Progression of AI Agents vs Human Experts: Analysis of Pick and Place Material Downtime Data for Proper Root Cause Analysis (RCA) and Guided Actions

Abstract

Material-related downtime on pick-and-place (PnP) equipment is prevalent and costly, yet difficult to diagnose because symptoms span feeders, reels, splicing/presentation, nozzle pick dynamics, component variability, and operator actions. Expertise to correlate these heterogeneous signals is increasingly scarce. This paper evaluates whether large language model (LLM) agents can deliver repeatable, serialized root-cause analysis (RCA) and operator-ready actions when paired with domain data structure.

We present a practical data pipeline that transforms raw PnP event streams into expert-processed evidence: time-aligned alarms, reconstructed feeder/reel/part lineage, and contextual metadata. Using this evidence, we compare two prompting strategies: naïve versus expert-guided.  We compare three data regimes: raw, processed, and processed+aggregated. Across multiple production days and experimental conditions, naïve prompting on raw streams yields generic advice lacking actionable specificity; expert prompting without processed evidence remains brittle. In contrast, the processed-data + expert-prompt configuration consistently produces precise RCA tied to serialized assets and clear corrective actions, achieving and exceeding human expert performance.

These results indicate a clear path to scalable AI-driven downtime reduction on modern P&P lines.