General intelligence is a research goal. The systems that survive a factory acceptance test are narrow, bounded, and dull by comparison.
Two different conversations
There is a conversation about artificial general intelligence — software with human-like breadth, able to teach itself tasks nobody trained it for — and there is a conversation about the model that decides whether a bearing is about to fail. They use the same two letters and almost nothing else in common.
The first is a research programme. Nothing shipping today has autonomous self-control or the ability to carry knowledge from one domain into an unrelated one. The second is ordinary engineering: a defined input, a defined output, and a test that says whether it works.
What "narrow" actually buys you
Narrow is usually said apologetically, as though it were a limitation to apologise for. On a production programme it is the feature. A model trained to classify solder joints cannot write your work instructions, and that is precisely why you can qualify it.
A bounded problem has a bounded failure mode. You can assemble a held-out set of real boards, measure how often the model is wrong, decide what an acceptable escape rate is, and write that number into an acceptance criterion. None of that is available for a system whose scope is "understands manufacturing".
The narrower the claim, the easier it is to prove — which is the only reason it gets past a customer audit.
Where the ambition leaks in
The failure we see most often is not a bad model. It is a good model asked to carry a decision it was never scoped for. A vision system qualified to detect one defect class gets pointed at a new product variant. An anomaly detector validated on one machine gets rolled out across a fleet with different duty cycles.
In both cases nothing was retrained and nothing was revalidated. The scope quietly widened, the acceptance evidence stayed where it was, and the first anyone hears about it is a warranty claim. Widening scope is a change, and it needs the same treatment as changing a component.
How to scope one properly
Write down the decision the model is making, in a sentence, before choosing a technique. "Flag boards whose thermal profile deviates from the qualified window" is a specification. "Use AI for quality" is a budget line.
Then decide three things: what evidence proves it works, what happens when it is wrong, and who is accountable for the decision it informs. If a model is advisory, say so in the interface. If it stops the line, it needs the validation of anything else that stops the line. The honest version of an AI programme looks a lot like every other engineering programme — which is a good sign, not a disappointing one.
Written by the engineering delivery team. Published articles carry a named author once attribution is confirmed.
