rather than trying to hand over full autonomous control immediately. It's all about respecting the existing physics and processes before adding AI on top.