The challenge
Quality holds lived in a spreadsheet that three people edited by hand, and corrective actions were tracked over email threads that regularly went stale. PLC-driven inspection stations on the line already captured pass/fail data, but it sat in a proprietary log nobody exported, disconnected from the nonconformance record. Audit prep meant two engineers spending three weeks assembling evidence from five different folders. Supplier corrective actions closed late often enough that the same defect class kept recurring, and the quality director had no reliable way to prove otherwise to the plant's largest customer.
The approach
We built the nonconformance-to-CAPA workflow as the system of record, with the PLC inspection stations connected through a lightweight edge connector so a failed check opens a draft NCR automatically instead of waiting for a human to notice. A governed AI agent triages incoming supplier NCRs against defect history and drafts a first-pass severity and routing recommendation, which a quality engineer confirms or overrides — every recommendation is logged with the data it used. Every CAPA is bound to evidence at close, so an auditor can walk a single record from PLC event to signed closure without leaving the system.
The outcome
“The agent doesn't decide anything. It reads the same history our engineers would read and hands them a starting point with its reasoning attached — that's the only reason legal let it near the quality record.”