Most “AI in SMT assembly” projects fail for boring reasons: bad data joins, messy feeder metadata, and no change control. This post shows where machine learning actually improves pick-and-place programming, placement quality, and predictive maintenance—using real 2024 evidence and hard limits.