Cards have been scored in real time, at millisecond latency, continuously, since the 1990s. Optical sorters make accept or reject decisions on individual grains of wheat at industrial speed. Bearings have been listened to for failure for decades.
None of it is marketed as AI. All of it is load-bearing in the strict sense: the process stops without it.
The pattern is the same everywhere
The model was the last and cheapest part. The data pipeline underneath it took years, and in several cases more than a decade of instrumentation before anything could be predicted at all.
Which is the opposite of what the market believes
Business cases are written about model quality. Deployments succeed or fail on whether the sensing, the integration and the decision path were built first.
That is not an argument against ambition. It is an argument for sequencing, and for being honest in the business case about which part is actually expensive.