TL;DR — AI-based AOI keeps the existing camera and motion platform and replaces the rule library with a learned model. On SMT lines the gain is concentrated in false calls: typical reduction is 60–85%, which directly cuts re-inspection labour and stops operators from disabling the machine. Real defect detection improves more modestly, typically 2–6 percentage points. Retrofit cost is ¥80k–250k per machine, and payback usually lands in 6–12 months on rework and inspection labour.
Where traditional AOI breaks down
Rule-based AOI compares each pad, lead and component against a golden board using colour, area and offset thresholds. Two things break it: normal process variation (solder volume, paste spread, component tolerance) and cosmetic differences that are not defects. Engineers respond by loosening thresholds until the machine stops complaining, which trades false calls for missed defects — a silent quality risk.
What the AI layer changes
The model learns from labelled images what a real defect looks like on your boards, including the variation that rules cannot express. Practically, that means three things on the line:
- Fewer false calls on solder joints, tombstones, shifted components and slight misalignment.
- Stable behaviour across board variants, because the model generalises instead of matching a golden board.
- Explainable rejects: each call comes with a heat map, so operators can confirm or override in one click — and every override becomes training data.
Defect classes it handles well
| Defect | AI performance vs rules |
|---|---|
| Solder bridge / insufficient solder | Clear improvement, fewer false calls |
| Tombstone, billboard, shifted part | Clear improvement |
| Missing / wrong component | Comparable; rules already strong |
| Polarity and orientation | Improvement, especially on dense boards |
| Cosmetic marks, flux residue | Large improvement (rules over-flag) |
| Cold joint under shield | Limited; needs X-ray, not AOI |
Be explicit about the last row with any vendor. AOI cannot see what optics cannot see, and no model changes that.
Integration with MES and rework
The AI layer should push results to MES with board serial, station, defect code and confidence. Rework stations then pull the image and heat map, which shortens repair time and produces a clean record for traceability. Without this loop, false-call reduction is real but the labour savings are harder to prove.
Cost and rollout
Retrofitting an existing AOI machine with an AI inspection layer runs ¥80k–250k per machine depending on camera resolution and board complexity. A new AI-native AOI platform runs ¥400k–1.2M. Rollout takes four to eight weeks: two to three weeks of image collection across board variants, one to two weeks of labelling with the process engineer, then shadow mode before enabling automatic reject.
What to measure
Track false calls per thousand joints, escaped defects at functional test, rework minutes per board and operator override rate. A retrofit that halves false calls but doubles overrides has not solved the operator-trust problem, which is the actual goal.
Frequently asked questions
Can we keep our current AOI machines? In most cases yes. The AI layer runs on an added edge PC and consumes the existing camera feed.
How many boards do we need to train on? 200–400 boards covering all variants and defect classes is a practical starting set.
Will it work on high-mix low-volume? Yes, but plan for per-family models. A single model across unrelated product families usually underperforms.
Does it replace X-ray or ICT? No. It reduces false calls at the AOI stage; electrical and hidden-joint verification still needs X-ray or in-circuit test.
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.
