AI AOI for Manufacturing Quality Control

Title

AI AOI for Manufacturing Quality Control

Content

Compal Electronics, one of Taiwan's largest contract electronics manufacturers, operates more than 100 production lines across Taiwan, China, Vietnam, and the Americas and started its AI transformation in 2020. The trigger was QA data showing that more than 100 AOI (Automated Optical Inspection) stations were producing persistently high false-positive rates: Compal's high-mix production forced the AOI parameters to be re-tuned by human experts on a constant basis, which drove up maintenance cost and slowed line efficiency. The breakthrough came from a "Data PK" exercise—a head-to-head comparison of AI detection versus human inspectors on the same sample set under identical conditions. The actual human error rate turned out to be 5%, roughly 100x higher than the 0.02% that line operators had believed. That single empirical result changed the organisation's view of AI from "suspect" to "needed." Compal cut the AOI false-positive rate by more than 97% (stabilised at 0.5%-1%) and raised assembly AOI test coverage by more than 30%. The internal AI team has grown to more than 60 people, and the system now runs on more than 500 sites globally. Compal uses a hub-and-spoke model: the central AI team develops the platform, then onboards sister company AcBel Polytech (康舒科技) as the first external deployment. The system is built on a multi-agent architecture for smart manufacturing, with specialised AI agents handling equipment diagnostics, repair recommendations, and parts identification. To control cost and keep proprietary process data off the public cloud, Compal runs the system on-premise on small-to-mid language models fine-tuned for domain knowledge. In 2025, the Institute for Information Industry's Digital Transformation Research Institute (資策會數轉院) supported a joint R&D project between Compal and AcBel Polytech to broaden AI deployment across the Kinpo Group, and the same playbook is now used to bring new operators from a 3-6 month learning curve to competence much faster.