Evaluating Human-Executed and LLM-Based SMED Methods for Changeover optimalization in a Lathe line

This study investigates the application of Large Language Models (LLMs), specifically GPT-5, to optimize changeover processes using the Single Minute Exchange of Dies (SMED) methodology on a lathe line with a 14.5-hour baseline. Traditional human SMED, implemented by engineers, reduced changeover time by 44.4% (5 hours 6 minutes) through externalizing tool preparation and calibration, waste elimination, and ergonomic tool upgrades, yielding €6,200 annually with a €550 investment. GPT-5, analysing the same data, classified activities as 70.25% value-adding, 27.52% non-value-adding, and 3.92% downtime (vs. human 50%/17%/33%), achieving 89% deficiency alignment and proposing a 30–50% reduction with modular assemblies and AR, aligning ~80% with human waste strategies. The LLM operated on an anonymised textual representation of 170 recorded activities and was guided by a structured prompt to classify each activity and summarise deficiencies and improvement opportunities. The comparison underscores human SMED’s practical gains and GPT-5’s scalable insights, supporting a hybrid optimization approach.