Open Solder Joints: How AOI Detects Visible Defects
See how AOI identifies visible open solder joints, where fixed color thresholds struggle, and how a pretrained 2D model handles common and special cases.
September 17, 2026
No component library to build or maintain, and no XY data required: how golden-board AI programming actually works step by step, and how it compares to the traditional 3–5 hour AOI workflow.

Golden-board AOI programming works by having the system learn a board directly from one physical "golden" sample instead of from engineering data. A camera captures the known-good board, AI automatically locates components and generates inspection regions, and operator feedback refines the model over time. This removes the data import, component-library build, and manual threshold tuning that make traditional AOI setup take anywhere from a few hours to, for complex boards, several days.
For high-mix EMS shops, the way an AOI machine gets programmed often matters more than how it inspects. A line that changes over several times a day lives or dies by setup time. This article explains, step by step, how golden-board programming actually works, how it differs from the traditional workflow, and the questions engineers should ask before trusting it on the floor.
Traditional AOI programming is slow because it rebuilds a board's inspection logic from engineering data every time. A skilled engineer imports CAD/Gerber and placement data, verifies or extends a component library, teaches fiducials, defines board regions, sets per-component checks against IPC-A-610 criteria, and then tunes thresholds to suppress false calls. Published programming guides describe this as a multi-hour job for a straightforward board, stretching to days for complex, dense assemblies.
The traditional workflow generally runs in six stages:
Every stage assumes you have clean CAD data and time to tune. In high-mix, low-volume production — where CAD may be missing, incomplete, or from a customer who won't share it — stages 1 and 2 alone can stall a changeover.
Workflow steps drawn from published AOI programming guides (e.g., Aivon's step-by-step guide and Unisoft's programming overview), 2024–2025. Setup time varies widely with board complexity; treat any single figure as board-dependent.
Golden-board programming teaches the AOI system from a single physical known-good board — the "golden board" — instead of from engineering data and a component library. The system photographs this reference, and modern AI automatically detects components, extracts their visual features, and builds inspection regions by comparing production boards against the golden reference. The golden board is the specification, so there is no component library to build or maintain, and no XY data is required to start.

The "golden board" concept itself is not new — a known-good reference has long been part of AOI setup. (Most traditional AOI installed in the last decade actually builds its inspection model primarily from CAD/algorithmic data, using the reference board mainly for alignment and threshold optimization.) What changed with AI is what the software can do from that single board: older systems still needed a component library on top of it, whereas newer AI systems can generate a usable program from the golden board alone.
Golden-board AI programming replaces manual setup with automatic feature recognition. In practice it runs in four steps: scan one golden board; the AI auto-detects components and generates inspection frames; it sets initial thresholds from the reference image; and it learns from operator feedback on false calls, refining the model as production runs. Setup that took hours collapses to minutes because no data import, library build, or manual framing is involved.
This is the same broad direction several vendors are moving in — Saki, Marantz, and Mycronic all offer forms of auto-programming. The practical differences are in what inputs are still required: many auto-programming tools still expect XY data or a maintained component library, whereas a golden-board approach like DaoAI's has no component library to build or maintain and needs no XY data. If you're evaluating systems, that input requirement is the question to pin down.
The core difference is the input: traditional AOI programming is driven by placement data and a component library, while golden-board programming is driven by a single golden board plus AI feature recognition. This changes setup time from hours to minutes and shifts who can run a changeover from a dedicated programming engineer to a line operator.

| Dimension | Traditional AOI programming | Golden-board AI programming |
|---|---|---|
| Primary input | CAD / Gerber + placement data + component library | One physical golden board |
| Typical setup time | 3–5 hours per board (DaoAI-cited; varies with board complexity in independent guides) | ~5 minutes, operator-run (DaoAI-reported) |
| Who runs it | Dedicated AOI/programming engineer | Line operator |
| XY data availability | Required | Not required |
| New/custom components | Manual library entry | Auto-recognized from the board |
| False-call handling | Manual threshold tuning | Feedback learning over time |
| Best fit | Stable high-volume lines | High-mix, frequent-changeover lines |
For a deeper head-to-head, see AI AOI vs. traditional rule-based AOI: the hidden costs.
Yes — accuracy depends on the quality of the golden board and the AI's ability to recognize components, not on a hand-built component library. Library-driven programming leans on where parts should be and their specified values; a golden-board approach leans on what a good board actually looks like, which is what inspection ultimately compares against. Either way, acceptance criteria can be aligned to IPC-A-610, the industry standard for PCBA visual acceptability.
This matters for trust on the floor. IPC-A-610, "Acceptability of Electronic Assemblies" (current revision IPC-A-610J, released 2024), is the globally recognized benchmark for what counts as an acceptable solder joint, placement offset, or defect class. Its criteria can be translated into AOI rule sets regardless of how the program was built — so golden-board programming does not mean abandoning a recognized quality standard. DaoAI states its P Series achieves ≥98% detection accuracy, while false calls come down as operators give feedback; shops should validate performance on their own boards during a demo.
Golden-board programming mainly pays off in changeover time, labor, and false-call rework. Moving from a traditional 3–5 hour setup to a ~5-minute, operator-run golden-board workflow reduces programming time, while feedback learning brings false calls down as operators review results. A line operator can become inspection-ready in about 30 minutes instead of relying on a dedicated programming engineer.
| Metric | DaoAI-reported result |
|---|---|
| Programming-time reduction | up to ~97% (3–5 hours → ~5 minutes) |
| False-call reduction | Improves over time via feedback learning |
| Overall operating-cost reduction | ~60% |
| Labor savings | ~USD 15,000–40,000 per person / year (re-inspection + line-stoppage) |
| Operator ramp-up | inspection-ready in ~30 minutes; no AOI programming engineer |
These are vendor-reported figures, drawn from DaoAI's production deployments — including with a major Asian electronics manufacturer — where board programming that previously took 3–5 hours is completed by a line operator in about 5 minutes. For the mechanics behind those numbers, see how AI cuts PCBA AOI programming from 3 hours to 5 minutes. Any shop should confirm the numbers on its own boards during a demo — which, because setup takes minutes, is realistic to do live.
No. A golden-board approach starts from one physical known-good board. XY or BOM data can optionally improve labeling, but it is not required to generate a working inspection program.
Feature-recognition AI identifies components by their physical characteristics rather than matching a stored footprint, so genuinely new or custom parts are detected from the golden board without a manual library entry.
That is the intent of golden-board programming. Because there is no data import or manual framing, setup is reduced to scanning a golden board and reviewing flagged images — tasks an operator can perform. DaoAI reports operators are inspection-ready in about 30 minutes of training.
Yes. IPC-A-610 acceptance criteria define what an acceptable assembly looks like; they can be mapped into the inspection logic regardless of how the program was created.
Many auto-programming tools still require a maintained component library. The key question when comparing systems is which inputs are truly optional. A golden-board workflow has no component library to build or maintain and needs no XY data. You can review DaoAI's approach and full specs in the resource center.
DaoAI's P Series runs the golden-board workflow across offline (P1/P1L/P2), inline 2D (P3/P3D) and inline 3D (P5/P5D) configurations — the same golden-board programming for high-mix benches and high-volume lines alike.
AI inspection insights, product news, and industry analysis. Once a month, no noise.
We use your details to reply to your enquiry. See our Privacy Policy.