AI quality inspection: when can smaller batches become economical?

AI quality inspection can improve small-batch manufacturing economics when it reduces the complete cost of qualifying and maintaining an inspection task across product variants. Lower model-training effort alone is insufficient. Imaging, fixtures, false alarms, missed defects and acceptance responsibilities must leave enough contribution for the additional orders to remain commercially worthwhile.

Inspection setup costs across product variants

A short production run spreads inspection preparation across relatively few units. If commissioning costs €12,000, an illustrative 2,000-unit programme carries €6 per unit before operating inspection costs; a 200,000-unit programme carries six cents. Reducing preparation therefore matters most where other production activities are already flexible. It does not resolve expensive tooling, minimum material orders or customer qualification elsewhere. The relevant opportunity is a family of products for which inspection has become a disproportionate constraint on commercially useful variety.

Bosch’s account of stator inspection describes creating synthetic fault images to prepare a model for a new application. It also describes engineers correcting alerts on acceptable welds. The reported shortening of project duration and financial gains were expectations, not disclosed realised returns. The case shows how preparation can become easier while engineering judgement remains necessary. A manufactured image can represent a known fault; it cannot establish the incidence of faults in ordinary production or the full range of acceptable variation.

Different approaches address different problems. Rule-based vision can remain economical where the acceptance condition is simple and stable. Supervised recognition learns from labelled examples, whereas methods such as PatchCore investigate anomaly detection using normal examples. Detecting an unusual appearance is not equivalent to establishing that the part fails its specification. A supplier should demonstrate the relevant function on representative production conditions, rather than offer a generic manufacturing accuracy score.

False alarms, missed defects and inspection workload

False rejections can dominate the economics when defects are rare. In an illustrative 100,000-part population with a 0.1% defect rate, 99% detection of defective parts finds 99 defects and misses one. Incorrectly flagging 1% of the 99,900 acceptable parts adds 999 false alerts. Only about 9% of the resulting 1,098 alerts concern actual defects. One minute of review per alert requires 18.3 hours, before handling and records. These are assumed rates and calculated consequences, not supplier measurements. Automatically scrapping all alerts would replace review cost with lost material and capacity.

Visibility is a separate physical constraint. A camera cannot reveal a fault hidden from its viewpoint simply because the model improves. Additional views, illumination, handling or another measurement technique may be necessary. Their cost can exceed software preparation. Product families sharing geometry and acceptance requirements offer more reuse than an assortment requiring new fixtures each time. The economically important design choice can therefore occur before software selection, when the company decides which variants can share a dependable inspection arrangement.

Qualification must address the defects that matter and the consequences of missing them. A few successful examples do not establish a low escape rate for rare events. Repeated images of the same specimen provide less independent evidence than genuinely varied faults. A model may be useful for routing questionable parts to review before it is suitable for autonomous release. That intermediate arrangement can create value, but its remaining labour and delay must be included. Product quality control should also be distinguished from functions intended to keep people safe around machinery.

Maintaining qualified inspection and profitable variety

Continuing changes create another cost. New suppliers, surface finishes, lighting and software revisions can alter the relationship between images and accepted output. The manufacturer needs a way to preserve known configurations, evaluate changes and recover when performance worsens. A low licence price can become expensive if each variant requires substantial outside engineering. Conversely, an integrator can support a viable recurring service by maintaining qualified performance across similar tasks, provided the customer does not still carry nearly all the difficult adaptation work.

The commercial test is whether the factory can accept worthwhile orders it previously could not serve, at a satisfactory contribution after inspection and quality costs. Evidence should follow engineering effort per variant, reusable hardware, review workload, defects escaping to customers and retained margin through repeat runs. A larger catalogue is not sufficient. AI changes the economics of variety when the complete system makes reliable small-scale production repeatable, rather than making the first demonstration easier while leaving every subsequent introduction bespoke.