The debate between AI vision and traditional machine vision is usually framed as a comparison of detection accuracy at the same task. That framing misses the more important point: the two architectures compete on different classes of tasks. Traditional machine vision with rule-based contrast-threshold detection handles routine inspection well — single product family, predictable defect classes, stable lighting, high contrast between defect and substrate. It handles complex inspection poorly — mixed materials, irregular defect morphologies, high product-mix lines, low-defect-rate products.
AI vision handles both, but the value proposition in the routine class is incremental. The value proposition in the complex class is categorical: the hardware-locked incumbents do not compete there. They assess the problem, determine that the architecture cannot reliably solve it at the customer's requirements, and decline the deployment. That is not a failure of any individual platform — it is the operating limit of the rule-based approach applied outside its design domain.
The Semiconductor Parts customer (Client B) is the cleanest illustration. A Korea plant of a Japanese customer operating precision small-component inspection. The hardware-locked vision incumbent evaluated the line and proposed a 3D vision rebuild at three times the 2D-solution cost — because the 2D architecture could not handle the irregular defect signatures on the small product geometry. A second AI vendor in the bid echoed the 3D recommendation. HyperQ AI Vision delivered the inspection on a 2D setup at roughly a third of the proposed 3D capital cost, with two days on-site for installation and commissioning. The competitive displacement was not an accuracy story. It was an architecture story: the learned model absorbed the variation that the rule-based architecture tried to solve through additional hardware.
What rule-based vision is excellent at
Traditional machine vision — contrast-threshold detection with structured lighting and a matched camera-lens configuration — is a mature, proven architecture. Decades of refinement. Large installed base. Reliable on-line performance in the operating envelope it was designed for.
That operating envelope is bounded by four conditions: a stable inspection geometry across the deployment's life, an opaque surface with high contrast between defect and substrate, a product-variant count low enough that per-variant recipes are feasible, and a defect-class distribution that is well characterised in advance. When all four conditions are met, the rule-based system is reliable and the AI vision alternative offers marginal improvement at meaningful additional cost. Stay with what works.
The failure modes appear at the boundary. Transparent or mixed-material surfaces put multiple optical layers in a single pixel, which contrast-threshold detection cannot disambiguate — we covered this in detail in the post on AI vision for glass and flat panel display manufacturing. High product-mix lines exhaust the per-variant recipe configuration budget, as Client A discovered at 8,000 variants. Irregular or atypical defect morphologies fall outside the threshold rules that were specified against the training set. Low-defect-rate products produce training sets too sparse for supervised classification.
What AI vision changes architecturally
The fundamental architectural change is the treatment of variation. Rule-based vision encodes the acceptable-variation boundary at configuration time, by hand, against a reference sample. The boundary is correct for what the engineer specified it against. It breaks when the production conditions move outside that reference — when the surface changes, when the defect morphology evolves, when a new product variant arrives.
AI vision encodes the acceptable-variation boundary by learning from the production-line distribution. The model trains against what the line actually produces — not against a vendor reference set. When the production conditions change, the model retrained against the new distribution generalises to the new conditions. The configuration burden moves from manual threshold-tuning at each product change to a retraining cycle on the customer's own production data.
The practical implication on a mixed-material line: the rule-based system needs two cameras and two lighting geometries to disambiguate metal and rubber surfaces at adequate contrast for both materials. The learned model handles the mixed-material surface variation inside the inference layer, with a single camera and structured illumination. Client A's reduction from two-camera-two-light to single-camera-single-light per station is the production proof of the architectural principle.
The two cases that mark the displacement boundary
Two customer cases define the operating boundary where rule-based vision stops competing.
The first is the Client A line at 8,000 variants and 30-plus changeovers per shift. Two successive hardware-locked vendors assessed the auto-switching requirement at that scale and declined to commit. Both platforms handled the lower-variant configurations adequately. At 8,000 variants with the changeover frequency the OEM scheduling required, the recipe-configuration architecture ran out of engineering hours. The displacement was not a question of detection accuracy. It was a question of operational sustainability at the variant scale.
The second is the Client C (Display Panel) line at one to two defects per year. The supervised-classification training requirement ran to years of accumulation time before a minimally viable dataset would exist. Multiple incumbent vendors assessed the line and declined. The anomaly-detection architecture trained on the line's good distribution — which the line produces continuously — and reached production-grade accuracy from the initial deployment. The displacement was not a detection-rate story. It was a data-architecture story: the incumbent's training requirement was structurally incompatible with the defect rate.
In both cases, the incumbent vendors made reasonable assessments that their architectures were not suited to the problem. The problems are not unusual in manufacturing. Mixed materials, high variant counts, and low defect rates are common in precision components, electronics, display manufacturing, and pharmaceutical packaging. The displaced platforms did not fail — they operated correctly within their design domain and correctly assessed that the problem was outside it.
The comparison framework that matters
For buyers evaluating the two approaches, the accuracy comparison on a single defect class is the wrong starting point. The right framework has five questions, which we covered in detail in the post on AI vision vendor evaluation and the 5 questions that expose hardware lock-in.
But the most direct version of the question is: what happens to the inspection capability when the production conditions that existed during the proof-of-concept change. New product variants introduced. New defect morphologies appearing on the line. Camera hardware upgraded. Those three scenarios occur on every production line within a three-to-five-year deployment horizon. The architecture whose capability degrades at each of those events is not the architecture for that line's operating life.
The hardware-locked vendor's accuracy in a controlled proof-of-concept is real and often competitive or superior on the specific defect class the POC was designed around. The question the POC does not answer is the three-year trajectory. AI vision's advantage in the complex class is not peak accuracy on a single defect class. It is the ability to improve on the customer's cadence, across the customer's evolving product range, without a vendor-managed migration at each evolution point.
Where the comparison is still open
For buyers operating simple, stable lines — single product family, predictable defect distribution, established lighting setup, low changeover frequency — the rule-based architecture remains a reasonable and often cheaper option. The AI vision value proposition on those lines is the potential for incremental accuracy improvement and the lower training-data requirement at product introduction; neither of those arguments is as compelling as the arguments on complex lines.
For buyers considering AI vision as a platform for multiple lines with different complexity profiles, the hardware-agnostic architecture scales across both the simple and complex cases — the same platform that handles the routine line also handles the 8,000-variant line. That platform scalability is the argument for standardising on AI vision even in cases where the individual simple line does not make the strongest case for the switch.