The public image of artificial intelligence is dominated by frontier models: enormous systems, expensive training runs and general-purpose interfaces. In practice, many of the applications that change daily work may come from the opposite direction.
A small model running close to the user can be cheaper, faster and easier to control. It can classify a document before it leaves a device, summarize a narrow stream of internal data, detect an anomaly in a machine or power an interface that would never justify a large inference bill.
The economics change the product
When inference becomes inexpensive, developers stop reserving AI for high-value moments. It can run continuously. That enables products built around ambient assistance rather than explicit prompts.
The trade-off is obvious: smaller systems know less and fail in narrower but still important ways. The design problem shifts from asking whether a model is intelligent to defining exactly where it is trusted.
The useful unit of AI deployment may not be the smartest model. It may be the cheapest reliable decision.
That distinction matters because ubiquity is usually driven by economics before spectacle.