TL;DR — AI quality control is not one model, it is four layers — inspection, process monitoring, traceability and root-cause analysis. Most factories should automate inspection first, because it produces measurable results in one quarter: scrap rate down 20–40%, false calls cut by 80%+, and inspection labour down by half. Budget ¥150k–500k for the first line, and plan a 90-day cycle from kickoff to stable production.
What AI quality control actually covers
Vendors often use the term loosely. In practice there are four distinct layers, and they fail for different reasons:
- Inspection — detecting defects on the finished or in-process part. Fastest to deploy, clearest ROI.
- Process monitoring — predicting out-of-spec output from machine signals (pressure, torque, temperature) before the part is finished.
- Traceability — linking each part to its images, parameters and operator, so a complaint can be traced to a batch.
- Root-cause analysis — correlating defect patterns with process variables to find the actual driver.
Starting with root-cause analysis is the most common way projects stall: it depends on the other three layers already collecting clean data.
Which layer to automate first
Start where the defect is visible, the volume is high and the current method is human eyes. That combination delivers a defensible number within 60–90 days and builds internal credibility for harder layers. Process monitoring is the natural second step, because the inspection data it needs is already being collected.
The four metrics that prove it works
| Metric | Definition | Typical target |
|---|---|---|
| Defect recall | True defects caught / all true defects | ≥98% |
| False call rate | Good parts rejected / all good parts | ≤2% |
| First pass yield | Parts passing without rework | +3–8 points |
| Scrap cost per unit | Material and labour lost per part | −20–40% |
Track these weekly. A system that improves recall while tripling false calls has simply moved the cost, not removed it.
Cost bands
A single inspection station with camera, lighting, edge computing and software runs ¥150k–300k. A line-level system with MES integration and dashboards runs ¥500k–1.5M. Annual maintenance is typically 5–10% of the hardware value, mostly for optics cleaning, threshold re-validation and model updates.
90-day rollout
Days 1–14: collect 2,000–5,000 images across good parts and known defects, define the defect taxonomy with the quality team, agree acceptance thresholds in writing.
Days 15–45: train the first model, validate offline against a held-out set, then run shadow mode next to human inspection without acting on results.
Days 46–75: enable assisted mode — the system flags, the operator confirms. Log every disagreement; this is the most valuable data of the whole project.
Days 76–90: move confirmed categories to automatic reject, keep human review on ambiguous classes, and hand over dashboards to the quality team.
Common failure modes
Three problems account for most disappointing projects: acceptance criteria never written down, so “accuracy” is argued rather than measured; the imaging setup changed after validation; and nobody owns the model after handover. All three are organisational, not technical.
Frequently asked questions
Do we need to replace our existing vision system? Usually not. AI inspection is added as a second stage after rule-based pre-screening, which keeps cycle time short and cost contained.
How much data before go-live? 2,000–5,000 labelled images across all defect classes is a practical minimum for a single station.
Can one system cover multiple lines? Yes, if takt time allows shared inference. Multi-line deployment typically uses one central training pipeline and per-line edge devices.
What is the realistic payback? Eight to eighteen months on labour savings and scrap reduction alone, faster where customer complaints or recalls carry a high cost.
Next step
If you are scoping a first deployment, send us the line layout, defect list and target takt time. We return a feasibility note with expected recall, false-positive budget and a cost band within three working days. See also our AI visual inspection guide for architecture and pricing detail.
