A hardware stamping factory in Zhongshan. Revenue: roughly 50 million yuan annually. 12 QA inspectors working three shifts. Manual visual inspection error rate: 2 to 3 percent. Annual quality-related losses: approximately 1.5 million yuan. The owner thought the problem was staffing.
He was wrong.
The Truth Hidden by Sunk Cost
Week one revealed two management blind spots nobody had noticed.
Blind spot one: the same workstation at 3 AM had nearly 5x the error rate at 3 PM. Physiological fatigue impact on quality control was far greater than management realized.
Blind spot two: 31 percent of defects caught were not QA problems at all — they were cumulative dimensional deviations from upstream die wear. After running a full month of data, the AI system auto-flagged: Die mold number 3 wear curve begins accelerating at day 17. Preemptive mold replacement dropped upstream defect rate from 12 percent to 4 percent.
AI Did not Replace People — It Showed Them What They Could not See
Three takeaways:
First: The first month AI did not save labor — it exposed two management blind spots nobody knew existed. The second layer of AI QA value is implicit and far more valuable: die wear early warning, shift fatigue curves, defect correlation maps. This is not QA — it is production system health diagnostics.
Second: Eighty percent of peers still think AI QA is only for big factories. A full system costs under 20K yuan now. This information gap is his biggest competitive moat.
Third (most valuable): He thought he bought an inspection tool. It turned out to be a CT scanner for his entire production line.
Three Rules for Peer Factories
- Start with QA, not full-factory AI. QA is the highest-tolerance, fastest-ROI entry point.
- Choose open-interface systems, not black boxes. You need upstream signals, not just inspection results.
- Run a full month of data before calculating ROI. The first month's biggest value is not how many people you saved but how many problems you discovered that you did not know you had.