Compal cuts AOI false positives by over 97% with AI and rolls out on-premise multi-agent system to AcBel

Compal Electronics began its AI transformation in 2020, after QA data revealed that false-positive rates on more than 100 AOI machines had stayed high for a long time. Because of its high-mix production mode, product variation was large, and AOI parameters had to be tuned frequently and by hand from experience, which raised maintenance cost and held back line efficiency. Management then assigned the effort to the common software division, headed by a senior director. The biggest early resistance came from the production floor, where staff believed manual re-judgement had an error rate of only 0.02%. A proof of concept started with logo inspection, a lower-risk task with clear rules. The turning point was a "Data PK": AI and humans cross-checked the same samples under the same conditions, and the actual human misjudgement rate turned out to be 5%, about a hundred times the figure the floor believed. As models improved and data accumulated, Compal reduced the AOI false-positive rate by more than 97%, stabilising it at 0.5% to 1%, and raised assembly AOI test coverage by more than 30%. Compal has more than 100 production lines in Taiwan, China, Vietnam and the Americas. Its AI team now has more than 60 people, and AI has been introduced at more than 500 sites worldwide. Compal describes a hub-and-spoke structure with itself as the technical core, and it developed an in-house multi-agent collaboration platform. New production-line staff typically need three to six months of learning and practice, so that senior director decided to develop Agentic AI that turns senior staff's repair knowledge, fault-diagnosis logic and operating experience into an assistant supporting the floor. Compal then helped affiliate AcBel Polytech introduce the system, in which several AI agents handle equipment diagnosis, repair suggestions and part judgement. To avoid high cloud AI costs and the risk of sending data to the cloud, it fine-tunes small and mid-sized language models on domain knowledge so that they run on-premise. In 2025 the Digital Transformation Research Institute of the Institute for Information Industry supported a Compal and AcBel R&D project.